About 500 fibre-optic cables carry 97 percent of intercontinental data across the seabed, each roughly the width of a garden hose and lying unprotected on the ocean floor — the new front of geopolitical infrastructure warfare. On Christmas Day 2024, a rusting Cook Islands-flagged tanker named the Eagle S began dragging its anchor across five of them in the Gulf of Finland, at the heart of the Baltic’s unique geographic and infrastructure vulnerability, while Finnish helicopters scrambled to intercept it.
That single incident should have been a success story. Special forces rappelled onto the deck. The vessel was seized. The anchor-drag scar on the seabed provided physical evidence. And yet, ten months later, a Finnish court dismissed the case. The evidence was strong, seabed scars, position data, seized equipment. But the case collapsed on a question no amount of forensic work could answer: whether Finnish law could reach the conduct at all.
If you have ever assumed that catching a saboteur means they face consequences, the Eagle S case will disabuse you of that notion. The real question is not whether you should be worried about undersea cables. It is what you should actually be watching for.
On 25 December 2024, the Eagle S dragged its anchor for roughly 100 kilometres across the Gulf of Finland seabed, damaging five separate cables: the Estlink 2 power interconnector, the C-Lion1 data cable, the Balticonnector gas pipeline link, and two additional telecommunications cables between Finland and Estonia. Finnish special forces rappelled from helicopters to board the vessel, and the National Bureau of Investigation launched a criminal investigation.
The physical evidence was strong. Seabed surveys showed the drag scar. Position data placed the ship directly over the cable corridor. Investigators found transmitting equipment, laptops with Russian and Turkish keyboards, and sensor-type devices aboard, raising suspicions the vessel had a secondary espionage role. The captain was reportedly instructed by radio to destroy charts depicting local subsea cables while under detention.
Finnish intelligence agency SUPO complicated the picture before the court even ruled, stating publicly that investigations found no evidence of deliberate Russian state activity. SUPO attributed the damage to poorly maintained vessels and untrained crews. The statement sits uneasily alongside documented AIS deactivation patterns, but it highlights exactly the problem: a state actor using deniable, poorly maintained proxies is, by design, operating below the threshold where intelligence agencies can make a public call.
On 3 October 2025, the Helsinki District Court dismissed the case, ruling Finland lacked jurisdiction over a negligent act in its Exclusive Economic Zone rather than territorial waters. Under UNCLOS, jurisdiction lay with the flag state, the Cook Islands, or the crew’s nationalities, Georgia and India, neither of which had any incentive to pursue prosecution. The court stressed the allegations involved negligent rather than intentional conduct. Proving intent in grey-zone operations is what makes these cases impossible to prosecute under the legal framework that makes prosecution so difficult.
The Fitburg incident reinforces the pattern. Finnish special forces boarded another Russia-linked vessel in the same EEZ corridor. The ship was carrying Russian steel subject to EU sanctions, which gave Finland grounds for seizure, but the criminal investigation dragged on for 19 months before suspects were referred to prosecutors.
The Eagle S was one vessel in a much larger fleet, one designed from the ground up for deniability. To understand why the case failed, you need to understand the fleet itself.
Russia’s shadow fleet, also called the dark fleet or ghost fleet, is a network of 600 to 1,600 ageing oil tankers and cargo vessels registered under flags of convenience with ownership obscured through cascading shell companies. Its primary purpose is sanctions evasion: following the 2022 G7 oil price cap, these vessels export Russian crude outside Western insurance and shipping frameworks, generating more than €10 billion in monthly revenue. About 48 percent of Russia’s oil exports transit the Baltic Sea.
The fleet’s sabotage capability exploits the Baltic’s shallow average depth of 55 metres through the anchor-drag method. A vessel drags an anchor across a known cable corridor, tearing through fibre-optic cables on the seabed. The method needs no specialised equipment and is inherently deniable: anchor loss is common in commercial shipping, and the Baltic’s narrow cable corridors make it a target-rich environment.
What separates accidents from sabotage is the AIS data. Vessels routinely deactivate their Automatic Identification System transponders before entering cable corridors and reactivate after passing through, behaviour inconsistent with ordinary navigation. Baltic and North Sea states formally identified AIS manipulation as a maritime safety threat in January 2026. Since October 2023, the shadow fleet has been linked to at least eleven cable disruptions in the Baltic.
The fleet’s opacity is intentional. Vessels fly Cook Islands, Panama, or Liberia flags. Ownership runs through shell companies in multiple jurisdictions. Crews are recruited internationally, sometimes via Telegram. Inspections have found equipment atypical for commercial shipping, and some crew members have had links to Russian military or mercenary groups. In May 2025, Russia deployed an Su-35 fighter jet to overfly a shadow fleet vessel, causing Estonian authorities to abandon an interception attempt.
At the other end of Russia’s capability spectrum sits the Main Directorate of Deep-Sea Research, GUGI, operating the spy ship Yantar, deep-sea submersibles, and specialist submarines. A UK-Norway naval operation in late 2025 tracked an Akula-class attack submarine and two GUGI surveillance submarines over Western cable routes. The shadow fleet handles the low-tech, high-volume end; GUGI handles the sophisticated end. Together they demonstrate full-spectrum seabed warfare capability.
If the Baltic is where Russia tests the limits of attribution, the Taiwan Strait is where China rehearses for something more consequential.
China’s threat profile is different in kind, not just degree. Where Russia relies on deniable, distributed sabotage, China has invested in capability built for military contingency, centred on the Taiwan Strait.
The most significant capability is a deep-sea cable-cutting submersible, tested to 3,500 metres depth, well beyond the reach of the fewer than 60 commercial cable-repair ships operating globally. A coordinated deep-water cut could leave regions offline for months.
The Matsu Islands incident in February 2023 demonstrated intent. Chinese vessels severed both fibre-optic cables connecting Taiwan’s Matsu Islands to the main island, causing a 50-day internet blackout for 14,000 residents. Analysts widely assess it as a practice run for wider Taiwan isolation in a conflict scenario. Taiwan has 14 external cable connections; severing most of them would cut the island off from global communications.
China’s primary operational tool is the People’s Armed Forces Maritime Militia, PAFMM, fishing vessels civilian on paper but operating as a PLA subsidiary. The Shunxing 39 incident in January 2025, in which a Chinese-flagged vessel loitered over cable routes in the Taiwan Strait, reinforces the pattern.
The Hong Tai 58 case in February 2025 produced a rare prosecution. A decrepit Chinese-flagged vessel, its cargo holds rusted shut, severed a Chunghwa Telecom cable off Taiwan’s Penghu Islands. Taiwan’s Coast Guard seized it, and the captain received a three-year jail sentence. The vessel appeared to be what analysts describe as a pawn sacrifice, a deliberately expendable asset.
In the Baltic theatre, Chinese-linked vessels complicate the Russia-only narrative. The Newnew Polar Bear was implicated in the October 2023 Balticonnector damage and was observed sailing in tandem with the Russian nuclear icebreaker Sevmorput. The Yi Peng 3 was detained by Denmark in November 2024 for dragging its anchor through Baltic cables before being released.
There is an industrial dimension as well. HMN Tech, formerly Huawei Marine, the state-owned firm leading China’s Digital Silk Road cable projects, has been excluded from Western cable projects on US pressure over espionage concerns. The result is a bifurcated global infrastructure: Chinese-aligned networks operate in a separate technology ecosystem with different security assumptions, and that bifurcation makes unified defence harder.
Russia’s model is distributed deniability. It relies on the shadow fleet’s scale, anchor-drag tactics needing no specialised equipment, and flags of convenience that frustrate attribution. The strategic logic is grey-zone harassment: inflict persistent, low-cost disruption that degrades confidence in infrastructure without crossing the threshold of armed conflict. The cost asymmetry makes this sustainable. Dragging an anchor across a cable corridor costs essentially nothing beyond fuel. Keeping naval frigates on constant alert to shadow hundreds of potential saboteurs is not.
China’s model is centralised capability. It operates state-directed vessels with demonstrated submersible cutting technology. The strategic logic is military contingency: develop and rehearse the ability to isolate Taiwan by severing its cable connections, either as a pre-invasion shaping operation or as a standalone coercion tool. The deeper waters of the Taiwan Strait drive investment in submersible rather than surface-level capability.
Anchor dragging, Russia’s signature method, is deniable and cheap but limited to shallow waters. Submersible cutting, China’s demonstrated edge, is more attributable but effective at depths commercial repair cannot reach, making it more consequential despite less frequent use.
Theatre differences compound the challenge. The Baltic Sea is NATO territory with a multilateral defence architecture, allied naval presence through Baltic Sentry, and the possibility of incident response. The Taiwan Strait is contested, with limited international patrol presence, most cables in international waters, and Taiwan’s SAWS monitoring system lacking physical intervention capability. NATO’s cable defence architecture was not designed for both theatres simultaneously.
Both threat models converge on the same weak point, and the Eagle S case exposed what it looks like when attribution fails at every tier.
Attribution operates on three tiers. Technical attribution, forensic evidence, AIS data, vessel position, is achievable and was demonstrated in the Eagle S case. Legal attribution, prosecutable evidence meeting criminal standards, fails because of jurisdictional architecture. Political attribution, willingness to name a state actor and impose consequences, is constrained by escalation management and alliance consensus.
Flags of convenience are the primary obfuscation mechanism. A vessel registered in the Cook Islands, Panama, or Liberia falls under the flag state’s jurisdiction, and those states have neither capacity nor incentive to investigate sabotage in European waters. The EEZ jurisdictional gap, explained in the Eagle S case above, is the legal vulnerability at the centre of it all: under UNCLOS, coastal states have limited criminal jurisdiction in their EEZ beyond economic offences.
AIS deactivation patterns illustrate the gap. A vessel going dark before entering a cable corridor and reactivating afterwards is operationally damning but legally circumstantial. A defence can always claim equipment failure. The Helsinki court’s ruling means criminal jurisdiction, when tethered to flag state or nationality, is ill-suited to deter grey-zone maritime operations.
As noted in the Eagle S case, SUPO’s finding of no evidence of deliberate state activity captures the intelligence-evidence gap: state actors using deniable proxies operate by design below the threshold where agencies can make a public call.
The Hong Tai 58 prosecution in Taiwan is the rare success, achieved through Taiwan’s results principle. Article 4 of Taiwan’s criminal code extends jurisdiction when the result of an offence occurs within its territory, even if the conduct occurred beyond it. Baltic Sea states are studying this model but have not adopted it.
There is also a resilience dilemma: the more effectively countries build redundancy into their cable networks, the harder it becomes to meet the legal standard for severe disruption. If traffic reroutes within milliseconds, a court may find the damage insufficiently serious, even though the adversary achieved the attrition it intended.
The Eagle S case was a structural inevitability. The system, detection, interception, evidence collection, worked. What failed was the legal architecture it operated within. Russia’s distributed deniability in the Baltic and China’s centralised military contingency in the Taiwan Strait exploit the same vulnerability: the gap between knowing what happened and being able to act on it.
Taiwan’s results principle and the Hong Tai 58 prosecution point toward a possible path, extending domestic criminal jurisdiction to cover damage whose effects are felt within a state’s territory regardless of where the conduct occurred. Whether NATO and allied states are willing to reform the jurisdictional architecture that makes frigate deployments impressive but insufficient is the question that matters for the institutional and legal responses underway.
The cables are still there, still carrying the internet, still as vulnerable as they were on Christmas Day.
When a fibre-optic cable is severed, internet traffic does not stop: it is automatically rerouted through other cables in the global network, often within milliseconds. Users may notice slower speeds or increased latency, but rarely a total blackout. The exception is regions with limited cable connections. Taiwan’s Matsu Islands experienced a 50-day internet outage in 2023 because only two cables served 14,000 residents, and both were cut simultaneously.
Repair typically takes two to four weeks, depending on water depth, weather conditions, and the availability of a cable repair ship. Fewer than 60 such vessels operate globally, and demand spikes after multiple incidents. Shallow-water breaks in the Baltic are usually faster to repair than deep-water cuts, which is partly why China’s demonstrated submersible cutting capability at 3,500 metres is so concerning: it targets depths where commercial repair simply cannot reach.
Yes. While sabotage dominates headlines, signals intelligence collection is the quieter and arguably more persistent threat. Russia’s GUGI-operated spy ship Yantar is equipped with deep-sea submersibles capable of accessing and splicing into fibre-optic cables to intercept data without leaving visible damage. The same vessels that can cut cables can also tap them, making physical inspection and integrity verification extremely difficult. This is why some Western cable routes now avoid waters accessible to known surveillance vessels.
Undersea cables carry approximately 99 percent of all intercontinental internet traffic because they offer vastly higher bandwidth, lower latency, and greater reliability than satellites. A single modern fibre-optic cable can transmit hundreds of terabits per second, while satellite links typically operate in the gigabit range and introduce noticeable delay. Satellites serve as a backup for isolated locations, but they cannot replace the cable network’s capacity. The physical vulnerability of cables is the trade-off for their unmatched performance.
The cost asymmetry is extreme. Dragging an anchor across a cable corridor costs essentially nothing beyond fuel and crew wages, and the vessel can continue its commercial voyage. Repairing a single cable break can cost between one and three million dollars, depending on depth, location, and repair ship availability. When multiple cables are damaged in a single incident, as with the Eagle S dragging its anchor across five separate cables, the repair bill multiplies accordingly. This asymmetry makes sabotage a strategically sustainable attrition campaign for adversaries.
No single attack could take down the entire internet. The global network of approximately 500 active undersea cables includes substantial redundancy, and traffic reroutes automatically when cables fail. However, a coordinated simultaneous attack on multiple cables in a concentrated region could isolate a country or continent. Taiwan’s 14 external cable connections are a specific concern for this reason: severing most or all of them in a coordinated operation would effectively cut the island off from global communications, which is precisely the scenario analysts believe China is rehearsing.
NATO launched Baltic Sentry in early 2025, deploying frigates, maritime patrol aircraft, and uncrewed surface vessels to monitor cable corridors in the Baltic Sea. The operation focuses on deterrence through visible presence and rapid response rather than physical protection of every cable. NATO is also investing in distributed acoustic sensing technology that converts existing fibre-optic cables into seabed surveillance arrays, capable of detecting approaching vessels. However, Baltic Sentry covers only one theatre, and the Taiwan Strait lacks any comparable multilateral protection framework.
Where prosecution fails, states have several non-legal responses. Finland expelled the Eagle S from its waters and its crew remain under travel restrictions. Port states can detain vessels for safety violations discovered during inspection, effectively removing them from operation for months. Diplomatic escalation, including coordinated statements like the Baltic-North Sea joint declaration of January 2026, raises the political cost for flag states. Insurance sanctions are also emerging: some Western insurers now refuse to underwrite vessels with suspicious AIS patterns, making shadow fleet operations more expensive.
Cable operators locate breaks using optical time-domain reflectometry (OTDR), which sends a light pulse down the fibre and measures the time it takes for the reflection to return from the break point. This pinpoints the damage location to within tens of metres. Repair ships then use remotely operated vehicles to inspect the seabed, retrieve the severed ends, and splice in replacement cable. The process is well established, but it assumes a repair ship is available. With fewer than 60 vessels globally, a multi-cable incident creates a queue that can extend outages by weeks.
The Baltic Sea’s cable corridors are the most active sabotage theatre, with shallow depths averaging 55 metres making them accessible to anchor-drag attacks. The 14 cables connecting Taiwan to the outside world are considered the most strategically exposed, because their severance would isolate the island in a conflict scenario and most pass through international waters where protection is minimal. The Red Sea and South China Sea cable routes face a different vulnerability: they are chokepoints where dozens of cables converge, and a single incident could affect connectivity across multiple continents.
Why the Baltic Sea Has Become the Epicentre of Undersea Cable SabotageNinety-nine percent of the world’s intercontinental data travels through fibre-optic cables the diameter of a garden hose, laid on the seabed and unguarded. Since 2022, the Baltic Sea has become the place those cables get severed with striking regularity — part of what has become the emerging front of geopolitical infrastructure warfare. Seven of about ten Baltic incidents occurred between November 2024 and January 2025. Here is why.
In October 2023, the Balticconnector pipeline and a data cable between Finland and Estonia were damaged after the Newnew Polar Bear dragged its anchor 100 nautical miles.
On 17 November 2024, BCS East-West Interlink was cut, reducing a fifth of Lithuania’s internet capacity. C-Lion1 was severed the next day. The Yi Peng 3 from Ust-Luga was over both routes. On Christmas Day, the Eagle S dragged 62 nautical miles, severing EstLink 2 and four data cables. Finnish authorities found transmitting gear aboard. In December 2025 the Fitburg left a 10-kilometre scar. In January 2026 a vessel sailed over an inactive cable before turning toward BCS East.
Median anchor-drag exceeds 60 kilometres. One incident could be chance. Seven in three months, shadow-fleet vessels each time, is not. So why now?
More than 500 cables carry 99% of intercontinental data. The Nord Stream sabotage in September 2022 was the opening. Russia has since turned to grey-zone tactics, operating below armed conflict to avoid NATO Article 5. Cable sabotage damages while preserving deniability, alongside GPS jamming, drone incursions, arson, and cyber operations. The Red Sea cuts at Bab el-Mandeb severed four cables in early 2024. Narrow corridors where density and tension overlap are vulnerable. The Baltic is the most concentrated. What makes it so exposed?
The Baltic’s average depth is 55 metres. In deeper seas, cutting a cable requires submersibles. Here, a standard anchor reaches any cable. Chokepoints compound this: only three Danish Strait passages force roughly 4,000 vessels daily into constrained routes. More than 35 submarine cables share the seabed with pipelines, interconnectors, and wind cables. The Gulf of Finland concentrates the risk: Helsinki-Tallinn cables and EstLink run in parallel, with Russian ports hours away. Then there is the law. Cables transit EEZs, territorial seas, and international waters. The Eagle S case was dismissed in October 2025 because the damage occurred in Finland’s EEZ, not territorial waters. Geography makes attacks simple. Law makes prosecution unreachable. Exploiting both is Russia’s fleet.
Russia’s shadow fleet comprises roughly 600 vessels with opaque registration. Cook Islands flags, AIS spoofed or disabled, shell-company ownership — these are what cable sabotage uses to remain between accident and act of war. At least 50 Russian-linked ships have operated without AIS. The Eagle S (see above) flew a Cook Islands flag and was registered in the UAE. The captain was instructed via radio to destroy cable charts. The Fitburg carried sanctioned Russian steel. Both were detained. Neither detention reached Moscow. Catching the ship isn’t proving who sent it.
Three indicators matter. First, anchor-drag: accidental marks are metres; the Yi Peng 3 dragged roughly 300 kilometres. Second, AIS: disabling transponders near cables and reappearing past the route is inconsistent with innocent passage. The January 2026 Liepāja incident showed a vessel sailing over an inactive cable before turning toward the active one. Third, provenance: Russian port departures, Cook Islands flags, shell companies, sanctions evasion shift probability from accident toward sabotage. The prosecution gap remains wide. If attacks cannot be deterred, what happens next?
Optical time domain reflectometry locates the break. A repair vessel is mobilised from roughly 75 ships. Grapnel hooks retrieve the segment; replacement sections are spliced, each joint up to 16 hours. Restoration takes at least 14 days; the global median is 40. A new vessel costs over €50 million, and the market is sized for routine faults. Redundancy helps: well-connected states reroute instantly. The EU’s €347 million subsea package is a start, not a fix.
Cable landing stations connect submarine cables to terrestrial networks. They are unremarkable coastal buildings in Helsinki, Stockholm, Tallinn, and Liepāja. One station terminates multiple cables, so ground-level access can sever several connections. Physical security is the cable owner’s responsibility, not the state’s, and most lack hardened perimeters. The January 2026 Liepāja incident, where the suspect vessel docked next to the damaged station, shows how port proximity blurs maritime and shore-based vulnerability. The seabed gets the attention. The shore may be next.
The Baltic became the epicentre because its geography, Russia’s fleet, and the region’s legal architecture form a self-reinforcing system. Shallow water makes attacks simple. Spoofed transponders and shell companies make them unattributable. A patchwork of jurisdictions makes prosecution unreachable. The incidents are not scattered accidents. They are a strategy. This article has examined why the Baltic is uniquely exposed; the broader strategic context of undersea cable security reveals how this vulnerability fits into a global pattern of infrastructure warfare. And geography is only the starting point: the vessels and methods driving the Baltic’s attack surge are the next piece of the puzzle.
Internet traffic does not stop; it reroutes. Modern cable networks are built as meshes, so when a cable like C-Lion1 between Finland and Germany is cut, data automatically shifts to alternative routes across Sweden, Denmark, or Poland. This occurs in milliseconds. Baltic nations with fewer backup cables experience slower speeds and higher latency during repairs, but connectivity is not lost entirely.
No, satellites carry less than one percent of intercontinental data. Undersea cables move approximately 99 percent of all international internet traffic, including financial transactions, diplomatic communications, and everyday browsing. Satellites serve remote regions and broadcast applications, but their bandwidth is a tiny fraction of a fibre-optic cable’s capacity. The global internet is a seabed network, not a space-based one.
No. NATO has never invoked Article 5, its collective defence clause, in response to undersea cable sabotage. The alliance’s January 2025 Baltic Sentry mission represents a step change: increased naval patrols, surveillance drones, and a dedicated maritime watch. However, the ambiguous nature of grey-zone attacks, designed to fall below the threshold of armed conflict, has so far kept Article 5 off the table.
Almost all Baltic undersea cables are privately owned, typically by telecommunications consortiums and specialist cable operators such as Arelion, Cinia, and Elisa Corporation. Governments do not own them, and protection is overwhelmingly the cable owner’s financial responsibility. This creates a structural gap: the operator bears the repair cost, while the state bears the strategic consequence of lost connectivity.
Cable burial is possible but not a complete solution. Standard ploughs trench cables one to two metres into soft sediment, which protects against fishing gear but not a multi-tonne ship anchor dragged for kilometres. In the Baltic’s hard glacial seabed, burial is technically difficult and expensive. Even buried cables are vulnerable when anchors penetrate deeper than the trench or when the cable crosses bedrock where trenching is impossible.
The Baltic is the most active theatre but not the only one. The February 2024 Red Sea cable cuts at the Bab el-Mandeb chokepoint severed multiple cables connecting Europe to Asia, demonstrating that any narrow maritime corridor where cable density and geopolitical tension overlap is vulnerable. The South China Sea, the Strait of Malacca, and waters around Taiwan are assessed as high-risk zones with similar geographic and strategic profiles.
A single cable repair typically costs between one and three million euros, but the indirect economic cost is far larger. Financial markets rely on cable routes for low-latency transactions, and even brief rerouting adds milliseconds that algorithmic trading systems cannot tolerate. For a cable serving multiple nations, the disruption to banking, cloud services, and enterprise connectivity can accrue losses in the tens of millions daily.
Not with standard monitoring. Most cable operators detect a break only after the optical signal is lost, triggering an optical time domain reflectometry scan to locate the fault. Distributed acoustic sensing (DAS) technology can theoretically detect the sound of an approaching anchor while it is still dragging, but DAS is not widely deployed on Baltic cables. Current detection is reactive, not preventative.
The primary response is NATO’s Baltic Sentry mission, launched in January 2025, which deploys frigates, patrol aircraft, and naval drones to monitor cable corridors. The EU has committed 347 million euros to subsea infrastructure resilience for 2026 to 2027, including a 20 million euro rapid repair pilot. Finland and Sweden have also tightened port inspection regimes, particularly for vessels departing from Russian ports with shadow-fleet characteristics.
Fishing trawlers do damage cables accidentally, particularly in shallow waters where bottom trawling overlaps with cable routes, and such incidents have been recorded for decades. The Baltic’s recent sabotage cluster is distinguished by extreme anchor-drag distances exceeding 60 kilometres and systematic AIS manipulation, neither of which is consistent with fishing activity. Accidental trawler damage and deliberate anchor sabotage are operationally distinct.
Proving state direction requires evidence of a direct order, official knowledge, or operational control, which is almost impossible to obtain without intelligence penetration or a defector. Even when a vessel is caught in the act, as with the Eagle S, ownership layering through shell companies and opaque flag registrations severs the legal chain to Moscow. Attribution under criminal standards of proof remains the central unsolved problem.
A coordinated multi-cable attack would be significantly more disruptive than a single cut. While well-meshed nations such as Germany and Sweden would retain connectivity, Estonia, Finland, and Latvia, which have fewer alternative routes across the Gulf of Finland and the eastern Baltic, could experience meaningful internet degradation. Simultaneous cuts to power interconnectors like EstLink would compound the impact by affecting electricity supply.
The AI Industry Goes Public: Anthropic, OpenAI, and the Biggest Capital Event in Tech HistoryIn the space of seven days in June 2026, the AI industry crossed a threshold that will reshape technology investment. Anthropic filed its confidential S-1 on 1 June. OpenAI followed on 8 June. Both offerings are projected to exceed $100 billion each, potentially ranking among the three largest IPOs in history. Together with SpaceX’s $1.75 trillion June IPO, they represent roughly $4 trillion in new public-market capitalisation concentrated within a few quarters, the densest cluster of mega-IPOs in market history.
They arrive against a backdrop of capital concentration without precedent: AI absorbed 80% of all global venture capital in Q1 2026 while hyperscalers committed over $700 billion to infrastructure through 2027. This is the moment venture-funded AI labs become public-market institutions, and every structural dynamic of the industry, from how capital is raised to how governance functions to how risk is priced, will be tested by quarterly earnings calls.
This page maps the entire landscape. The cluster articles below examine each dimension in depth: the S-1 race itself, the governance structures that have no public-market precedent, the capital flood that produced the wave, and the evaluation frameworks you need to navigate what comes next.
Inside the IPO Race Between Anthropic and OpenAI — How the confidential S-1 process works, why both companies filed within a week of each other, and how their business models and valuations compare.
Can AI Governance Survive the Public Markets — What Anthropic’s Public Benefit Corporation and OpenAI’s foundation-controlled structure mean for shareholders, and whether governance designed to constrain profit can endure quarterly earnings pressure.
Why AI Companies Took 80 Percent of Global Venture Capital in 2026 — The capital concentration that produced the IPO wave, how it compares to the dot-com era, and what hyperscaler infrastructure spending signals about the industry’s trajectory.
How to Evaluate an AI Company Before It Goes Public — The risk factors and revenue-quality metrics that traditional tech IPO frameworks miss, from model obsolescence to compute supply chain dependency.
Anthropic’s confidential S-1 filing on 1 June 2026 marked the first pure-play frontier AI company to begin the public-offering process. The filing came just four days after the company closed a $65 billion Series H at a $965 billion post-money valuation, the final private-market milestone before the public transition. OpenAI followed on 8 June, confirming the filing itself rather than letting it leak. Both companies are projected to list at valuations exceeding $1 trillion, with Anthropic targeting a December 2026 debut and OpenAI signalling a March 2027 timeline.
The filings are confidential under the JOBS Act, which means no financials are public yet. The Rule 135 announcements from both companies confirmed only that filings exist. Both chose the “option” language, making clear that a public offering depends on market conditions and SEC review. Goldman Sachs and Morgan Stanley are bookrunning both deals, an unusual concentration of advisory power, and Anthropic added JPMorgan Chase as a third lead within 48 hours of filing.
What we actually know right now is limited but telling. Anthropic projects its first-ever operating profit of roughly $559 million for Q2 2026 on nearly $11 billion of quarterly revenue, though the company told investors heavy compute commitments will likely erase margins in late 2026 and early 2027. The operating profit benefits from a compute ramp-up discount period, so it is a timing effect, not evidence of a structurally profitable business. OpenAI generates roughly $2 billion in monthly revenue but carries a projected $25 billion-plus GAAP loss for 2026. Unit economics, compute costs, and customer concentration remain unknown until the public S-1s drop. The confidential review period typically runs three to six months, and investors are operating on incomplete information that will persist until the roadshow phase.
Read the full breakdown: how the confidential filing process works and what investors know right now
The JOBS Act allows Emerging Growth Companies, those with revenue under $1.235 billion, to submit a draft Form S-1 to the SEC for confidential, non-public review. Both Anthropic and OpenAI qualify. The SEC reviews the filing, issues comment letters, the company amends, and the cycle repeats until the regulator is satisfied. Only then does the S-1 become public, typically 15 days before the roadshow launches.
The confidential route exists to reduce the penalty for companies that test the public-market waters and decide not to proceed. For Anthropic and OpenAI, that optionality is explicit in their own language: the offering “will depend on market conditions and other factors.” But the procedural benchmarks are getting clearer. SpaceX moved from confidential filing to public S-1 in roughly 72 days. Airbnb ran four months end to end. Klaviyo and ARM both ran three to four months. Reddit dragged on for nine months across multiple comment rounds. For Anthropic investors, the SpaceX timeline is the one to watch.
The Rule 135 announcement that both companies published is a legal formality. It tells the market only that a filing exists without triggering securities-offering restrictions. It reveals nothing about revenue, compute costs, governance mechanics, or risk factors. That information vacuum creates a window where secondary-market trading, prediction markets (Polymarket gives 69% odds on Anthropic listing by October 31), and analyst estimates are the only pricing signals available. The most consequential regulatory question in the Anthropic filing is how the SEC treats its gross-vs-net revenue recognition (see the FAQ below for details).
Explore the S-1 mechanics in detail: Inside the IPO Race Between Anthropic and OpenAI
Four forces converged in mid-2026 to make this the moment. Capital ceilings came first. Anthropic’s $65 billion Series H and OpenAI’s $122 billion raise pushed private valuations to levels where limited partners are maxed out. Venture capital funds have concentration caps. Pension funds have allocation limits. Sovereign wealth funds have governance requirements that constrain how much they can deploy into a single private company. When a funding round reaches $65 billion, the pool of investors who can write that cheque and stay within their mandates shrinks. The public market solves this by dispersing ownership across millions of shareholders.
Competitive timing is the second driver. The company that prices first establishes the valuation benchmark against which every subsequent AI IPO will be measured. This matters because Anthropic and OpenAI are projected to list at similar valuations on vastly different revenue bases. The first S-1 to go public will reveal unit economics, compute costs, and customer concentration that will flow through the entire AI funding ecosystem. The Polymarket odds reflect this: 83% chance Anthropic beats OpenAI to the bell.
The third force is the market window itself. SpaceX’s $1.75 trillion IPO in June 2026 demonstrated three things: public markets can absorb trillion-dollar technology offerings, retail demand is real (30% allocation, four times oversubscribed), and unconventional governance structures do not deter institutional capital. It cleared the path by demonstrating the demand exists. It also consumed institutional capital that now must be replenished before the next mega-IPO lands.
The fourth driver is employee liquidity. Both companies have large workforces compensated in equity that becomes liquid only through a public listing. After years of paper wealth tied to private valuations, the pressure from within to create a path to liquidity is real and growing.
Learn more about the strategic calculus: why both companies filed within a single week
Anthropic and OpenAI represent competing bets on the foundational AI market. Anthropic is a safety-first enterprise play: API revenue via Claude, a narrower product suite, Public Benefit Corporation governance, and an expected valuation of $1 to $1.2 trillion. OpenAI is a platform play spanning ChatGPT consumer subscriptions, API developer ecosystem, and enterprise partnerships with Microsoft, supporting a $1.1 to $1.5 trillion projection. Both can credibly target hundred-billion-dollar-plus offerings, but their paths to sustainable public-market economics diverge.
The revenue profiles tell the story. Anthropic’s revenue is concentrated in enterprise API contracts, stickier and more predictable but with narrower customer diversification. About 80% of revenue comes from enterprise customers, with more than a thousand businesses now spending $1 million or more annually. OpenAI’s revenue spans consumer subscriptions (900 million weekly active users, 50 million-plus subscribers), API consumption, and enterprise deals through Microsoft. Broader diversification, but higher volatility in the consumer segment. OpenAI also carries a structural overhang: a Microsoft revenue-share agreement reportedly granting Redmond 20% of revenue through 2030.
Cost structures add another layer of divergence. Both companies face the same structural dynamic: compute is the dominant input. NVIDIA’s data-centre revenue (which hit $62.3 billion in a single quarter) represents an effective tax on every AI company. The difference is in how each manages the compute-to-revenue ratio. Anthropic’s enterprise focus may produce higher revenue-per-compute-dollar efficiency. OpenAI’s consumer free tier consumes massive inference compute without direct revenue. Anthropic’s Q2 2026 operating profit of $559 million looks like a profitability signal, but it benefits from a compute ramp-up discount period, and the company told investors heavy compute commitments are expected to erase margins in late 2026 and early 2027.
The valuation context makes this comparison more than academic. The foundational AI market is projected at a trillion dollars-plus over the next decade, making both companies’ valuations defensible against total addressable market. But the margin of error is thin. The revenue recognition question, the compute-cost trajectory, and model obsolescence risk all introduce valuation uncertainty that traditional comparables analysis cannot capture. When the S-1s go public, the market will see these numbers for the first time, and the gap between the two companies’ economics will determine whether one becomes the benchmark and the other the discount.
See the full business-model comparison: Inside the IPO Race Between Anthropic and OpenAI
Neither company uses a conventional Delaware C-corp. Anthropic is a Public Benefit Corporation, meaning directors are legally obligated to balance shareholder returns against the company’s stated public benefit: the safe development of AI. OpenAI, after restructuring in October 2025, now operates a similar PBC structure with a nonprofit foundation holding equity valued at approximately $130 billion and retaining authority to appoint board members and address safety concerns. Both structures are untested at hundred-billion-dollar-plus public-market scale. For investors, the question is not which structure is “better” but whether the governance mechanism that attracted private capital, mission constraints on profit maximisation, will survive quarterly earnings pressure, activist campaigns, and the market’s expectation that every quarter’s numbers improve on the last.
The structural details matter. Anthropic’s PBC under Delaware law requires directors to balance pecuniary shareholder interests against the corporation’s public benefit purpose. Key features include the Long-Term Benefit Trust, whose trustees can recruit and remove board members if the company deviates from its safety mission, along with a staggered board and supermajority voting requirements for structural changes. Amazon holds an estimated mid-to-high-teens equity stake and Google holds roughly 14% (hard-capped at 15%). Neither holds voting rights, board seats, or observer rights. Anthropic’s structure is more restrictive than any tech IPO in recent memory.
OpenAI’s foundation-controlled model places ultimate authority with a nonprofit board that can step in to address safety concerns. The foundation can appoint members of the for-profit board and, through a special committee, intervene if AI safety is at risk. The key difference from Anthropic: Anthropic’s constraint is baked into the corporate form itself (directors have a legal duty to balance), while OpenAI’s constraint is structural (the nonprofit holds ultimate authority).
The quarterly earnings tension is where theory meets reality. Will Anthropic’s Long-Term Benefit Trust block a deployment that could generate hundreds of millions in quarterly revenue but raises safety concerns? Will OpenAI’s nonprofit board constrain the pace of consumer product launches to manage misuse risk? If either company misses an earnings estimate for safety-related reasons, will the market treat it as a one-off event or reprice the entire governance discount?
SpaceX provides a partial precedent. Its dual-class share structure insulated Elon Musk’s mission control, and the market priced it, investors accepted the tradeoff. But SpaceX’s mission (colonise Mars) carries no tension against quarterly revenue. AI governance introduces a dimension no public company has ever had to manage: the safety-vs-revenue tradeoff where a company may choose to forgo revenue for safety reasons. The disclosure requirements around it, what safety incidents must be reported, what revenue was forgone, are uncharted territory for the SEC and for investors.
Dig deeper: Can AI Governance Survive the Public Markets
The governance questions matter because of the sheer scale of capital flowing into the companies that will face them. Three dynamics drove the concentration.
First, compute costs. Frontier model training runs cost hundreds of millions of dollars, and inference at scale costs more. Every dollar of AI revenue requires massive upfront infrastructure investment, and that investment flows through to a single point: NVIDIA, which reported $62.3 billion in data-centre revenue in one quarter. The compute requirement creates a natural barrier that concentrates funding in the companies that can demonstrate a path to frontier-model capability.
Second, the late-stage funnel narrowed sharply. Ninety percent of Q1 2026 AI VC went to late-stage deals across only 10% of total deal volume. The number of deals actually declined relative to 2022 levels. The record was driven entirely by larger round sizes rather than broader ecosystem growth. Fewer companies are getting larger cheques, and the six fund managers who raised $36.4 billion in Q1 (Andreessen Horowitz, Thrive Capital, Founders Fund, and a handful of others) controlled 76% of all venture fundraising for the quarter.
Third, three companies dominated. OpenAI ($122 billion), Anthropic ($30 billion), and xAI ($20 billion) together accounted for roughly two-thirds of the $42 billion deployed across just three deals out of more than 1,500 total. All the while, the non-AI remainder, $58 billion, was spread across every other sector in existence: healthcare, climate, fintech, robotics, biotech.
What this means for the IPO wave is straightforward. The VC concentration means the IPO of a small number of firms reconfigures the entire AI funding ecosystem. When Anthropic and OpenAI’s S-1s go public, the transparency shock (unit economics, compute costs, customer concentration) will flow through to every private AI valuation. Companies that raised at premium multiples based on private-market benchmarks will find themselves repriced against public-market comparables. The capital concentration analysis examines this dynamic in depth.
Read the full capital concentration analysis: Why AI Companies Took 80 Percent of Global Venture Capital in 2026
The parallels are structural: extreme capital concentration in a single technology theme, narrative-driven valuations, and an infrastructure buildout preceding widespread monetisation. The divergences are material: today’s AI companies have real revenue, and the capital requirement is driven by compute costs rather than customer acquisition costs. The dot-com era peaked at roughly 50% of VC flowing to internet companies. AI’s 80% share is higher, but concentrated in far fewer companies with functioning business models.
Both eras feature a genuine technology revolution coinciding with capital concentration that raises sustainability questions. The dot-com era saw the internet’s infrastructure built (fibre optic networks, data centres, e-commerce platforms) before the revenue models caught up. The AI era is building compute infrastructure (GPU clusters, data centres, energy capacity) before AI-native business models are fully proven at scale. The key structural difference is that dot-com infrastructure was built by a different set of companies from the ones that ultimately captured the value. In AI, the companies building the infrastructure (hyperscalers) and the companies building the models (Anthropic, OpenAI) are distinct but interdependent.
The risk divergence is where the comparison gets most useful. Dot-com companies failed because they had no revenue, no path to revenue, and burned through capital on customer acquisition that never converted. AI companies have functioning business models, but those models are structurally exposed to forces that traditional valuation frameworks do not capture. Model obsolescence can render a product line non-competitive inside a single quarter. Compute-cost inflation can destroy gross margins regardless of revenue growth. A safety incident can trigger regulatory intervention that freezes revenue overnight.
At the peak of the dot-com era, about 7% of the fibre-optic network was being used. Today, data centre utilisation hovers around 80%. The infrastructure is being used as it is built, not speculatively deployed. Overvaluation remains possible, but the mechanism of potential failure differs from the dot-com era, so the frameworks you use to assess it need to account for that.
Explore the dot-com parallels in depth: Why AI Companies Took 80 Percent of Global Venture Capital in 2026
Combined 2026 capital expenditure from the four major hyperscalers is tracking roughly $700 billion, the largest concentrated infrastructure cycle in tech history. The spending leaders: Amazon at approximately $200 billion, Alphabet at $180 to $190 billion (up from $91 billion in 2025), Microsoft at approximately $190 billion, and Meta at $125 to $145 billion. Goldman Sachs projects annual AI infrastructure spending growing to $1.6 trillion by 2031. The spending signals that AI infrastructure is structural, not cyclical, and that even two-trillion-dollar-plus incumbents are straining against debt-market limits to fund it.
Alphabet’s capital-structure decision is the clearest signal. The company is raising $80 billion in equity capital, including $30 billion in underwritten stock offerings and $40 billion through an at-the-market offering programme beginning in Q3, rather than issuing debt. Equity is more expensive than debt in nominal terms, but choosing it preserves balance-sheet flexibility and credit ratings. It also signals that Alphabet views AI infrastructure spending as a multi-year structural commitment, not a one-off investment cycle that debt could bridge.
Berkshire Hathaway’s $10 billion commitment to Alphabet’s raise marks a pivot. Warren Buffett has historically avoided technology investments, and his Apple position was justified as a consumer-brand play rather than a technology bet. The Alphabet commitment signals that even value-oriented institutional capital, the capital that sat out the dot-com era, is being pulled into the AI capex cycle. Whether this is a confirming signal of durable value or a late-cycle indicator depends on your read of the capex-to-cash-flow funnel, and that funnel is getting wider, not narrower. Sequoia’s David Cahn calculated approximately a $600 billion annual revenue gap between what hyperscalers are spending on AI infrastructure and what the AI ecosystem is generating in actual revenue, and the gap is widening in 2026.
Then there is the physical constraint that capital cannot manufacture. The IEA projects data-centre electricity consumption roughly doubling to nearly 1,000 TWh by 2030, equivalent to the current electricity demand of Japan. Data centres already consumed about 4.4% of total U.S. electricity in 2023 and could reach 12% by 2028. Anthropic estimated that training a single frontier AI model will require five gigawatts of power by 2027. Capital can fund GPU purchases. It cannot manufacture additional grid capacity on the same timeline, and that energy constraint is the practical limit of AI infrastructure spending.
See the full hyperscaler breakdown: Why AI Companies Took 80 Percent of Global Venture Capital in 2026
Traditional tech IPOs carry market adoption risk, competitive moat durability risk, and unit-economics-at-scale risk. AI IPOs add four categories with no SaaS analogue. Model obsolescence: a competitor’s model release can render your product non-competitive in a single quarter, and the depreciation cycle in AI is measured in months, not years. Compute supply chain dependency: every AI company is structurally exposed to NVIDIA’s pricing and availability, and there is no short-term substitute at frontier-training scale. The safety-vs-revenue tradeoff: a safety incident (model misuse, harmful outputs, regulatory intervention) can freeze revenue overnight in ways that have no parallel in SaaS or consumer tech. And the valuation information gap: public-market investors will price these risks with less information than private investors had.
Anthropic’s regulatory situation illustrates how the safety-vs-revenue tradeoff operates in practice. A March 2026 Pentagon supply-chain designation bars military use of Claude. A June 2026 Commerce Department export control directive forced two Claude models offline for foreign nationals. The D.C. Circuit Court has the case under advisement. For a company approaching a trillion-dollar valuation, the government’s demonstrated willingness to restrict its products represents a risk category with no settled SEC disclosure template. OpenAI, with its different government relationships, faces a different regulatory risk profile.
Whether public AI investment offers better risk-adjusted returns than private venture capital depends on whether the public market properly prices these novel risk categories. Historically, novel risk categories tend to be mispriced, in either direction, for years. Private investors accepted model obsolescence risk and compute dependency in exchange for information asymmetry and board access. Public investors get liquidity and transparency but face quarterly-volatility risk and the possibility that IPO pricing embeds private-market optimism that public scrutiny will correct. The detailed framework, including the retail investor lens and the revenue-quality metrics that matter most, is in the full evaluation article.
Dig deeper: How to Evaluate an AI Company Before It Goes Public
Traditional SaaS metrics (ARR, NRR, CAC payback, gross margin) are necessary but insufficient for AI companies. Three additional lenses matter most. First, the API-vs-enterprise mix: usage-based API revenue is inherently more volatile than contracted enterprise revenue. A customer running an AI workload can switch models or reduce usage instantly. An enterprise contract with a 12-month commitment provides visibility. Anthropic’s 80% enterprise concentration may produce higher revenue durability than OpenAI’s consumer-heavy mix, but the S-1s will reveal whether that thesis holds.
Second, the revenue-to-compute ratio. AI companies operate at 50 to 60% gross margins versus 80 to 90% for SaaS. Every AI query incurs real compute costs. NVIDIA’s pricing power means an AI company’s gross margin can compress without any change in its own operations. An AI company with 50% gross margins after compute costs is a fundamentally different business from a SaaS company at 80%, and the revenue-to-compute ratio, how much revenue each dollar of compute cost generates, is the AI equivalent of unit economics.
Third, model-version revenue attribution. How much current revenue depends on a specific model version that could be superseded? SaaS companies do not face this question because their product is cumulative. AI companies’ revenue is tied to a specific model iteration, and when the next competitor model ships, revenue can reset. This is a depreciation cycle with no SaaS analogue, and you should expect it to produce revenue volatility that standard SaaS frameworks do not anticipate.
Free-tier dynamics add another dimension. AI companies with large free user bases (ChatGPT’s 900 million weekly active users, Claude’s free tier) have conversion funnels that look more like consumer internet than enterprise SaaS. The free-tier cost in inference compute and the conversion rate to paid tiers are metrics not captured by standard SaaS frameworks. And then there is the disclosure question: does the company quantify the revenue it forgoes for safety reasons? If not, investors are pricing an unknown liability.
Read the full evaluation framework: How to Evaluate an AI Company Before It Goes Public
Three dynamics will shift the AI industry once public-market reporting begins. First, the transparency shock. When Anthropic’s S-1 goes public, it will disclose revenue concentration, compute costs as a percentage of revenue, customer retention metrics, and safety-related risk factors that no AI company has ever reported publicly. This information will function as a benchmark against which every private AI company will be repriced, including those that raised at premium multiples based on selective disclosure to private investors. The transparency cascade will be particularly acute for companies that followed the same capital-concentration pattern as Anthropic and OpenAI but lack their revenue scale.
Second, the governance stress test will move from theory to practice. The tension between mission-driven governance and quarterly earnings is not abstract. Will Anthropic’s Long-Term Benefit Trust block a deployment that could generate hundreds of millions in quarterly revenue but raises safety concerns? Will OpenAI’s foundation board constrain the pace of consumer product launches to manage misuse risk? If either company misses an earnings estimate for safety-related reasons, the market must decide whether to treat it as a one-off event or reprice the entire governance discount. These are questions private markets never had to answer because reporting was voluntary and timelines were flexible.
Third, the competitive reset will determine the structure of the AI industry for the next decade. Once Anthropic and OpenAI report quarterly, the valuation gap between the largest labs and everyone else will either widen or compress. If the market rewards revenue scale and punishes smaller competitors for compute-cost disadvantage, the winner-take-all dynamic accelerates and makes it impossible for new entrants to raise capital. If the market prices in model obsolescence risk, treating every AI company’s revenue as contingent on the next model release, the valuation gap compresses and the entire sector trades at a discount to traditional tech, making AI a structurally cheaper sector for acquirers.
Either outcome changes the AI funding ecosystem. The path that materialises depends on whether public-market investors price the risks that private investors accepted in exchange for access and information. The governance analysis and the evaluation framework provide the tools to navigate whichever path unfolds.
Explore the second-order effects: Can AI Governance Survive the Public Markets · How to Evaluate an AI Company Before It Goes Public
Inside the IPO Race Between Anthropic and OpenAI — How the confidential S-1 process works, why both companies filed within a single week, and how their business models and projected valuations compare. Read this first if you want to understand the mechanics of the dual filing and what the Rule 135 announcements actually tell you.
Why AI Companies Took 80 Percent of Global Venture Capital in 2026 — The capital concentration that produced the IPO wave: the $42 billion Q1, three companies absorbing two-thirds of it, and hyperscalers committing over $700 billion to infrastructure. Read this second to understand the market structure that funnelled capital into a handful of companies and what it means for the IPO pipeline.
Can AI Governance Survive the Public Markets — Anthropic’s Public Benefit Corporation versus OpenAI’s foundation-controlled structure: what each means for shareholder rights, how the safety-vs-revenue tradeoff will be tested by quarterly earnings, and whether governance designed to constrain profit can endure activist pressure. Read this to understand the structural question that separates AI IPOs from every tech IPO before them.
How to Evaluate an AI Company Before It Goes Public — The risk factors and revenue-quality metrics that traditional tech IPO frameworks miss: model obsolescence, compute supply chain dependency, API-vs-enterprise revenue mix, and the safety-vs-revenue disclosure gap. Read this last to equip yourself with the evaluation lens you need to assess any AI IPO, starting with the two happening now.
Suggested reading order: Start with “Inside the IPO Race” to understand the event, then “80 Percent of Global Venture Capital” for the capital context, then “Can AI Governance Survive” for the structural novelty, and finish with “How to Evaluate” for the action framework.
When will investors see the actual financials in the S-1 filings?
The confidential review process typically runs three to six months, during which the SEC issues comment letters and both companies amend their filings. The public S-1, including financials, risk factors, and business description, will be filed only after the SEC is satisfied with the review. Based on SpaceX’s 72-day benchmark, Anthropic’s public S-1 could appear as early as August 2026, with OpenAI’s following. Until then, secondary-market trading and prediction markets are the only pricing signals available. The IPO race explainer covers the full process timeline.
Could Anthropic’s long-term benefit trust block a takeover bid that shareholders want?
Yes, that is part of its design. The governance deep-dive examines how these protections interact with shareholder rights in detail.
Is the 2026 AI IPO wave more like the dot-com bubble or a genuine technology buildout?
It contains elements of both, which is what makes the question difficult to answer with a binary yes or no. The dot-com era saw real internet infrastructure built alongside widespread overvaluation. The AI era features real revenue and functioning business models alongside capital concentration that historically correlates with bubble dynamics. The key difference is compute costs: AI’s capital concentration is driven by the physical cost of training and running models, not by speculative customer-acquisition spending. The capital concentration analysis provides the full comparison framework.
Which hyperscaler is spending the most on AI infrastructure, and what does it signal?
Amazon leads at roughly $200 billion, with Alphabet and Microsoft close behind at $180 to $190 billion each, followed by Meta at $125 to $145 billion. Alphabet’s $80 billion equity raise is the signal: choosing equity over debt at this scale indicates that the company views AI infrastructure as a multi-decade structural commitment rather than a cycle that debt could bridge. Berkshire Hathaway’s $10 billion participation reinforces the signal. The capital flood article details the spending comparisons and capital-structure decisions.
What happens to OpenAI’s capped-profit structure if the company needs more capital than the cap allows?
This is the structural question that the public markets will eventually test. After its October 2025 restructuring, OpenAI operates a PBC with a nonprofit foundation holding ultimate authority. If the company needs capital beyond what the structure makes attractive to investors, it faces options that are all difficult at trillion-dollar-plus scale: renegotiate the structure (which the foundation controls), convert to a different form (legally complex), or access debt markets. The governance analysis examines how each structure handles capital-constraint scenarios.
How dependent are AI companies on NVIDIA, and what happens if GPU supply is constrained?
Structurally dependent. There is no short-term substitute for NVIDIA’s data-centre GPUs at frontier-training scale. NVIDIA’s $62.3 billion in single-quarter data-centre revenue reflects this dependency. If supply is constrained, every AI company’s gross margin compresses simultaneously. This is a systemic risk, not a company-specific one. The evaluation framework article explains how to assess compute supply chain exposure when evaluating an AI IPO.
Which offers better risk-adjusted returns: investing in AI through VC funds or buying AI IPOs?
There is no settled answer, and the question itself is the point. VC investors get information asymmetry and board access but face illiquidity and concentration risk. Public-market investors get liquidity, transparency, and diversification but face quarterly-volatility risk and the possibility that IPO pricing embeds private-market optimism that public scrutiny will correct. The outcome depends on whether the public market properly prices AI-specific risks (model obsolescence, compute dependency, safety-vs-revenue tradeoffs). Historically, novel risk categories are mispriced for years. The evaluation article provides the full public-vs-private framework.
What regulatory risks could derail these IPOs beyond the standard SEC review?
Anthropic faces two distinctive risks: a March 2026 Pentagon supply-chain designation barring military use of Claude, and a June 2026 Commerce Department export control directive forcing two Claude models offline for foreign nationals. The D.C. Circuit Court has the case under advisement. For a company approaching a trillion-dollar valuation, the government’s demonstrated willingness to restrict its products represents a risk category with no settled SEC disclosure template. It is a risk that OpenAI, with its different government relationships, does not share in the same form.
The AI industry’s structural future will play out not in boardrooms but in the quarterly earnings cycle. Anthropic and OpenAI’s S-1 filings are the opening shot. The capital concentration that produced them, the governance structures they bring to market, and the risk frameworks investors use to price them will determine whether this moment is remembered as the maturation of a transformative industry or the densest cluster of overvaluation in market history. The four articles in this cluster give you the framework to navigate either outcome. Start with whichever dimension matters most to you.
How to Evaluate an AI Company Before It Goes Public: Metrics, Risks, and Due DiligenceWhen an AI S-1 lands on your desk, the due diligence checklist you have refined over two decades of SaaS IPOs has a gap. No box for “what happens if a competitor ships a model that destroys this company’s competitive position within 90 days?” No line for “how much of this revenue comes from customers who also hold equity in the company?”
The frameworks you already know (ARR, NRR above 120%, CAC payback, gross margin) remain the foundation. But AI companies carry risks those frameworks were not designed to price — risks made urgent by the moment when AI companies begin trading on public exchanges. Here are the questions that expose what they miss, organised around dimensions traditional evaluation cannot capture.
Traditional tech IPO risks (market adoption, moat durability, unit economics, execution) do not vanish when the company builds AI. They compound with four new dimensions.
Model obsolescence: a frontier model is a depreciating asset with a competitive lifespan counted in quarters. Software products accumulate features. AI models get superseded by a single release.
Compute supply chain dependency: every AI company depends on NVIDIA. A supply disruption or pricing shift flows directly to gross margin. SaaS companies face infrastructure costs but not single-supplier dependency at this scale.
The safety-versus-revenue tradeoff: a safety incident (model misuse, harmful outputs, regulatory intervention) can destroy revenue overnight. Compliance budgets alone cannot address this.
Public-versus-private dynamics: venture investors accepted these risks for board access and information advantage. Public markets will price them with less information and more quarterly-volatility pressure. Novel risk categories tend to be mispriced for years before finding equilibrium, and quarterly earnings — central to the biggest capital event in tech history — will be the mechanism that exposes the gap.
SaaS revenue quality is assessed through ARR, NRR, CAC payback, and gross margin. These remain relevant but are insufficient for AI companies.
The dominant AI revenue model is consumption-based. Customers pay per token, not per seat. A customer can reduce spending to zero in a month, making API revenue more volatile than subscription SaaS. Enterprise contracts with minimum commitments provide more visibility. The ratio between these two revenue types is a quality signal.
Then there is the revenue-to-compute ratio. A SaaS company with 80% gross margins is different from an AI company with 50% margins, because compute costs consume half of every revenue dollar. Those costs may rise if NVIDIA pricing power persists, creating a structural margin ceiling that SaaS companies do not face.
Net revenue retention above 140% signals genuine enterprise compounding through usage expansion. Databricks and Anthropic both sit at roughly that level. OpenAI‘s enterprise NRR has never been disclosed, which makes it difficult to assess whether revenue growth reflects durable enterprise relationships or one-time consumer usage. When a company asks you to price a business at 42 times ARR without disclosing its most important retention metric, the appropriate response is to wait for the S-1.
The revenue-quality risks above are determined, in large part, by how an AI company chooses to charge for its products. That brings us to pricing.
SaaS pricing is built on per-seat licensing. Adding users increases revenue with near-zero marginal cost. That model produced 80%+ gross margins and made SaaS investing legible to public markets.
AI pricing introduces six archetypes: hybrid tiered subscriptions, usage-based per-token billing, credit and token pools, outcome-based pricing, seat-based plus AI add-on, and freemium. 92% of AI software companies now employ mixed models combining subscriptions with consumption fees.
The pricing model you choose is not just a billing decision. It determines your margin trajectory. A company charging flat subscriptions for variable-cost compute services will see margins deteriorate as usage scales. Outcome-based pricing, like Intercom‘s Fin AI agent at $0.99 per resolution, aligns costs with revenue but requires accurate cost-per-task measurement. GitHub Copilot initially lost roughly $20 per user monthly when charging $10 per month, as compute costs ran $20 to $80 per user. That is an AI-specific unit economics failure.
If pricing models determine margin trajectory, and revenue quality depends on consumption patterns SaaS frameworks cannot capture, there is a deeper structural force that makes both problems worse: model obsolescence.
The depreciation cycle in AI is measured in months, not years. This has no precedent in any prior technology cycle that public markets have priced.
Unlike traditional software, where products accumulate features and switching costs create retention, AI models can be superseded by a single release. Each new frontier model raises the performance baseline all competitors must meet. Morgan Stanley identifies the structural conflict: innovation increases growth potential but also raises the probability of disruption. The same forces that make AI transformative also make individual AI companies fragile.
The valuation implication is direct. A 24x revenue multiple on AI revenue that could become non-competitive within months is different from a 24x multiple on SaaS revenue with multi-year contracts and 120%+ NRR. And the gap between accounting depreciation (4 to 6 years for AI chips) and faster operational obsolescence creates hidden balance-sheet risk. The assets supporting revenue generation may become economically impaired long before they are written down.
Where model obsolescence threatens the revenue side, compute dependency threatens the cost side. Every AI company’s margin trajectory is constrained by NVIDIA’s pricing power and cloud provider terms.
NVIDIA holds roughly 65% of the data-centre AI-chip market by revenue, with some estimates placing its share of data-centre GPUs at more than 90%. Its data centre gross margins exceed 75%. Every AI company faces a single vendor whose pricing decisions flow directly to their cost structure.
Four questions illuminate the dependency. What percentage of compute spend goes to external cloud providers versus company-owned infrastructure? What are the contractual compute obligations? OpenAI’s roughly $390 billion in commitments to Microsoft and Amazon exemplify extreme dependency. Does the company have multi-cloud distribution that creates bargaining power, or is it locked into a single provider? What custom silicon or ASIC programmes exist to diversify away from NVIDIA?
Circular financing compounds compute dependency. When cloud providers invest billions into AI labs that then spend that capital back on the same provider’s compute infrastructure, the financial interdependency obscures whether revenue growth reflects genuine end-demand or vendor-financed self-dealing. GMO describes the pattern as reminiscent of the circular financing of the internet bubble. iShares and BlackRock frame it as legitimate financing for long-lived assets, comparable to Boeing providing asset-backed financing so airlines can take delivery of new planes. Both interpretations can be true. Neither makes the dependency go away.
Traditional tech moats include switching costs, network effects, scale economies, and brand. These remain relevant but are insufficient for AI companies because the core product (the model) faces commoditisation pressure from open-weight alternatives that give enterprise customers credible walk-away alternatives.
DeepSeek’s V3.2 demonstrated that open-weight models can achieve frontier-level performance at a fraction of the compute cost. Models from DeepSeek, Qwen (Alibaba’s open-weight model family), and Meta are, as Stanford HAI puts it, unavoidable in the global competitive AI landscape.
Four AI-specific moat dimensions matter. Data flywheels: does more usage generate proprietary training data that improves the model, attracting more usage? Integration depth: how embedded is the AI platform in enterprise workflows through APIs, SDKs, and partner networks? Ecosystem lock-in: does the company control distribution channels that make it the default choice regardless of marginal model performance differences? Model architecture defencibility: do proprietary training techniques create sustainable advantages, or can open-weight alternatives match performance at lower cost?
The moat durability test is simple. If a credible open-weight model matched this company’s performance at half the cost tomorrow, would enterprise customers stay? If the answer depends on integration depth and ecosystem lock-in rather than model performance alone, the moat has durability. If customers would switch based on performance-per-dollar alone, the company is a model vendor, not a platform.
Anthropic now holds roughly 40% of enterprise LLM API spend while OpenAI has dropped to 27%, down from roughly 50% in 2023. Enterprise moat is shifting from brand to integration. Anthropic‘s tri-cloud availability on AWS, GCP, and Azure means enterprise customers can use its models regardless of their existing cloud commitments, creating a distribution advantage that pure model performance cannot replicate.
Even the strongest moat is only as durable as the governance structure that controls it.
AI companies are going public with governance structures not tested at public-market scale. OpenAI’s unresolved nonprofit-to-for-profit conversion, Anthropic’s public benefit corporation (PBC) charter with the Long-Term Benefit Trust controlling board composition, and no-vote structures for strategic investors (Amazon and Google in Anthropic) create governance risk categories that standard shareholder-rights frameworks were not designed to evaluate.
As a PBC, Anthropic’s board is legally permitted to balance shareholder returns with broader stakeholder interests, including employees and the public good. The Long-Term Benefit Trust, not shareholders, selects board members based on safety mission alignment. This raises a question with little precedent: what happens when a company explicitly allowed to prioritise public benefit begins answering to quarterly earnings expectations?
Five questions illuminate the governance picture. Who controls board seats, and under what conditions can they be removed? What voting rights do common shareholders hold versus founders and strategic investors? What conversion complexity exists, and could it delay or derail the IPO? Do hyperscaler equity stakes create competing interests between infrastructure vendor, equity stakeholder, and model competitor? Does the company disclose the revenue it forgoes for safety reasons? If not, shareholders are pricing an unknown liability.
PitchBook’s AIBQ framework weights governance optionality at 20% of business quality, reflecting that governance structure is not a legal footnote but a valuation input. The first governance crisis at an AI public company will set the template for everything that follows.
With these five dimensions mapped out (revenue quality, pricing, model obsolescence, compute dependency, moat durability, and governance), the practical question becomes: what does a retail investor actually do when an AI S-1 lands?
Before the S-1 is public, almost nothing is available. The confidential filing process means no financials, no customer metrics, no governance detail. Treat any pre-S-1 valuation narratives as unverifiable.
Once the S-1 is public, five areas warrant scrutiny. Revenue concentration: how much comes from the top three to five customers? AI API revenue tends toward extreme concentration through a few large enterprises or hyperscaler partnerships. Compute cost as a percentage of revenue: the AI equivalent of cost of goods sold, structurally higher and more volatile than any SaaS COGS. R&D spend trajectory: AI companies must continue spending on frontier research to avoid obsolescence. R&D is not discretionary. Cutting it destroys the investment thesis. Governance structure: what rights do common shareholders actually have? Dual-class shares, PBC charters, benefit trusts, and nonprofit-to-for-profit conversions all affect whether public shareholders have meaningful say. Safety-versus-revenue tradeoff: companies that do not quantify the revenue they forgo for safety reasons leave shareholders pricing an unknown liability.
Retail investors should approach AI IPOs as high-risk, high-reward positions with structural information asymmetry. Position sizing should account for the probability of significant repricing events. The quarterly earnings cycle will provide rapid feedback on whether the investment thesis holds. Many institutional investors wait for the first two or three quarterly reports to expose the gap between narrative and numbers before committing capital, and retail investors have good reason to do the same.
Evaluating an AI company before it goes public is not about rejecting traditional tech evaluation. It is about layering AI-specific questions onto it. Revenue quality requires new lenses: consumption volatility, compute-to-revenue ratios, model-version dependency, and free-tier economics. Moat durability requires a test no SaaS company has faced: would customers stay if an open-weight alternative matched your performance at half the cost? Governance requires asking who controls the company when the safety mission and shareholder returns conflict.
Model obsolescence is the defining structural tension. You cannot simultaneously pursue high growth and avoid obsolescence because the same forces that drive industry progress destroy individual company advantages. Compute dependency is the dimension most directly tied to margin expansion. Every AI company’s gross margin trajectory is constrained by NVIDIA’s pricing power. Pricing architecture is itself a revenue-quality input. Whether a company charges per token, per outcome, or per seat determines whether its margins improve or degrade with scale.
The first governance crisis at an AI public company, the first quarterly earnings miss driven by model obsolescence, and the first safety incident that destroys revenue overnight will each teach public markets lessons no pre-IPO framework can fully anticipate. Investors who enter these positions with structured scepticism will be better positioned to navigate the structural forces reshaping technology investment.
The AI Business Quality (AIBQ) framework, developed by PitchBook, weights five dimensions of AI company quality: governance optionality at 20%, compute independence at 15%, and three other factors covering revenue durability, competitive positioning, and talent retention. It is not a scoring tool but a prioritisation map. Use it to ensure you are not over-indexing on traditional SaaS metrics while ignoring governance and compute dependency, the two dimensions with no SaaS precedent.
Circular financing describes the pattern where cloud providers invest billions into AI labs, which then spend that capital back on the same provider’s compute infrastructure. Amazon’s Anthropic stake and Microsoft’s OpenAI position are the canonical examples. It matters because it obscures whether revenue growth reflects genuine end-customer demand or vendor-financed self-dealing. Investors must investigate how much of an AI company’s compute spend flows back to entities that hold equity in it.
Waiting is almost always the better strategy for retail investors. AI IPOs introduce novel risk categories that public markets are likely to misprice for several quarters. The first two or three earnings reports will provide audited data on revenue concentration, compute costs, and net revenue retention that the S-1 may not fully reveal. Let the quarterly earnings cycle expose the gap between narrative and numbers before committing capital.
Structural profitability for frontier AI companies is likely measured in years, not quarters. The combination of ongoing frontier research spend, compute costs that may not decline faster than API pricing compresses, and free-tier compute burdens means the path to GAAP profitability is longer than any SaaS analogue. Investors should approach AI IPOs with a minimum three-to-five-year holding horizon and accept that quarterly losses are not a failure signal but a structural feature of the business model.
No, but they change what a durable moat looks like. Open-weight models from DeepSeek and Qwen commoditise raw model performance, which destroys moats built on model superiority alone. However, moats built on data flywheels, deep enterprise integration, and ecosystem lock-in survive commoditisation because customers stay for the platform, not the model. The correct question is not whether open-weight models destroy moats but whether the company’s moat depends on the model or the ecosystem around it.
NVIDIA’s monopoly ending would be a structural positive for AI company margins but a transitional shock for valuations. Companies with massive forward compute commitments at current pricing would face stranded-cost risk, while companies with multi-cloud distribution and custom silicon programmes would benefit fastest. The transition would also compress the competitive gap between well-funded and capital-constrained AI labs, potentially triggering a wave of model commoditisation that benefits platform companies over pure model vendors.
Core AI companies (LLM vendors, infrastructure providers) require platform-level scale and face winner-takes-most dynamics where model obsolescence risk is existential. Applied AI companies (vertical workflow tools, industry-specific AI) have lower capex requirements, more predictable revenue from enterprise contracts, and moats built on domain expertise rather than model performance. Core AI offers higher upside but higher obsolescence risk; Applied AI offers more durable economics but lower ceilings. Portfolio construction should reflect this tradeoff explicitly.
Lock-up periods, typically 90 to 180 days post-IPO, prevent pre-IPO shareholders from selling. For AI companies, these periods carry unusual risk because model obsolescence cycles can move faster than the lock-up window. A competitor could release a frontier model during the lock-up that materially damages the investment thesis while insiders remain unable to sell. Investors should factor this timing mismatch into position sizing and avoid treating lock-up expirations as automatic buying opportunities.
Three categories demand attention. First, export controls on advanced chips could restrict compute access and advantage competitors in less regulated jurisdictions. Second, copyright and training-data litigation (such as the New York Times lawsuit against OpenAI) could impose retroactive licensing costs or force model retraining. Third, antitrust action targeting hyperscaler equity stakes in AI labs could unravel the circular financing structures that currently fund compute obligations. Each represents a binary risk that traditional tech valuation frameworks do not price.
An AI ETF solves the single-company model obsolescence problem through diversification: if one lab’s model is superseded, another holding benefits. However, AI ETFs currently have limited exposure to pure-play frontier AI companies because the sector is only now going public, and many ETFs are weighted toward established tech giants with AI adjacencies rather than the companies whose IPOs this article analyses. A barbell approach of a broad AI ETF plus targeted individual positions in companies you have evaluated through the frameworks here balances diversification with conviction.
Focus on what the S-1 may have obscured. Ask about net revenue retention (if undisclosed, why not), compute cost as a percentage of revenue and its trendline, free-tier to paid conversion rates, and the dollar value of revenue forgone for safety reasons. Also probe whether revenue concentration among the top three customers has increased or decreased since the S-1 filing. These questions test whether the company’s public-market narrative is tightening or diverging from its pre-IPO audited numbers.
The platform-versus-feature test for AI companies has two parts. First, does the company control distribution, or does it depend on someone else’s platform for customer access? A company whose product lives inside another vendor’s ecosystem is a feature, not a platform. Second, if a competitor’s model matched this company’s performance at half the cost, would customers stay because of integration depth and switching costs? If the answer is no, you are investing in a feature with a model dependency, not a platform with a durable moat.
Why AI Companies Took 80 Percent of Global Venture Capital in 2026 and What Investors Need to KnowQ1 2026 saw roughly $240 billion of the $297 billion in global venture capital flow into AI companies. That is about 80% of every venture dollar deployed, to a single sector, in a single quarter. There is no direct historical parallel for concentration at this scale.
The surface-level reaction is predictable: bubble warnings, dot-com comparisons, predictions of collapse. But the explanation is more structural than a simple bubble narrative allows, and for investors, more actionable. The capital is chasing compute hardware, not narrative, and that distinction reshapes how you should think about portfolio risk as the capital event reshaping the technology landscape approaches.
Three reinforcing dynamics produced the 80% figure.
The first is compute intensity. Frontier model training runs cost hundreds of millions of dollars. Inference at scale costs more than training. Every dollar of AI revenue requires upfront infrastructure investment that traditional software companies never needed. The round sizes reflect data centre commitments, NVIDIA chip procurement, and energy contracts at a scale that rivals national grid projects. This is industrial capital intensity, not venture-style growth spending.
The second is the late-stage concentration dynamic. Late-stage rounds accounted for roughly 82% of all venture capital deployed in the quarter, a 205% year-over-year increase, with $235 billion of that coming from just 158 rounds of $100 million or more. Deal volume actually declined relative to 2022 levels. The funnel is narrowing: fewer companies receive larger cheques because the cost of competitive participation is now measured in billions, not millions.
The third is three-company dominance. OpenAI‘s $122 billion raise at an $852 billion valuation, Anthropic‘s $30 billion Series G at $380 billion, and xAI‘s $20 billion Series E accounted for 67.3% of megadeal capital across just three transactions. Four of the five largest venture rounds in history closed within the same 90-day window. These are the same two companies now racing to convert private capital into public listings, and the IPO wave that follows will determine whether those private valuations hold.
AI VC-backed firms now represent 46.5% of the total US venture capital market, roughly $2.35 trillion in aggregate value as of 31 March 2026. This is not a sector rotation. It is a structural reallocation of institutional capital toward compute-intensive businesses.
NVIDIA sits at the centre. Every AI company’s compute bill flows partially to NVIDIA, which recorded $75.2 billion in quarterly data centre revenue. As one analysis puts it, that figure is the most concrete evidence that capital is purchasing physical infrastructure rather than narrative.
This level of single-sector dominance invites an obvious question: how does it compare to the last time investors concentrated this heavily on one technology theme?
The dot-com comparison reveals how the two eras diverge.
At the dot-com peak, internet companies captured roughly 50% of venture capital. AI’s 80% is 60% higher as a share. In Q4 1999 the top five deals captured 39% of quarterly US venture capital; in Q1 2026 the top five captured 75%. The concentration is broader as a sector share and narrower in company count.
But the revenue picture is different. Dot-com companies were speculating on future revenue that often never arrived. Anthropic crossed $30 billion in annualised revenue by April 2026 on roughly 1,400% year-over-year growth, a revenue ramp that enterprise software has not previously sustained. OpenAI generates billions in API and subscription income. These are real businesses with real customers.
The capital purchases different things, too. Dot-com-era funding went to customer acquisition: marketing, sales, subsidies. AI-era funding purchases GPU clusters, data centres, and power procurement, an investment profile closer to industrial buildout than software development. Physical infrastructure carries longer capital-recovery timelines and different risk characteristics than marketing spend.
The concentration also resides in different markets. Dot-com capital was distributed through IPOs, spreading risk across public shareholders. AI’s concentration remains locked in private VC portfolios. The risk transfer to public markets has not yet occurred. The IPO wave forming in the second half of 2026 is the moment that changes.
That risk transfer is being pulled forward by the largest coordinated infrastructure investment cycle in corporate history. The hyperscaler balance sheets tell the story.
Combined capex from the largest cloud providers is projected at $660 to $690 billion in 2026, nearly double 2025’s $415 billion. Amazon leads at a projected $200 billion, a more than 50% increase from the $131 billion it spent in 2025. Alphabet doubled its guidance to $175 to $185 billion, with Google Cloud backlog surging past $460 billion. Meta raised guidance to $125 to $145 billion, and its shares fell 9.25% on the announcement. Microsoft is tracking above $120 billion with its AI business crossing a $37 billion annualised run rate, up 123% year-over-year.
The capex ratios tell their own story. Meta is tracking toward capex equal to 54% of sales, Microsoft at 47%, and Alphabet at 46%. Those ratios are normally seen in industrial utilities and regulated telecommunications companies, not software businesses. They signal multi-decade infrastructure commitments where returns accrue over 15 to 20 years, not quarterly earnings cycles.
Alphabet’s capital-structure decisions show how seriously management treats the spending commitment. The company announced an $80 billion equity raise plan: $30 billion in underwritten stock offerings, $10 billion from Berkshire Hathaway through a private placement, and $40 billion through an at-the-market programme. An ATM programme lets a company sell newly issued shares into the existing trading market at prevailing prices over time, avoiding the price impact of a large one-off issuance. Alphabet had already raised more than $66 billion in debt across multiple currencies, and the equity raise preserves credit-rating flexibility while signalling that AI infrastructure spending is structural, not cyclical.
Berkshire Hathaway’s $10 billion commitment is the institutional-validator signal. The firm had invested $4.3 billion in Alphabet in November 2025, and its stake was valued at roughly $20 billion before the new commitment. This marks a pivot from Apple concentration toward AI infrastructure, demonstrating that value-oriented capital is being pulled into the capex cycle.
NVIDIA remains the single largest beneficiary across all this spending. That raises the question every investor should be asking: if infrastructure spending is accelerating this fast, is revenue keeping up?
Sequoia Capital’s David Cahn has laid out the quantitative tension in the AI investment thesis: there is roughly a $600 billion annual gap between what hyperscalers are spending on AI infrastructure and what the AI ecosystem is generating in revenue. The gap is widening in 2026 as capex has accelerated faster than revenue projections.
Allianz Research puts the divergence at roughly 46%, which already exceeds the 32% gap observed before the 2001 telecom correction. That earlier gap preceded a sector-wide revaluation that erased value even for sound telecommunications companies. Infrastructure spending is now scaling roughly 50% faster than revenue, pushing the payback period further out each quarter.
Anthropic’s $30 billion annualised revenue on 1,400% growth demonstrates that the gap can close when product-market fit is genuine. The question is whether enough companies can achieve similar trajectories before capital markets demand returns.
Agentic AI is the mechanism most widely cited to close the gap. These autonomous systems can complete multi-step tasks (booking a full travel itinerary, resolving a customer service chain, managing a procurement workflow end to end) rather than responding to single prompts. They would command higher per-unit pricing than current API and subscription models. But meaningful enterprise adoption is estimated at 12 to 24 months away, making it the variable on which most AI monetisation forecasts depend.
For you as an investor, the gap’s practical implication is segmentation. Semiconductor companies and cloud platforms are monetising the buildout directly and carry lower gap-risk. Foundation-model companies sit on the revenue-generating side but must sustain growth rates that enterprise software has not previously sustained. Application-layer companies face the widest outcome distribution because their revenue depends on adoption timelines still being established.
All of this capital flowing into AI has a mirror image: the sectors receiving almost none of it.
When 80% of global VC floods into one sector, the remaining 19%, roughly $58 billion in Q1 2026, is spread across robotics, biotech, fintech, enterprise software, climate tech, defence tech, and manufacturing automation. That $58 billion exceeds the entire annual US venture market before 2018, but it is overshadowed by AI concentration at the top.
The capital is not zero. The distortion is in pricing. Non-AI late-stage companies are raising at 2022 valuation levels despite having 2026 traction metrics, because institutional LP preference for established AI-heavy managers starves generalist and sector-specialist funds of capital. Enterprise software companies with multi-million dollar ACVs are being valued at 3 to 5x ARR, half the multiple of horizontal AI tools with identical growth rates.
Defence tech illustrates the contrarian opportunity. Companies with $50 million plus ARR and 100% year-over-year growth are raising at valuations that would be multiples higher in consumer AI. But ITAR restrictions, security clearance requirements, and regulatory moats create barriers AI foundation models cannot replicate. Shield AI‘s $2.3 billion Q1 raise demonstrates capital is available for the right combination of technology and structural defensibility.
If you are constructing or auditing a venture portfolio in this environment, six questions clarify your true concentration exposure. First, what percentage of your fund’s NAV is tied to OpenAI, Anthropic, or xAI? Second, what valuation methodology would apply if those positions were marked down 30%? Third, what is your total exposure to the six managers who controlled 76.2% of Q1 fundraising? Fourth, what are the secondary-market discounts on your top holdings? Fifth, what is your rationale for any non-AI allocation? And sixth, what distribution timeline do you expect given that the IPO wave is the primary liquidity path?
Harvard Management Company invested in three non-AI late-stage rounds in Q1, citing “valuation compression and reduced competition.” When sophisticated LPs are deliberately underweighting the market’s most popular trade, it is worth understanding why.
That underweighting becomes particularly relevant as the IPO pipeline opens.
The concentration in private markets creates an IPO pipeline where diversification within the AI sector is structurally difficult. If three companies represent two-thirds of AI venture capital, a portfolio constructed from AI IPOs will be inherently concentrated by design, not by choice.
The underwriting dynamic shapes access. Goldman Sachs, JPMorgan Chase, and Morgan Stanley are the lead banks on both Anthropic and OpenAI IPOs. Institutional allocations will be rationed, and firms without existing relationships with these banks will face reduced access to the largest listings. In practice, the banks’ top-tier institutional clients receive the bulk of the allocation at IPO price, while smaller or less-established firms buy in the aftermarket where the discount has already narrowed.
Lockup-expiration overhang is the primary post-IPO risk. When 12 to 18 month lockup periods expire, the supply of shares from early investors and employees could pressure prices. More than 600 current and former OpenAI employees have already sold $6.6 billion in company stock through the secondary market ahead of the IPO, an early signal of selling appetite.
Secondary-market pricing provides a reality check. 2023-vintage AI names trade at roughly a 19% discount to their last primary round marks. The top 20 names account for 81.1% of all secondary trading value. These discounts signal what sophisticated pre-IPO buyers believe about private valuations. You should demand secondary-market comparables before committing to IPO allocations.
Yale’s endowment disclosed in March 2026 that it had reduced AI exposure from 35% to 22% of its venture portfolio, rotating into robotics and manufacturing automation. It is a leading signal of how sophisticated LPs are managing concentration risk through deliberate underweighting relative to market benchmarks.
Your practical response to the IPO wave begins with four steps: model lockup-expiry scenarios for any AI name you hold or plan to hold, assess secondary-market discounts against primary marks, audit your underwriting-bank relationships to understand where you sit in the allocation queue, and benchmark your AI exposure against LP reallocation signals like Yale’s rotation. For a structured approach to these assessments, see how to evaluate an AI company before it goes public.
The 80% figure reflects rational structural forces. It also carries real portfolio risk. Both assessments hold simultaneously, and investment decisions must account for both as AI’s most capitalised companies begin answering to shareholders.
It reflects a structural shift, not a quarterly blip. The underlying driver (compute costs measured in hundreds of millions per training run) is not reversing, and the narrowing funnel of late-stage deals (90.6 percent of AI VC in Q1 2026 concentrated in 10 percent of deal volume) has been building since 2024. The $660 to $690 billion in projected hyperscaler capex for 2026 ensures capital intensity persists through at least 2027, making sector-level concentration durable even if the three largest recipients rotate.
They consolidate, get acquired, or become stranded assets with technology that works but cannot compete at scale. The market is bifurcating between the three or four capital-rich leaders and everyone else. Mid-tier AI labs with credible technology but insufficient capital are becoming acquisition targets for hyperscalers seeking talent and intellectual property, while application-layer companies that build on existing models require far less capital and can operate outside the megadeal dynamic entirely.
Direct exposure to the largest private AI companies is unavailable to retail investors, but indirect exposure exists through several channels: NVIDIA (which captures a portion of every AI compute dollar), cloud hyperscalers (Microsoft, Alphabet, Amazon) whose infrastructure underpins AI workloads, and semiconductor ETFs. Secondary market platforms like Forge and EquityZen occasionally offer pre-IPO shares, but minimums are typically high and the 19 percent discount on 2023 vintage AI names signals the risk retail buyers should price in.
The answer is segmented rather than binary. Semiconductor and infrastructure companies are priced against observable revenue (NVIDIA’s $75.2 billion in quarterly data centre revenue is concrete), while frontier model companies embed assumptions about sustained 1,000-plus percent revenue growth and agentic AI monetisation that remain unproven. The secondary market’s 19 percent discount on 2023 vintage AI names suggests sophisticated pre-IPO buyers believe private marks exceed fair value. Both genuine transformation and overvalued assets coexist, exactly as they did during the dot-com buildout.
They are among the largest limited partners fuelling the megadeal dynamic. Sovereign wealth funds, particularly from the UAE, Saudi Arabia, and Singapore, have allocated tens of billions to the same six VC managers that controlled 76.2 percent of Q1 2026 fundraising. Unlike traditional institutional LPs, sovereign funds often prioritise strategic technology access and economic diversification over pure financial returns, meaning their capital is less valuation-sensitive and more structurally committed to AI regardless of near-term mark-to-market fluctuations.
It creates an uncomfortable dependency: most AI safety research is funded by the same companies whose commercial incentives may conflict with its findings. OpenAI, Anthropic, and Google DeepMind all maintain internal safety teams, but independent safety research relies on grants from these organisations or from philanthropic arms of the same institutional capital pool. The concentration means there is no adequately funded safety ecosystem operating outside the commercial AI stack, concentrating both the technology and its governance in the same entities.
Overwhelmingly on compute infrastructure. The primary allocation goes to GPU procurement and data centre capacity commitments, followed by power procurement agreements and networking infrastructure. Personnel costs for research scientists and engineers represent a meaningful but secondary expense. Marketing and customer acquisition are minimal line items compared to software-era startups. This makes the spending profile resemble industrial infrastructure investment more than traditional venture-backed company operations, with multi-year capital recovery timelines to match.
A poor reception for the largest AI IPOs would compress valuations for the entire technology IPO pipeline and likely freeze non-AI listings as institutional investors absorb losses. Investment banks would delay or reprice smaller offerings. The lockup expiration overhang (12 to 18 months after listing) would amplify selling pressure as early investors rush to capture remaining gains. The 2001 precedent suggests the correction would be sector-wide even for fundamentally sound companies, though the damage would concentrate in AI names trading at the widest gap between private marks and public-market comparables.
No, but the distortion is severe. Approximately $58 billion deployed to non-AI startups in Q1 2026 represents 19 percent of global VC, spread across biotechnology, defence technology, fintech, enterprise software, robotics, and climate technology. The capital is not zero but it is significantly compressed relative to historical norms. The real constraint is not abandonment but allocation: six fund managers control 76.2 percent of VC fundraising, and they are overwhelmingly deploying into AI, starving sector-specialist funds of limited partner capital regardless of the quality of non-AI opportunities.
They are pivoting toward the application layer, where capital requirements are lower, or toward sectors with regulatory moats that AI cannot easily penetrate. Defence technology (ITAR restrictions and security clearances), biotechnology (FDA pathways and clinical data exclusivity), and manufacturing automation (physical-world complexity) are attracting generalist VC attention. Some are also building secondary-market strategies, acquiring pre-IPO AI shares at the 19 percent discount rather than competing for primary allocations where the six dominant managers control access.
Yes, and the mechanism is already visible. If the capex-to-revenue gap (estimated at $600 billion annually, with infrastructure spending growing roughly 50 percent faster than ecosystem revenue) fails to narrow, capital markets will enforce discipline through lower IPO valuations, higher debt costs for hyperscalers, and LP reallocations away from AI-heavy managers. The Yale Endowment’s rotation from 35 percent to 22 percent AI exposure is an early signal. The correction would be painful for foundation-model companies but potentially beneficial for application-layer startups that can ride cheaper infrastructure without bearing the buildout cost.
It creates a severe funding disadvantage. European, Asian, and Australian AI startups compete for a shrinking pool of non-US venture capital in a market where 80 percent of global VC and effectively all megadeal capacity is concentrated in American companies. The practical response has been relocation: founders establish Delaware corporations and seek US-based lead investors, effectively ceding domestic ecosystem development. Sovereign AI initiatives in France (Mistral) and the UAE represent attempts to counteract this dynamic, but their capital pools are orders of magnitude smaller than the megadeal scale now required.
Can AI Governance Survive the Public Markets? How Anthropic and OpenAI’s Unconventional Structures Will Test Wall StreetI’ll construct the complete article with all pillar/cluster links added and output it directly.
Sometime in the next twelve months, an Anthropic director will log onto an earnings call and explain to a room full of analysts why quarterly revenue missed expectations because the board chose safety over deployment. The question will be some version of “walk us through the tradeoff” and the answer will have no precedent. That conversation has never happened at public-company scale. It will, and nobody knows what happens next.
Over 190 companies are waiting to go public in 2026, but only two raise governance questions the market has no framework for — and the confidential filing that kicked off the race is already behind them. Anthropic is a Delaware Public Benefit Corporation with a Long-Term Benefit Trust that can override shareholder preferences. OpenAI uses a capped-profit model where returns are contractually limited and a nonprofit board holds ultimate control. Neither fits the governance template that public markets are built to price.
Traditional tech IPOs (Google, Meta, Snap) used dual-class shares to preserve founder control over commercial strategy. AI governance adds a dimension that dual-class structures never addressed: legally enforceable mission constraints on profit. Here is what each structure looks like, how they compare, and how to think about the gap between what listing standards assume and what these companies have hard-coded into their charters — the governance puzzle at the centre of the public-market experiment unfolding across the AI industry.
Anthropic is organised as a Delaware Public Benefit Corporation, which means its directors have a statutory duty to balance shareholder returns against the company’s stated public benefit: the safe development of artificial general intelligence. They are not required to maximise shareholder value. Under Delaware law (DGCL §§361–368), this duty is enforceable. Shareholders owning 2% or $2 million in stock can bring derivative suits alleging the board failed to balance pecuniary and benefit interests.
The Long-Term Benefit Trust adds another layer. Five trustees can recruit and remove board members if they determine the company is deviating from its safety mission. Combined with a staggered board and supermajority voting requirements for structural changes, Anthropic’s governance creates a set of checks that no public company has.
If you are accustomed to annual director elections, majority voting, and shareholder proposal rights, you will find this structure more restrictive than most tech IPOs of the last decade. The question is whether you price that restriction as a protective mechanism or a liability.
The two structures solve the same problem differently. Anthropic’s approach is directional: the board must balance profit against safety, with the benefit trust providing enforcement. OpenAI’s approach is quantitative: investor returns are capped at a multiple of investment (reportedly 100x). Excess profits flow to the nonprofit parent.
Anthropic’s PBC offers standard equity returns but introduces governance risk. The board can legally refuse a lucrative deployment on safety grounds, and the Long-Term Benefit Trust can replace directors who prioritise profit over mission. OpenAI’s capped-profit structure offers governance certainty at the cost of a hard upside ceiling. Its nonprofit board has ultimate authority over the for-profit subsidiary regardless of investor preferences, and the operating agreement states explicitly that the mission “takes precedence over any obligation to generate a profit.”
SpaceX provides the closest precedent. Its dual-class structure concentrates 85% of voting power with Elon Musk, making it a “controlled company” that the market has accepted with a governance discount. But SpaceX’s mission (colonise Mars) carries no tradeoff against quarterly revenue. AI governance introduces a safety-versus-revenue tension that SpaceX never had to navigate.
Neither structure is clearly “more investor-friendly.” They represent different risk profiles. Anthropic offers uncapped returns with governance risk; OpenAI offers governance certainty with capped returns. The market has no precedent for pricing either one.
NYSE and Nasdaq listing rules require majority-independent boards, independent audit committees, and shareholder approval for equity compensation plans. None address what happens when a company’s charter requires the board to prioritise something other than shareholder returns.
The gap is visible in the proxy advisor vacuum. ISS and Glass Lewis have voting guidelines for dual-class structures and board independence, but neither has published a framework for evaluating PBC boards or capped-profit governance. If you rely on these firms for voting recommendations, you are left without a standardised assessment tool. 67% of US investors evaluate AI issues on a case-by-case basis, and 29% have no benchmarks or voting policies for AI at all.
54% of S&P 100 companies now disclose board-level AI oversight in their proxy statements. But that standard addresses how boards monitor AI risk, not how AI companies are governed. The governance question is upstream of the oversight question. The SEC Investor Advisory Committee recommended in late 2025 that issuers disclose board oversight mechanisms for AI, but the focus is on risk disclosure, not structural governance design. The gap between what listing standards assume and what AI governance requires is wide. What happens when that gap meets quarterly earnings pressure?
The pressure will arrive through multiple channels at once. Activist investors will demand strategy changes when safety decisions reduce revenue. Proxy advisory firms will issue voting recommendations against directors who prioritise mission over returns. Analysts will downgrade following quarters driven by mission-constrained decisions.
Standard public-company governance assumes the board’s fiduciary duty runs to shareholders first. A PBC board must articulate why balancing against the public benefit serves long-term shareholder value, and no company has had to deliver that narrative at public-company scale. The closest benchmark is Once Upon a Farm, a small PBC whose S-1 risk factors (discussed in detail below) explicitly address the tension between public benefit duties and shareholder value. Once Upon a Farm is far smaller than any AI lab, but its S-1 remains the only PBC filing that grapples explicitly with the fiduciary tension, and the language it uses is the disclosure floor, not the ceiling.
Quarterly disclosure becomes a stress test. Safety incidents, compute allocation decisions, and revenue concentration will be visible in ways they never were as a private company. The governance structure determines who controls the narrative around those disclosures. If the structure gives you enough information to price what you are buying, the market can accept a governance discount the way it accepted dual-class discounts for Google and Meta. If it does not, the discount becomes a crisis.
You need to assess AI governance across three dimensions, and none of them appear on a standard governance checklist.
First, structural durability. Can the governance mechanism survive an activist campaign? Dual-class shares, staggered boards, and supermajority provisions are the standard toolkit for activist-proofing. AI companies layer mission constraints on top. You need to evaluate whether these layers create barriers to shareholder influence that go beyond what dual-class structures have delivered in the past.
Second, the safety-versus-revenue tradeoff. Model scenarios where a lucrative deployment is blocked by governance. A defence contract refused on ethical grounds. A commercial model release delayed for safety testing. What revenue could the company forgo, and how does the market price that risk? Anthropic’s own forecast puts the probability of a $100 million-plus defence contract by May 2027 at just 3%, reflecting governance constraints interacting directly with revenue opportunity.
Third, transparency. The governance structure determines who controls the narrative when safety decisions become quarterly disclosures. You need enough information rights to make informed decisions. Dual-class shares address founder control, not mission constraints. The gap between those two purposes is where governance risk lives. Here is how to spot that risk in an S-1 filing — and the capital event reshaping the AI industry makes this skill more urgent than any prior tech cycle has.
Five red flags separate mission-preserving governance from accountability-evading opacity.
First, information rights. If the governance structure gives the board or a trust authority to block commercial decisions without requiring disclosure of the reasoning, you cannot price the risk. Second, director removal barriers. If you cannot remove directors who consistently prioritise mission over returns, the structure has no accountability mechanism. SpaceX’s governance shows what extreme accountability failure looks like: with 85% voting control, removing the CEO requires a majority of shares he alone controls.
Third, undefined benefit scope. If the public benefit purpose is stated in broad, aspirational terms without specific, measurable commitments, the board has unlimited discretion to justify any decision as mission-aligned. Fourth, no sunset provisions. If mission-constraint mechanisms are permanent rather than subject to periodic shareholder reapproval, there is no market check on whether the constraints still serve their purpose.
Fifth, S-1 risk factor language. Once Upon a Farm’s S-1 sets the disclosure benchmark: it states its directors are “not merely permitted, but obligated, to consider our specific public benefit” and that “[our] duty to balance a variety of interests may result in actions that do not maximise stockholder value.” Any AI lab S-1 with weaker disclosure than a baby food company leaves you with less information to price governance risk.
Whether AI governance survives the public markets depends on whether each company’s transparency and accountability mechanisms give you enough information to price the mission-vs-revenue tradeoff. The structures will survive if they produce enough disclosure for the market to do its job. They will break if they use mission constraints as a shield against accountability.
As the SpaceX precedent shows, the market accepts governance discounts, but only when it can see what it is pricing. The proxy advisor vacuum means you must build your own evaluation framework, and the three dimensions (durability, tradeoff, transparency) are the starting point — but governance is just one axis of second-order effects that ripple far beyond governance.
Whether AI governance survives comes down to an informational question: can you see enough to price what you are buying? How governance risk factors into a broader evaluation framework is the question every AI IPO investor will need to answer.
Yes. The Long-Term Benefit Trust has the authority to recruit and remove board members if it determines the company is deviating from its safety mission. In a hostile takeover scenario, the trust could replace directors who support the acquisition with trustees who oppose it, effectively blocking any bid it deems inconsistent with Anthropic’s public benefit purpose. Combined with the staggered board and supermajority voting thresholds, the trust creates a takeover defence that is structurally stronger than any poison pill a traditional public company could deploy.
It depends entirely on what you compare it to. A 100x return on a late-stage investment implies OpenAI must reach roughly a $10 trillion valuation for its most recent investors to hit the cap, which makes the ceiling appear theoretical rather than practical. But as a structural feature, the cap means investors can never capture tail-risk upside of the kind that made early Google and Meta shareholders billionaires many times over. The cap is generous relative to most venture outcomes and restrictive relative to what uncapped equity in a world-changing technology company has historically delivered.
Delaware courts apply a two-part test under DGCL section 365. First, the court examines whether the board’s decision was rationally related to the corporation’s stated public benefit. Second, it assesses whether the decision was informed and disinterested, applying the business judgment rule. So far, no Delaware court has adjudicated a PBC balancing claim at trial, which means every AI governance dispute would be litigating on uncharted ground. Directors get significant deference, but the rational-relationship threshold is not a blank cheque.
Not easily. Converting from a Delaware Public Benefit Corporation to a standard corporation requires approval from holders of at least two-thirds of outstanding shares, the same supermajority threshold that protects the structure in the first place. The Long-Term Benefit Trust’s board influence makes assembling that two-thirds majority extraordinarily difficult. More fundamentally, the trust was designed precisely to prevent shareholder pressure from unwinding the mission constraints, and it would almost certainly mobilise against any conversion effort.
Not at Anthropic’s scale. Small PBCs have been acquired, but these were friendly transactions where the acquirer committed to maintaining the benefit purpose post-closing. No hostile takeover of a PBC has been tested in Delaware courts, and no PBC with a trust-based governance override like Anthropic’s has ever faced an unsolicited bid. The absence of precedent is itself a governance risk: nobody knows whether a Delaware court would prioritise shareholder value maximisation or public benefit preservation in a contested transaction.
OpenAI would need to restructure, and doing so mid-IPO would be a governance crisis. The capped-profit subsidiary’s contractual terms are embedded in its operating agreement, and modifying them would require consent from existing investors who accepted the cap as part of their bargain. If OpenAI hit the cap and still needed capital, it could theoretically raise debt, create a new uncapped vehicle, or negotiate a waiver from investors, but each path introduces legal complexity and potential litigation from parties who prefer the original terms.
They do not recognise the structure at all. European corporate law does not have an equivalent to Delaware’s PBC, and exchanges in London, Amsterdam, and Frankfurt have no listing rules addressing mission-constrained governance. An AI lab listing in Europe would either need to incorporate locally under a structure that lacks the same statutory protection or list as a foreign private issuer while explaining to European institutional investors why their governance expectations about shareholder primacy do not apply. The governance gap is wider in Europe than in the US.
Yes, and it appears to be moving in that direction. The SEC Investor Advisory Committee recommended in late 2025 that issuers disclose board oversight mechanisms for AI, but the current focus is on how traditional companies monitor AI risk, not on the structural governance of AI companies themselves. The next logical step is S-1 disclosure requirements that force AI labs to quantify the safety-versus-revenue tradeoff in specific deployment scenarios. That would transform governance from a theoretical concern into a set of priced, disclosable risks.
Yes, but none have announced IPO plans at comparable scale. Safe Superintelligence Inc. and several European labs use nonprofit or charitable structures with commercial subsidiaries, while Google DeepMind operates inside Alphabet’s standard corporate governance with an internal ethics review framework. The critical distinction is enforceability: a voluntary ethics board can be dissolved by management, while Anthropic’s benefit trust is a structural constraint that cannot be removed without a supermajority shareholder vote. Market participants are watching whether any third model emerges during the current pre-IPO pipeline.
Very little. OpenAI’s capped-profit subsidiary is governed by a nonprofit board whose fiduciary duty runs to the nonprofit’s charitable mission, not to the subsidiary’s investors. Investors accepted this arrangement contractually when they bought in, and the operating agreement explicitly subordinates their interests to the nonprofit’s determination of what serves humanity. The only available challenge would be to argue the board acted outside its authority under the operating agreement itself, a narrow and difficult claim that offers no remedy for ordinary business disagreements.
Inside the IPO Race Between Anthropic and OpenAI: Valuations, Business Models, and What It Means for InvestorsAnthropic confidentially filed its draft S-1 with the SEC on 1 June 2026. OpenAI followed exactly one week later, on 7 June. The headlines treated both as “AI IPOs incoming,” and the market reaction was predictable: combined private valuations approaching $2 trillion make this the biggest capital event in tech history, and the first time public-market investors can buy pure-play AI companies directly.
Both filings are confidential under the JOBS Act. The Rule 135 announcements that hit the wires in early June tell investors almost nothing: no financials, no risk factors, no business description. Neither company will trade for months. What matters now is why the two largest frontier AI labs both ran out of alternatives at the same moment, and what that moment means for everyone watching.
A confidential S-1 is a draft registration statement submitted to the SEC under the JOBS Act. It lets Emerging Growth Companies, defined as those with revenue under $1.235 billion, begin regulatory review without publicly disclosing their financials. Both Anthropic and OpenAI are expected to qualify, which means only regulators and advisers see the sensitive information during the initial review phase.
The process moves through several stages. The company submits its draft, the SEC reviews it and issues comment letters (typically within 30 days for the initial round), and the company responds with amendments. This back-and-forth continues for multiple rounds, usually three to six months for companies of this scale. Only after the SEC is satisfied does the S-1 “flip” to public on EDGAR, followed by a 15-day cooling-off period, the roadshow, book-building, pricing, and finally first-day trading.
SpaceX’s timeline offers a useful benchmark. The company’s confidential-to-public S-1 took roughly 11 weeks, which suggests a possible public-filing window in late August or September for Anthropic and OpenAI if their reviews move at a similar pace. But analysts point to specific complications that are likely to slow things down. The biggest is Anthropic’s gross-vs-net revenue accounting. In practice, this means Anthropic books the full amount its enterprise customers spend on cloud compute as its own revenue, even when that money passes straight through to AWS or Google Cloud. The SEC may decide only Anthropic’s margin on those deals counts as revenue, which could reduce the reported run-rate by 20 to 40 percent (more on this below). Add in the Department of War litigation requiring material risk-factor disclosure, AI safety risk factors as novel disclosure territory with no settled template, and customer concentration treatment, and you are looking at four to six months rather than eleven weeks.
The Rule 135 announcement tells the market only that a filing exists. You cannot buy shares, review financials, or assess risk factors yet. Both companies’ announcements explicitly state they “do not constitute an offer to sell or the solicitation of an offer to buy any securities.” Once the S-1s become public, investors will need a framework to evaluate what they are actually buying.
If the SEC process means neither company lists for months, the natural question is why file now at all. Four structural forces converged in mid-2026 to leave both frontier labs with no good alternative to the public markets.
The first is capital ceilings. Anthropic’s $65 billion Series H at a $965 billion post-money valuation and OpenAI’s $40 billion raise pushed private valuations to levels where venture and growth-equity limited partners are structurally maxed out. Mutual funds like Capital Group, Fidelity, and T. Rowe Price participated in Anthropic’s Series H not as long-term private investors but as a pre-IPO price anchor. Mutual funds do not pay $965 billion for long-term private positions. They pay to be first in line on the public listing. AI captured approximately 80% of global venture capital in Q1 2026, saturating LP allocations and making further private rounds impossible at the scale both companies now require.
The second is competitive timing. Being first to price establishes the benchmark valuation multiple for every AI IPO that follows. The lab that completes SEC review first can define the narrative for AI companies as a public-market asset class. Analysts describe it as “a race, not a parade, toward dual listings.”
The third is the market window. Equity markets in mid-2026 are receptive to high-growth tech issuance, but that window may not stay open. Beyond its value as a timeline benchmark, SpaceX also creates a competitive deadline. Neither Anthropic nor OpenAI wants to be the last mega-IPO in the wave when institutional demand may be exhausted.
The fourth is employee liquidity. Both companies have large workforces compensated substantially in equity. Without a public listing or tender offer, that equity is illiquid. IPOs unlock employee wealth and serve as a retention mechanism, particularly important when competing for AI talent. Post-IPO, employees face a 180-day lock-up before they can freely sell.
As the AI Funding Tracker notes, a clean first-day pop for SpaceX opens the window for every name that follows. A stumble resets 2026 expectations entirely.
The same structural pressures produced the same timing response for both companies. Beyond that, they are making different bets about where AI value accrues, and those bets will appeal to different types of investors.
Anthropic is a Public Benefit Corporation. Its board is legally required to balance shareholder returns with a stated public benefit: developing AI systems that are safe, interpretable, and aligned with human values. This affects fiduciary duties in ways no large-scale tech IPO has tested before. The board can reject acquisition offers that compromise the stated benefit even at a premium price, and shareholders have limited legal standing to challenge decisions made in pursuit of that benefit.
OpenAI operates under a capped-profit structure with the OpenAI Foundation, a nonprofit, retaining ultimate control. Microsoft owns approximately 27%, and its revenue-share agreement reportedly grants Microsoft 20% of revenue through 2030. Investors face return limits, with profits above a negotiated cap flowing back to the foundation. The S-1 will need to explain this structure in detail to public-market investors who are not used to capped upside.
The revenue profiles are equally divergent. Anthropic’s run-rate crossed $47 billion as of mid-May 2026, driven heavily by Claude Code. That terminal tool alone grew from $500 million to $8 billion ARR between September 2025 and May 2026. Anthropic holds 42 to 54% of enterprise coding market share versus OpenAI’s 21%, and coding is 51% of all generative AI enterprise usage. The revenue is concentrated, sticky, and enterprise-heavy. But the gross-vs-net accounting question means headline ARR could be reduced by 20 to 40% if the SEC forces a restatement to net revenue.
OpenAI reports roughly $2 billion in monthly revenue, about a $24 billion annualised run-rate, with 900 million weekly active ChatGPT users and 50 million paying subscribers. Enterprise revenue makes up more than 40% of that total and is on track to reach parity with consumer revenue by end of 2026. The diversification is broader, but consumer-revenue volatility is a reality. OpenAI lost nearly $9 billion in 2025 and does not expect break-even until 2029 or 2030.
Valuation-wise, Anthropic’s Series H implies roughly 21 times run-rate ARR, subject to the accounting question. OpenAI’s most recent round valued it at $852 billion. Both are well above traditional tech IPO multiples (the Magnificent Seven trade at 3 to 13 times forward sales). The comparison is not perfectly aligned (Anthropic’s multiple uses current run-rate rather than forward estimates), but the gap is wide enough that the direction is clear. The valuations have drawn dot-com-era comparisons from academics, though the revenue growth at both companies is real in a way that late-nineties internet startups rarely delivered.
What matters is that the two IPOs represent two incompatible theories of AI value. Anthropic is betting that enterprise agentic AI, sold through APIs and terminal tools into coding workflows, is where durable moats form. OpenAI is betting that platform-scale distribution, with 900 million users and a brand that has become synonymous with AI, creates competitive advantages that compound over time. The public market will have to choose which theory it believes.
The confidential filing process means every investor is operating on incomplete information. The real analysis begins when the S-1s flip public on EDGAR, likely between late August and December 2026. The one-week gap in June 2026 marked the moment structural pressure overwhelmed private-market capacity for both companies simultaneously. When the S-1s flip public, the investor community will see two distinct bets: a safety-governed enterprise API company whose revenue maths turns on cloud-reseller pass-through accounting, and a consumer-platform company whose subscriber conversion rate will be tested by quarterly earnings. Whichever company prices first will define the valuation multiples public-market investors use to judge every AI listing that follows. The race is not just about which stock trades higher. It is about which theory of AI value the public market validates.
Retail investors can purchase shares through any standard brokerage account on the first day of trading, the same way they would buy any newly listed stock. Most major brokers (CommBank, Stake, Interactive Brokers) will make shares available at the opening auction. Some platforms also offer conditional access to IPO allocations, though retail allocations for IPOs of this size are typically small and heavily oversubscribed.
The most likely window for public S-1 filings is late August to September 2026, assuming the SEC review follows a pace similar to SpaceX’s 11-week confidential-to-public timeline. From the public S-1 filing, a further 15-day cooling-off period plus a 1-2 week roadshow means first-day trading would fall between late September and mid-October 2026. However, complex reviews involving novel governance structures could extend this timeline by several months.
Neither company has disclosed GAAP profitability figures, and their confidential S-1s remain private, so the market does not yet know. What is clear is that frontier AI training runs cost billions of dollars, and neither company operates with the cost discipline of a mature public company. The gross-vs-net revenue accounting question at Anthropic further complicates any profitability estimate. The public S-1 flip will reveal whether either company is approaching break-even or burning cash at venture scale.
A Public Benefit Corporation like Anthropic legally requires its board to balance shareholder returns against a stated public benefit (AI safety), rather than prioritising shareholder value alone. This affects fiduciary duties in two practical ways: the board can reject acquisition offers that compromise the stated benefit even at a premium price, and shareholders have limited legal standing to challenge decisions made in pursuit of the public benefit. No large-scale tech IPO has tested this structure at public-market scale before.
OpenAI’s capped-profit model limits the return that investors can earn, with profits above a negotiated cap flowing back to the non-profit entity that governs OpenAI. The specific cap multiples have not been publicly disclosed for the most recent funding rounds. This structure means OpenAI is not a conventional for-profit company: Microsoft’s approximately 27 percent stake is subject to return limits, and the non-profit board retains ultimate control regardless of ownership percentage. The S-1 will need to explain this structure in detail for public-market investors.
Existing private investors (venture funds, mutual funds, strategic investors) will typically have their shares converted to common stock at the IPO. Standard IPO lock-up agreements prevent these shareholders from selling for 180 days after the listing, protecting the stock from immediate insider selling pressure. Some large pre-IPO investors like Capital Group and Fidelity may have negotiated different lock-up terms. When the lock-up expires, the market will watch closely for any large block sales.
If Anthropic books cloud-computing costs that pass through to enterprise customers as gross revenue, its headline ARR of $47 billion could overstate the company’s true economic scale by 20 to 40 percent. The SEC will scrutinise this treatment during review, and a restatement to net revenue would reduce the reported run-rate and raise Anthropic’s effective valuation multiple (currently approximately 21x ARR). This is not an abstract accounting debate; it directly affects how investors value the company against peers.
Yes. Both filings remain confidential, and the SEC can issue multiple rounds of comment letters requiring material amendments. If the SEC raises significant issues around governance structure (particularly Anthropic’s PBC model or OpenAI’s capped-profit structure), revenue recognition, or risk disclosure, the process could extend into 2027. External factors also matter: a market downturn, a poorly received SpaceX listing, or a macroeconomic shock could cause either company to delay pricing or withdraw entirely.
No. The simultaneous June 2026 filings are the result of structural convergence, not coincidence. Both companies hit private-capital ceilings at roughly the same time, both face employee equity liquidity pressure, and both recognise that the first to price sets the AI valuation benchmark for public markets. The one-week gap between filings likely reflects competitive intelligence rather than accident: each company wanted to establish its own narrative without appearing to react to the other.
The final IPO price per share is set during the book-building process in the final days before listing, and it depends on the total offering size divided by the number of shares offered. Neither company has disclosed these figures. For context, a company valued at $100 billion might price shares in the $30 to $50 range, but companies can also execute reverse stock splits to achieve a desired nominal price. No reliable estimate exists until the public S-1s disclose the proposed offering amount and share count.
The Botsitter Economy: When Managing AI Agents Becomes the New Full-Time JobBack in May and June 2026, Glean surveyed 6,000 digital workers and surfaced a statistic that reframes the whole enterprise AI conversation: the average knowledge worker spends 6.4 hours per week not using AI, but managing it. Feeding context. Correcting mistakes. Cleaning up errors. Switching between disconnected tools that each need the same organisational information fed to them separately.
The researchers called it “botsitting,” and the companion behaviour they uncovered, “botshitting” (shipping AI-generated work you cannot explain or defend), is admitted to by 69% of surveyed workers. The full Work AI Index 2026 report was published by the Work AI Institute, commissioned by Glean, and co-authored by researchers from Stanford, UC Berkeley, UC Santa Barbara, Emory, and several other universities. It is the first large-scale study to attach hours, percentages, and terminology to an experience workers have been living with but could not name.
These findings describe something more structural than a temporary adoption friction. They describe the emergence of a botsitter economy: a workplace in which supervising AI agents has become a material job function rather than an occasional task. This pillar page maps the three dimensions of that economy: the empirical reality, the organisational paradox it creates, and the multi-pronged response taking shape across psychology, regulation, and technical architecture. Each cluster article takes one dimension deep. Start here for the full picture.
Botsitting and Botshitting: The Empirical Picture from the Work AI Index 2026 The data behind the 6.4-hour botsitting tax, the 36% AI session failure rate, and the 69% of workers who admit to shipping AI work they cannot explain. This is where the numbers live.
The AI Productivity Paradox: Why Individual Gains Do Not Scale to Organisational Performance Why 87% of workers use AI and 73% report personal gains, yet only 13% see organisational improvement. The gap that makes the botsitter economy an investment problem, not just an academic curiosity.
Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis How moral disengagement, the EU AI Act, and enterprise graphs are reshaping AI oversight. The three dimensions of the response are converging faster than most organisations realise.
The botsitter economy describes a workplace where the labour of supervising AI agents has become a substantial, ongoing job function rather than an occasional cost of using new software. It entered the conversation in mid-2026 because Glean’s Work AI Index quantified it for the first time at scale, measuring 6.4 hours per week across 6,000 surveyed workers. That number made visible a category of labour that had been hiding inside “AI adoption.”
Before this survey, workers and managers had a diffuse sense that AI required more hand-holding than expected, but no single study had attached hours, percentages, and terminology to the experience. The Work AI Index 2026 changed that. Paul Leonardi, Duca Family professor of technology management at UC Santa Barbara and one of the co-authors, said it plainly: “Most people don’t realize the amount of time that they’re spending working on the tools to get the time savings that they’re professing.” Rebecca Hinds, head of Glean’s Work AI Institute, called it “a vicious cycle that feeds itself” and emphasised the “massive, massive human labor that’s at the core of this.” The LA Times coverage and CIO.com’s analysis both unpack what this survey revealed.
The 6.4-hour figure represents 37% of workers’ total AI time. It is not a rounding error or an onboarding friction that will disappear as tools improve. The 36% AI session failure rate means oversight is mechanically required by current AI reliability, not optional. More than a third of AI sessions fail outright, requiring a full restart or substantial rework. And the 69% botshitting rate means the alternative to botsitting is not smooth autonomous operation. It is unverified output entering organisational workflows at scale.
The botsitter economy has three dimensions that any organisation adopting AI at scale must understand. The empirical reality: what the data says about how much oversight AI actually demands. The organisational paradox: why individual productivity gains fail to aggregate into organisational performance. And the multi-pronged response: how psychology, regulation, and technical architecture are converging to make AI oversight manageable. Each dimension gets its own cluster article. This pillar page gives you the overview so you know which questions matter and where to go for the answers.
Go deeper: Botsitting and Botshitting: The Empirical Picture from the Work AI Index 2026 breaks down the data in full, including how top AI achievers’ habits differ from low achievers’.
Botsitting is the labour of keeping AI agents usable: feeding them context about your organisation, team, and goals; correcting their mistakes; cleaning up their output; and switching between the disconnected tools different agents require. Botshitting is its negative endpoint: shipping AI-generated work you have not verified, do not fully understand, or cannot defend if questioned. They are paired concepts because botshitting is what happens when the cognitive burden of botsitting exceeds a worker’s capacity or motivation to maintain it.
Botsitting encompasses three core activities identified in the Work AI Index 2026. Context-feeding is the most exhausting of the three. Workers must manually provide AI agents with organisational context they lack, repeatedly, across multiple disconnected tools. Error correction means identifying and fixing AI mistakes before outputs reach colleagues or customers. Output cleanup involves reworking AI-generated drafts into usable final form. Unlike traditional software configuration, which happens once at setup, botsitting is continuous. Every new task, every new context switch, every agent update can require fresh oversight labour.
The report itself describes it as a “thick, mostly invisible layer of human labor holding the whole thing together.” Rebecca Hinds pointed out something most organisations are not measuring: “It’s exhausting for workers to not only do this, but to have the work be unrecognized, often unrewarded and unacknowledged within the organization.”
Botshitting is the behaviour that emerges when botsitting labour is abandoned. Workers ship AI output they have not verified because the cognitive cost of verification exceeds their available attention or motivation. The 69% figure from the Work AI Index 2026 makes this a majority behaviour, not an edge case. Andrew Pope, a technology strategist, put it this way: “The smarter the tool, the sloppier we become. The apparent capability, helpfulness and humanness of our AI tools make us trust them more than we really should, becoming less critical of their outputs.”
The distinction between the two concepts matters. Botsitting is oversight labour. Botshitting is the failure mode that occurs when oversight is abandoned. They define two different conversations: one about reducing botsitting overhead (making oversight more efficient), and one about preventing botshitting (making oversight culturally and procedurally non-optional).
The accountability dimension creates a structural problem at organisational scale. The 40% of workers who blame AI for failures, versus only 29% who admit fault, reveals an accountability vacuum at the heart of the botsitter economy. Heavy AI users are 3.4 times more likely than light users to blame the tool when something goes wrong. When the majority of workers both ship unverified AI output and attribute failures to the AI rather than to their own supervision choices, organisations lose the ability to trace decisions, assess quality, or assign responsibility. Botsitting and botshitting are connected because the psychological pathway from one to the other is predictable: a work environment that demands AI usage, provides tools whose output looks plausible even when wrong, and offers no clear accountability framework for AI-mediated work, will produce exactly the behaviour the Work AI Index documents — a dynamic the three-pronged response article explores in depth.
Go deeper: Botsitting and Botshitting: The Empirical Picture from the Work AI Index 2026 has the full data breakdown, including how top AI achievers’ habits differ from low achievers’.
The Work AI Index 2026 found that workers save roughly 11 hours per week through AI use, but spend 6.4 of those hours on botsitting. The net gain is approximately 4.6 hours per week. That reframes AI from a productivity miracle to a narrow net improvement: real, solid, worth having, but far smaller than the headline “11 hours saved” suggests.
The arithmetic is straightforward: roughly 11 hours saved, minus roughly 6.4 hours of botsitting overhead, leaves about 4.6 hours of net productivity gain per week. The 6.4 hours represent 37% of total AI time. For every hour a worker spends getting useful output from AI, they spend roughly another hour making it usable. ITBrief’s coverage of the survey data put this net calculation front and centre.
Now here is where it gets uncomfortable for anyone who has signed off on an enterprise AI budget. Organisations budget for AI tooling licences, model API costs, and implementation. The human time cost of botsitting is rarely a budgeted line item. When you add the cost of 6.4 hours per worker per week to the direct AI expenditure, the total cost of AI adoption looks materially different from the tooling cost alone. Across a 1,000-person organisation, botsitting represents roughly 6,400 hours of labour per week that is not captured in any AI ROI calculation — which is exactly the kind of oversight cost the productivity paradox analysis argues organisations must start budgeting for.
Deloitte’s 2026 State of AI in the Enterprise report provides corroborating evidence that agentic AI adoption is outpacing governance and cost modelling. Worker access to AI rose by 50% in 2025, yet only one in five companies has a mature governance model. The botsitting tax is being paid in full. It is just not being measured.
Not all botsitting is equal. The Work AI Index 2026 reveals that top AI achievers spend less time on botsitting per productive AI session than low achievers. This gap compounds over weeks and months. The segmentation suggests that botsitting overhead is influenced by skill (knowing how to prompt and verify efficiently), tooling (access to context-rich environments versus fragmented tools), and organisational practices (whether AI outputs flow through structured review or ad-hoc checking).
There is also an equity concern here. If botsitting burden falls disproportionately on less-experienced or less-resourced workers, it becomes a compounding disadvantage. The workers who can least afford the overhead spend the most time on it. Top AI achievers, about 13% of the sample, apply active judgement and embrace what the researchers call “productive botsitting” rather than simply prompting and praying. Context-rich organisations see 64% less digital exhaustion and 31% less botshitting. The rest are paying the full tax.
Go deeper: Botsitting and Botshitting: The Empirical Picture from the Work AI Index 2026 covers the 36% session failure rate, the exhaustion multiplier, and the full segmentation data.
The AI productivity paradox is the gap between individual experience and organisational reality. 87% of workers use AI regularly. 73% report personal productivity gains. Yet only 13% say their organisation is performing significantly better. The primary explanation is coordination neglect: AI helps individuals complete tasks faster, but the integration, review, and reconciliation of those outputs across teams creates a new coordination burden that organisations rarely measure or account for.
The personal productivity gains are genuine. The failure is in organisational systems that cannot translate individual acceleration into collective throughput. Coordination neglect, a concept from organisational research, describes exactly this pattern: investing in tools that accelerate individual work without investing in the coordination mechanisms (handoffs, reviews, approvals, alignment) that integrate that accelerated work into organisational output.
Rebecca Hinds gave a concrete example that captures the absurdity. One worker converts a single bullet point into a five-page report with AI. They send it to a colleague, who uses AI to convert it back into a single bullet point. Each person looks productive. The organisation spins on an “AI slop” hamster wheel. When every individual uses AI to produce more, faster, while the organisational chokepoints (review, approval, alignment, decision-making) receive more volume without more capacity, this pattern becomes the norm.
The Faros AI analysis of more than 10,000 developers across 1,255 teams provides a concrete software-engineering example. Individual developers complete 21% more tasks. But review times increase by 91%, teams merge 98% more PRs, and DORA metrics show no correlation with AI adoption at the company level. Individual speed gains pile up at team boundaries that have not accelerated.
The ILO’s “Aggregation Paradox of AI” research brief confirms this pattern across the economy. AI delivers large productivity gains at the task level (10% to 70% improvement). At the firm level, evidence is mixed. At the macroeconomic level, no clear AI-driven productivity growth has appeared in official statistics. Berkeley researchers have noted that the Solow Paradox (“You can see the computer age everywhere but in the productivity statistics”) remains at least partly relevant in the age of AI.
Why does this matter practically? Because it is a resource-allocation problem. Organisations investing heavily in AI tooling without investing in the coordination infrastructure that integrates AI outputs are spending on acceleration without building the road network. The 6.4-hour botsitting tax from the empirical data on AI oversight is the individual-level cost. At organisational scale, it compounds into systemic drag. Executives who measure only individual AI adoption and output volume are seeing a carefully framed picture that omits the coordination costs consuming the gains.
Go deeper: The AI Productivity Paradox: Why Individual Gains Do Not Scale to Organisational Performance covers measurement frameworks, budgeting heuristics, and the coordination-neglect research in full.
The distinction is between output-volume metrics and outcome-quality metrics. Output-volume metrics (tokens generated, tasks completed, time saved per individual) tell you whether AI is accelerating activity. Outcome-quality metrics (revenue per employee, decision accuracy, time-to-market, error rates in shipped work, botsitting hours trended over time) tell you whether that acceleration is translating into organisational improvement. The single most diagnostic question is: does AI-generated work survive peer review without more rework than human-generated work? If the answer is no, your AI is generating volume, not value.
Most organisations track AI adoption rates and individual productivity self-reports. Few have established organisational-performance metrics that isolate AI’s contribution. This creates a dangerous information asymmetry. The data that looks good (adoption, usage, tokens, self-reported time savings) is abundant and easy to collect. The data that reveals whether AI is actually working (decision quality, error rates, coordination costs, rework volume) is sparse and requires deliberate instrumentation.
The result is that organisations can be perfectly satisfied with their AI programme right up until they examine the organisational outcomes. Only 31% of AI spend can be attributed to specific business outcomes. The rest is being spent on acceleration that may or may not be improving anything.
Token maxxing is the failure mode that flourishes in this measurement vacuum. It describes the pattern of optimising AI output volume (tokens generated, tasks completed, drafts produced) without regard for whether the output improves organisational outcomes. One engineer at a major tech company, quoted in the Work AI Index, described deliberately inflating token numbers by asking AI questions already answered in documentation, prototyping features with no intention of building them, and defaulting to using agents even when manual work would be faster.
When organisations reward AI output volume without assessing output quality or organisational impact, they incentivise token maxxing. Workers learn that more AI output equals better performance reviews. Both reinforce each other while avoiding the harder question: is any of this making us better? This is Goodhart’s Law in action: when a measure becomes a target, it ceases to be a good measure. Breaking this loop requires shifting measurement from “how much AI output are we producing?” to “are AI-assisted decisions measurably better than pre-AI decisions?” — the full measurement framework analysis walks through the practical heuristics for making that shift.
Berkeley’s executive education researchers recommend a multi-dimensional measurement framework that covers efficiency, quality, capability, strategy, and human metrics like employee satisfaction and learning velocity. The point is not to adopt any one framework but to measure anything beyond output volume at all. Organisations that can answer three questions (are botsitting hours trending up or down? does AI-generated work need more or less rework at peer review? can you trace a measurable outcome improvement to a specific AI-assisted workflow?) are measuring AI’s organisational impact. Those that cannot are measuring AI activity.
Go deeper: The AI Productivity Paradox: Why Individual Gains Do Not Scale to Organisational Performance walks through measurement frameworks and the token-maxxing analysis in full.
Traditional enterprise software is deterministic, versioned, auditable, and behaves identically across instances. It follows what infrastructure engineers call the “cattle” model (uniform, replaceable). AI agents are probabilistic, stateful, context-dependent, and behaviourally variable. They follow the “pets” model (individual, requiring bespoke attention). This distinction has concrete operational implications for how you test, audit, and budget for the systems your organisation depends on.
The Pets versus Cattle distinction is not a metaphor. It is an operating model. Standard software: you deploy it, monitor it for uptime, patch it on a schedule, and replace broken instances. AI agents: you deploy them, but then you must continuously feed them context, monitor their decision quality (not just their uptime), intervene when they drift from expected behaviour, and maintain their state across sessions. A conventional API either works or it does not. An AI agent can produce an output that is plausible, syntactically correct, and subtly wrong in a way that downstream systems may not catch immediately.
Palo Alto Networks has been making the case that businesses must stop treating AI agents like standard software. Through their Portkey acquisition in April 2026, they are establishing what they call an AI Gateway as a control plane for autonomous agents. Lee Klarich, their Chief Product and Technology Officer, put it directly: “As autonomous agents join the enterprise workforce, they also become a new, unmanaged attack surface.” Agents act as highly privileged insiders, executing a large volume of automated decisions across internal and external systems.
The operational implications are real and specific. Reliability: a standard software failure is typically binary (it works or it does not). An AI agent failure is often partial and ambiguous (the output looks plausible but is wrong). 70% of enterprise leaders name “non-deterministic outputs” as the number one production-readiness barrier for AI agents.
Cost: standard software costs are predictable (licences, hosting, maintenance). AI agent costs are variable (model calls, retries, tool usage, context size) and dependent on usage patterns that are hard to forecast. Governance: standard software governance is about access control and change management. AI agent governance must also address decision quality, output auditability, and liability for agent actions. Agents without named owners have 2.7 times lower production-conversion rates.
The vendor response to this distinction is coalescing around what you might call the managed-AI-operations model, covered in detail a few sections ahead. The short version: platforms and architectures that treat AI agents as a fleet requiring continuous stewardship rather than a system requiring occasional maintenance. Anthropic’s Managed Agents, UiPath’s AgentOps, Microsoft’s Agent Framework, and Galileo’s Agent Control are all building pieces of this new operational discipline — an architectural shift the response framework article covers in depth.
Go deeper: Read on below for the managed-AI-operations model, or jump to The AI Productivity Paradox for the full Pets vs. Cattle analysis.
The psychological pathway from botsitting to botshitting runs through cognitive offloading, digital exhaustion, and satisficing. Cognitive offloading (delegating not just task execution but thinking itself to AI) creates the condition for moral disengagement, where workers distance themselves from responsibility for AI-generated outputs. Digital exhaustion from continuous context-switching between agent supervision and actual work produces satisficing behaviour: settling for “good enough” AI output rather than investing the cognitive effort to verify it.
Moral disengagement, in the AI workplace context, is the process by which workers rationalise shipping unverified AI output by displacing responsibility onto the tool. “The AI did it” becomes a psychologically available excuse because the worker did not personally produce the output. The 69% botshitting rate and the 40% versus 29% blame-deflection split are the empirical evidence that this mechanism is operating at scale.
This is a predictable psychological response to a work environment that demands AI usage, provides tools whose output looks plausible even when wrong, and offers no clear accountability framework for AI-mediated work. Research published in the NIH has found that long-term AI use is significantly associated with mental exhaustion, attention strain, and information overload, and inversely associated with decision-making self-confidence. Context-feeding and error correction are identified as the highest-exhaustion botsitting activities in the Work AI Index 2026.
When workers spend hours switching between AI tools, feeding each one organisational context it lacks, and correcting outputs that look correct but are substantively wrong, the cognitive load accumulates differently than it does from equivalent hours of focused work. This exhaustion creates the condition for satisficing, the behavioural-economics concept where decision-makers accept “good enough” rather than optimal because the cognitive cost of optimisation exceeds the perceived benefit. When a worker has spent three hours correcting AI errors, the motivation to spend another hour verifying the fourth output collapses.
Leaders face a genuine dilemma. Detecting botshitting requires some form of output review, but aggressive detection can drive AI use underground. Shadow AI is pervasive: over 80% of workers use unapproved AI tools. 54% of high AI achievers already use unapproved tools. 36% hide how much AI helps them. Banning AI does not work: nearly half of employees continue using personal AI accounts after a ban.
The AI teammate model offers a cultural framework that makes verification feel like professional practice rather than surveillance. It frames AI as a collaborator whose output requires the same collegial scrutiny as human work. AI detection tools like Pangram Labs provide a technical layer, but their deployment can undermine the psychological safety needed for workers to volunteer uncertainty about AI outputs. The goal is creating conditions where workers say “I’m not sure the AI got this right” rather than conditions where they hide their AI use entirely.
Go deeper: Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis covers the full moral-disengagement analysis and the detection-versus-psychological-safety tension.
Once you understand the psychological toll botsitting takes, the next question is practical: what work should you even be handing to AI in the first place? The decision is not purely about whether AI can do a task. It is about whether automating the task preserves the organisational value that the human performance of it produces. Some tasks create value beyond their output: they build craft skill, transmit organisational knowledge, create worker engagement, and maintain institutional memory. Automating a customer-service interaction might be efficient, but if it erodes the relationship knowledge and diagnostic skill that senior representatives develop through those interactions, your organisation loses capability it cannot easily regenerate.
The Meaning versus Automation Trade-off recognises that efficiency is only one dimension of work value. 41% of AI startups in Y Combinator’s portfolio are automating tasks people would prefer to keep human, according to research cited in the Work AI Index 2026. Tasks that build diagnostic intuition, maintain client relationships, develop junior staff, or preserve organisational memory may be technically automatable but strategically worth keeping human.
The customer-service-representative example illustrates the trade-off with uncomfortable clarity. Automating routine queries may free representatives for complex cases. But if routine queries are how representatives learn to handle complex cases, automation can erode the skill pipeline. Klarna made headlines replacing its customer service team with AI, only to hire them back. Many companies discover chatbots handle simple queries fine but completely fail on anything requiring judgment or empathy. Customer service rep jobs are declining just 4%, beating the broader job-loss benchmark, despite companies trying to automate customer service with AI. The work proves more resilient than the automation narrative suggests.
Rebecca Hinds framed the retention dimension bluntly: “Workers are being asked to automate the parts of their jobs they find most meaningful, while keeping the parts they find least engaging.” Workers who spend significant time managing AI rather than doing work they value are more likely to leave. This creates a compounding risk: automate the meaningful parts of work to save time, lose the workers who valued those parts, then find that the remaining workforce is both smaller and less engaged.
The framework for thinking about this is task mapping: evaluating tasks not just by automation feasibility but by the full spectrum of value they produce (output quality, skill development, worker engagement, organisational knowledge, relationship maintenance). Stanford Professor Emeritus Bob Sutton calls the reflex to solve problems by piling more on top instead of subtracting what is already there “addition sickness.” The automation conversation often suffers from exactly this reflex, adding AI to every workflow without asking whether the workflow should exist.
As Hinds put it: “We don’t just assume that because the technology can do something, because it can automate something, it should be automated. So much of work is meant to be messy. It’s meant to be full of friction because that friction builds ownership, good judgment, purpose, and pride.”
Go deeper: Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis has the full task-mapping framework, the customer-service case study, and the Bob Sutton organisational-design perspective.
The EU AI Act’s Article 14 mandates that high-risk AI systems must be designed to enable effective human oversight. Designated human operators must be able to understand, monitor, and intervene in the system’s outputs. The rules covering high-risk AI systems take effect on 2 August 2026. For organisations deploying AI in any of the eight high-risk categories (employment, essential services, law enforcement, critical infrastructure, education, biometric identification, administration of justice, and migration), botsitting transforms from an operational headache into a legal obligation.
The Act, formally Regulation 2024/1689, entered into force on 1 August 2024 with a phased implementation. Prohibited AI practices have been banned since February 2025. The high-risk system obligations take effect 2 August 2026. Full compliance for remaining requirements lands August 2027. This is not a distant regulatory horizon. It is a compliance deadline your organisation should be preparing for now.
The eight high-risk Annex III categories cover a broad swathe of enterprise AI use. Employment and worker management is the one most relevant to the botsitter economy conversation, covering AI systems used in recruitment, promotion, task allocation, and performance evaluation. If your organisation uses AI to make or inform employment decisions, Article 14 applies. The human oversight mandate requires oversight “by individuals with appropriate competence, training, authority, and support” with meaningful ability to intervene and override outputs.
What compliance involves: a conformity assessment for high-risk systems, a Fundamental Rights Impact Assessment (FRIA) for certain deployers, and ongoing real-world testing and monitoring. The CEN/CENELEC standards bodies are developing the technical standards that will operationalise these requirements. The European AI Office provides implementation guidance, and the EU AI Board coordinates national supervisory authorities.
The Act has extraterritorial scope. It applies to any organisation placing AI systems on the EU market or whose AI output is used within the EU, regardless of where the company is headquartered. Penalties are substantial: prohibited-practice violations carry up to EUR35 million or 7% of global annual turnover. Non-compliance with high-risk obligations carries EUR15 million or 3% of turnover.
This matters for the botsitter economy because the Act formalises and mandates a function that the Work AI Index 2026 data shows is already being performed haphazardly. The August 2026 effective date creates a compliance deadline that will force organisations to move from ad-hoc botsitting (the 6.4 hours of unstructured oversight the survey documents) to structured human oversight (the documented, auditable, role-defined oversight the Act requires). The Act gives the botsitter role legal teeth and a compliance budget, transforming it from an invisible cost centre to a compliance line item.
Anu Bradford, Professor of Law at Columbia University and author of The Brussels Effect, captured the shift: “AI regulation is no longer a theoretical debate, it is an operational reality. Companies that treat the EU AI Act as a compliance exercise rather than a governance opportunity will find themselves perpetually catching up.”
The readiness gap is significant. 60% of Fortune 100 companies will appoint AI governance heads in 2026, per Forrester predictions. AI-native companies and legacy organisations face different governance challenges. AI-native companies built their operations around the assumption of AI’s probabilistic, stateful nature. Legacy organisations are retrofitting governance frameworks designed for deterministic software onto systems that do not behave deterministically. Some organisations will clear the August 2026 deadline comfortably while others scramble.
Go deeper: Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis covers the eight high-risk categories, the conformity assessment process, and the FRIA requirements in full.
An enterprise graph is a connected model of your organisation’s people, documents, projects, goals, tools, and permissions. It is the connective tissue that allows AI agents to understand context without you manually feeding it to them. The Work AI Index 2026 identifies context-feeding as the highest-exhaustion botsitting activity, and the data shows that context-rich environments see 64% less digital exhaustion and 31% less botshitting.
Context-feeding is the most exhausting part of botsitting because it is structurally unnecessary. The information exists in your organisation’s systems (documents, project boards, communication tools, permission structures). But it is fragmented across tools, inaccessible to AI agents, and maintained only in human workflows. Workers spend hours feeding the same organisational information into multiple disconnected AI tools. 53% of workers said critical information needed for their jobs was not accessible through their AI systems.
The enterprise graph solves this by creating a unified, machine-readable model of organisational context that agents can query directly. Glean, the organisation behind the Work AI Index, positions its enterprise graph as the primary technical solution to the botsitting problem it documented. The concept is straightforward: connect your organisation’s data, people, and tools into a graph that AI agents can traverse to understand who is working on what, what documents are relevant, and what decisions have been made. Eliminate the manual context-feeding that consumes the largest share of botsitting hours.
The numbers make the case. Context-rich environments see 64% less digital exhaustion, 52% less unexplained shipped work, 9% less time botsitting, and 31% less time botshitting compared to context-poor environments. Context anxiety, where AI agents wrap up tasks prematurely as they sense their context limit approaching, is a failure mode specific to context-poor environments. Agents with access to an enterprise graph can retrieve context on demand rather than having it pre-loaded into a finite context window.
The vendor implementation landscape is developing quickly. Anthropic’s Managed Agents architecture (June 2026) includes a “Harness” layer that manages context and state for Claude-powered agents, with scheduled deployments added in June 2026. The architecture decouples the agent’s reasoning engine (brain), tool-execution environment (hands), and session state (event log and context) so each can be managed independently. This reduces the coordination complexity that makes context-feeding so exhausting.
Agent-to-agent handoff protocols and agent sandboxing address related concerns about agent sprawl and uncontrolled agent proliferation. The Model Context Protocol has crossed 9,400 public servers, and multi-agent orchestration has reached 22% of production deployments. The infrastructure for context-aware agent management is being built. The question for most organisations is not whether to adopt it, but when and how.
Go deeper: Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis covers the enterprise graph architecture, the context-richness diagnostic framework, and the Claude Managed Agents implementation.
The managed-AI-operations model treats AI agents as a fleet requiring continuous stewardship rather than a system requiring occasional maintenance. It has three layers: a context layer (enterprise graph or equivalent that gives agents organisational awareness), a control layer (centralised policy enforcement that governs what agents can do, when they escalate to humans, and how their decisions are audited), and an operations layer (AgentOps, the discipline of monitoring agent performance, detecting drift, managing costs, and maintaining evaluation coverage across the agent fleet).
The context layer eliminates manual context-feeding by giving agents governed access to organisational data. Enterprise graphs and the Model Context Protocol are the primary implementations. The control layer converts unstructured human oversight into structured, auditable intervention. Galileo’s Agent Control enables centralised, hot-reloadable human-in-the-loop policies with confidence-based escalation. UiPath’s governance layers provide fleet-wide policy enforcement.
The operations layer, AgentOps, extends principles from DevOps and MLOps to agentic systems. It contends with non-deterministic behaviour, autonomous tool use, and context-dependent reasoning that conventional monitoring cannot address. Drift detection, cost management, evaluation coverage, and auditability are the core functions. Together, these three layers transform botsitting from a diffuse individual burden into a structured organisational capability.
The vendor response in mid-2026 signals that this model is coalescing rapidly. Anthropic’s Claude Managed Agents (June 2026) added scheduled deployments and the Harness control architecture, decoupling the agent’s reasoning from its execution environment. The system lets you describe an agent in a chat interface, and it writes the prompt, picks the model, wires up MCP tools, and creates a cloud environment. UiPath’s Agent Builder and Maestro provide AgentOps capabilities: agent creation, evaluation, governance, cost visibility, and fleet-wide orchestration. Microsoft’s Agent Framework provides multi-agent orchestration patterns (sequential, concurrent, group chat, handoff).
These products address different layers of the same managed-operations stack. The 12% of AI agent pilots that reach production share consistent traits that align with this model: 94% have a named agent owner with budget authority, 87% run automated evaluations on every change, 81% scope to a single workflow with binary success criteria, and 74% deploy with explicit human-in-the-loop checkpoints for the first 60 to 90 days.
Only 12% of agent pilots reach production. 88% fail. The top blockers: evaluation and observability (64%), governance and compliance (57%), model reliability and non-determinism (51%), and data quality and access (49%). The managed-operations model addresses each of these. It makes human oversight efficient, auditable, and scalable rather than eliminating it.
The August 2026 EU AI Act deadline and the June 2026 vendor releases are converging on the same operational model from different directions. Organisations that adopt the managed-operations model are building the infrastructure that converts the 6.4 hours of unstructured botsitting into structured, role-defined oversight labour. The same oversight the EU AI Act will require for high-risk systems. 56% of enterprises now name a dedicated AI agent owner, up from 11% in 2024, and the role of AI Agent Owner is projected to reach more than 80% penetration by 2027.
Go deeper: Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis covers the enterprise graph architecture, the Claude Managed Agents implementation, and the context-richness diagnostic framework in full.
The botsitter economy will persist as models improve because AI oversight is a structural feature of knowledge work, not a temporary adoption friction. It is the emerging shape of work in an AI-augmented organisation, a workplace where the labour of supervising AI agents is as material to organisational performance as the labour the agents perform. AI does not eliminate the need for human judgement in knowledge work. It relocates it from production to supervision. Understanding the dimensions of that shift is the first step toward managing it productively.
The three dimensions this pillar has mapped describe a single structural shift. The empirical reality — 6.4 hours, 36% failure rate, 69% botshitting — documented at scale in the Work AI Index 2026 establishes that AI oversight is a material category of labour, not an edge case. The organisational paradox — individual gains that fail to aggregate — explains why that labour matters at scale. Individual gains consumed by coordination overhead are not a measurement error. They are a structural feature of how AI-augmented work flows through organisations that have not adapted their coordination infrastructure.
The three-pronged response — addressing the psychology of disengagement, the regulatory mandate for oversight, and the technical architecture that makes oversight feasible — explored in full in the response framework — is three dimensions of the same question: how do you make AI supervision sustainable rather than exhausting, auditable rather than invisible, and productive rather than parasitic on the gains AI creates?
The EU AI Act’s 2 August 2026 effective date for high-risk AI system rules transforms the botsitter economy from an operational observation into a compliance reality. Organisations deploying AI in any of the eight high-risk categories will need documented human oversight. The Work AI Index 2026 data shows that most organisations are currently providing unstructured, unmeasured, and unsustainable oversight. The regulatory deadline creates urgency that did not exist when botsitting was purely an efficiency concern. It also creates a budget line: compliance-driven oversight is easier to resource than efficiency-driven oversight, even when the activity is operationally identical.
The vendor signal confirms the direction of travel. Anthropic’s June 2026 Claude Managed Agents release and the broader vendor response from UiPath, Galileo, and Microsoft are market signals that the managed-AI-operations model is coalescing. Palo Alto Networks’ thesis that AI requires a different operational model from standard software — the central insight of the productivity paradox analysis — is being validated by product releases. The 12% pilot-to-production conversion rate and the projected 80% plus AI Agent Owner penetration by 2027 suggest the organisational response is following the vendor response.
The botsitter economy is already here, and the organisations that recognise it as a structural shift rather than a temporary friction will be the ones that build the oversight infrastructure to sustain AI adoption at scale. The Work AI Index 2026 put it best: “Companies pulling ahead aren’t spending a greater share of their AI time using AI. They’re spending a greater share on the work around it: setting context, defining what ‘good’ looks like, building judgment, and deciding what should never have been handed to a model.”
Where to go next:
If you are new to the topic, start with the empirical picture from the Work AI Index 2026. The 6.4 hours, the 36% failure rate, and the 69% botshitting rate are the factual foundation everything else builds on.
If you are evaluating organisational AI strategy, the productivity paradox analysis provides measurement frameworks and budgeting heuristics for understanding why individual gains are not adding up the way your dashboards suggest.
If you are responsible for AI governance or implementation, the three-pronged response covers the psychology, regulation, and architecture dimensions you will need to address.
Botsitting and Botshitting: The Empirical Picture from the Work AI Index 2026
The full data breakdown from Glean’s survey of 6,000 digital workers. The 6.4-hour botsitting tax, the 36% AI session failure rate, the 69% botshitting admission rate, and what the net productivity calculation looks like after you subtract the oversight overhead. Includes segmentation data showing how top AI achievers’ habits differ from low achievers’. Estimated read: 8 minutes.
The AI Productivity Paradox: Why Individual Gains Do Not Scale to Organisational Performance
Why 87% of workers use AI and 73% report personal gains, yet only 13% see organisational improvement. Covers coordination neglect, token maxxing, the Pets vs. Cattle infrastructure distinction, and practical measurement frameworks for evaluating whether AI is improving organisational outcomes. Estimated read: 7 minutes.
Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting Crisis
How moral disengagement explains the psychological pathway from botsitting to botshitting, how the EU AI Act (effective August 2026) transforms oversight into a legal mandate, and how enterprise graphs and managed-agent architectures provide the technical infrastructure to make oversight sustainable at scale. Estimated read: 8 minutes.
Suggested reading order: Start with Article 1 for the empirical foundation, continue to Article 2 for the organisational implications, and finish with Article 3 for the integrated response.
The Work AI Index 2026 was published by the Work AI Institute, commissioned by Glean. The survey covered 6,000 digital workers across the US, UK, and Australia between December 2025 and January 2026. It is the source for the 6.4-hour botsitting statistic, the 36% AI session failure rate, the 69% botshitting admission rate, and the segmentation data comparing high AI achievers to low achievers. The full empirical picture article provides the key findings with source attribution.
Only 12% of AI agent pilots successfully reach production (88% fail). The 12% that succeed share consistent traits: a named agent owner (2.7 times higher conversion rates), scoped success criteria, automated evaluation coverage, and explicit human-in-the-loop checkpoints. Gartner, Forrester, and McKinsey have each published data supporting variations of this finding across different industry segments.
Coordination neglect is the pattern where organisations invest in tools that accelerate individual work (AI, automation) without investing in the coordination mechanisms (handoffs, reviews, approvals, alignment) that integrate that accelerated work into organisational output. Individual speed gains pile up at organisational chokepoints that have not been scaled, creating bottlenecks that consume the productivity gains. The productivity paradox article explores this in full.
Token maxxing describes the pattern of optimising for AI output volume (tokens generated, tasks completed, drafts produced) without regard for whether that output improves organisational outcomes. When organisations measure AI success by output volume rather than outcome quality, they inadvertently incentivise workers to produce more AI-generated content rather than better AI-assisted decisions. The measurement section of the productivity paradox article provides the framework for distinguishing output metrics from outcome metrics.
Context anxiety is the behaviour where large language models wrap up tasks prematurely as they sense their context window limit approaching. They rush to finish before running out of working memory. This creates unreliable outputs that require human intervention to complete or correct, adding directly to botsitting overhead. Context anxiety is most prevalent in context-poor environments where agents lack access to an enterprise graph and must operate within a finite context window. The enterprise graph section of Article 3 covers this in detail.
Pets vs. Cattle is an infrastructure paradigm that distinguishes between systems you treat as unique, stateful, and requiring individual attention (pets) and systems you treat as uniform, replaceable, and managed at scale (cattle). Traditional enterprise software follows the cattle model. AI agents follow the pets model because they are probabilistic, context-dependent, and behaviourally variable. You cannot test an AI agent once and assume consistent behaviour, you cannot budget for AI operations using standard software cost models, and you need governance frameworks that account for decision quality rather than just uptime. The productivity paradox article covers this in its analysis of how managing AI differs from managing standard software.
The official EU AI Act text (Regulation 2024/1689) is published in the Official Journal of the European Union and accessible through EUR-Lex. The European AI Office provides implementation guidance, and the CEN/CENELEC standards bodies are developing the harmonised technical standards that will operationalise the Act’s requirements. The EU AI Board coordinates national supervisory authorities. The EU AI Act section of Article 3 provides an overview of the human oversight mandates and the August 2026 compliance timeline.
The Work AI Index 2026 data shows that context-rich environments (where AI agents can self-contextualise through an enterprise graph or equivalent infrastructure) see 64% less digital exhaustion and 31% less botshitting compared to context-poor environments. Context-rich environments also reduce context anxiety, lower the repetition burden of feeding the same organisational information to multiple tools, and enable more reliable agent outputs because agents operate with current, connected organisational data rather than stale or fragmented context. The enterprise graph section of Article 3 provides the full comparison.
Psychology, Regulation, and Architecture: A Three-Pronged Response to the Botsitting CrisisSixty-nine percent of knowledge workers admit to delivering AI-generated work they cannot explain or defend. The Glean Work AI Index 2026 gave this behaviour a name, “botshitting,” and the numbers tell a larger story: 41% sometimes deliver work they could not explain if asked, and 28% have blamed AI for mistakes they themselves caused. The hidden labour of making AI useful, feeding it context, debugging its outputs, cleaning up its messes, consumes 6.4 hours per week. That is roughly half of all reported AI time savings.
The response cannot come from a single direction. Psychology explains why workers disengage from AI outputs. Regulation makes oversight a legal obligation. Architecture provides the infrastructure that makes oversight feasible at scale. Here is how each domain fits together within the full botsitter economy picture.
The first domain, psychology, explains why workers disengage from AI outputs in the first place.
Moral disengagement is Albert Bandura’s concept describing the cognitive mechanism by which people deactivate their moral self-regulatory processes, making unethical decisions without guilt. In AI work contexts, the mechanism is straightforward: when your team offloads thinking to AI, they also offload responsibility for the output. “The AI did it” becomes the cognitive off-ramp.
Botshitting is the endpoint of this process — documented in the empirical data at a 69% rate — but botsitting is where it begins. Botsitting is the constant oversight labour of supervising AI outputs: feeding the tool context, debugging what it produces, cleaning up its mistakes. When your team spends 6.4 hours a week on this work, the exhaustion is real.
The behavioural evidence is clear. When AI-generated work fails, 40% of workers blame the AI tool rather than their supervision choices; only 29% admit fault. Heavy AI users are 3.4 times more likely than light users to blame the tool when something goes wrong. As Zoe Rahwan of the Max Planck Institute for Human Development puts it, delegating tasks to AI is like having a buffer that lowers your own moral accountability.
The psychological pathway runs like this: botsitting produces digital exhaustion, exhaustion produces satisficing (workers settle for “good enough” AI output rather than verifying it), satisficing becomes botshitting, and botshitting erodes accountability across your organisation. Long-term AI use is associated with mental exhaustion and attention strain, and inversely associated with decision-making self-confidence.
There is a compounding factor. When human-AI teams collaborate, personal accountability for output quality diminishes compared to working alone. Responsibility diffuses across multiple actors in the AI’s design, deployment, and use. The “problem of many hands” means the more hands involved, human or synthetic, the thinner the accountability.
Detection seems like the obvious response, but it collides with psychological safety. AI detection tools like Pangram Labs can identify AI-generated content, but surveillance-heavy approaches create a predictable backlash. Over 80% of employees use unapproved AI tools, and nearly half of employees continue using personal AI accounts after a ban. Among self-identified high AI achievers, 54% are using unapproved tools or using approved tools in noncompliant ways. Thirty-six percent actively hide how much AI helps them.
The alternative is cultural rather than technical. The AI teammate model treats AI output as a collaborator’s contribution requiring the same collegial scrutiny as human work. This normalises critical evaluation without stigmatising AI use. The precondition is psychological safety, Amy Edmondson’s concept of an environment where team members feel safe admitting uncertainty without fear of blame.
The evidence supports this directly. 83% of executives believe psychological safety measurably improves AI initiative success, yet fewer than 39% rate their organisation’s current level as “very high.” Twenty-two percent of leaders admit they have hesitated to lead an AI project because they might be blamed if it misfires. As Rebecca Hinds of the Glean Work AI Institute observes, the ideal is people raising their hands and saying they contributed to botshitting and explaining why. That only happens in environments where disclosure is rewarded, not punished.
The goal becomes creating conditions where workers actively seek collegial review of AI-assisted work. Scrutiny becomes professional practice rather than punitive surveillance.
But even in a psychologically safe culture, you still need to decide which tasks should involve AI at all. That is where the Meaning vs. Automation Trade-off comes in.
The question is not whether AI can do a task. The question is whether automating it preserves the craft skill, organisational knowledge, and work meaning that the task’s human performance produces. This is the Meaning vs. Automation Trade-off.
Task mapping is the practical method. Catalogue what your organisation does and assess each task against two criteria: can AI do it adequately, and does human performance of the task produce value beyond the output itself? The customer service representative example is instructive. Automating routine queries increased throughput but eroded the deep customer knowledge representatives built through handling those queries themselves. For representatives asked to automate work they would rather do themselves, this alienation predicts both reduced engagement and increased turnover.
Some tasks should remain stubbornly human because their performance builds skill, preserves organisational memory, and sustains worker engagement. Efficiency metrics alone cannot capture these values. A useful diagnostic: would you bet your job on the output from this AI tool? If the answer is no, the task still requires human judgment.
There is a subtler risk. In organisations where demonstrating AI fluency has become a form of self-presentation, workers automate meaningful tasks to appear “AI-native” when the automation destroys the very value their role produced. Automation decisions are organisational design decisions masquerading as technology choices.
The EU AI Act transforms botsitting from an operational headache into a compliance obligation. Article 14 requires that high-risk AI systems be designed so designated human operators can understand, monitor, interpret, and override AI system outputs. This is not superficial rubber-stamping. Operators must have appropriate competence, training, authority, and support to intervene meaningfully.
The eight high-risk AI system categories cover a broad swathe of enterprise use: biometric identification, critical infrastructure, education and vocational training, employment and worker management, access to essential services, law enforcement, migration and border control, and administration of justice. Employment and worker management alone captures most enterprise AI deployment. AI used for recruiting, screening, performance evaluation, or other employment-related decisions is explicitly high risk.
Organisations must demonstrate compliance through documented conformity assessments. The CEN and CENELEC standards bodies are developing the technical specifications that operationalise these requirements. Certain high-risk systems also require a Fundamental Rights Impact Assessment before deployment, a governance mechanism enabling structured discussion about adverse impacts rather than a post-hoc justification.
The timeline creates immediate urgency. August 2, 2026 is the date high-risk AI systems must meet full compliance. (A November 2025 Omnibus revision proposed postponing this to December 2027, but the final timeline remains unsettled. Monitor the Official Journal for confirmation.) The Act has extraterritorial scope, similar to GDPR: any organisation placing high-risk AI systems on the EU market or whose AI outputs affect people within the EU is in scope. Penalties reach €35 million or 7% of global annual turnover.
Two governance frameworks provide the practical playbooks. The NIST AI RMF’s Govern-Map-Measure-Manage functions map to Article 14 oversight requirements. ISO/IEC 42001 provides a certifiable management system aligned with the Act’s requirements around risk management, data governance, and post-market monitoring. Yet only 37% of organisations have AI governance policies in place, and compliance programmes take 12 to 18 months to implement.
There is a structural dimension worth noting. AI-native organisations operate with continuous monitoring and agent-aware architecture built in from day one. Legacy organisations must retrofit these capabilities into systems never designed for AI oversight. Governance teams already spend 37% more time managing AI risk. The compliance burden falls disproportionately on the organisations with the least AI maturity.
Article 14 requires oversight — the productivity paradox makes coordinated oversight urgent — but oversight at scale requires infrastructure that most organisations do not yet have. That infrastructure starts with the enterprise graph.
The enterprise graph is the connective infrastructure that makes human oversight feasible at scale. Glean’s data model connects people, tasks, documents, projects, goals, technology, and organisational context into a queryable knowledge structure. AI systems drawing from an enterprise graph understand not just what information exists but its recency, authoritativeness, and relevance.
The reason this matters is that context-feeding is the single most exhausting botsitting activity. Workers spend their 6.4 weekly botsitting hours manually shovelling organisational context into fragmented AI tools that lack persistent awareness of who does what, what matters now, and which documents are authoritative. More than a third of AI sessions fail outright, requiring a full restart or substantial rework. The enterprise graph automates this context-feeding entirely.
The comparative data is striking. Context-rich environments, those with an enterprise graph and context engineering, see 64% less digital exhaustion and 31% less botshitting compared to context-poor environments where each AI tool starts from zero. When AI agents operate without organisational context, they produce outputs that sound right but are substantively wrong, creating exhausting verification labour for human supervisors. In the extreme, business becomes farce: a perpetual motion machine of AI-generated slop.
Anthropic’s Claude Managed Agents, announced in June 2026 with scheduled deployments, represent one vendor’s implementation of this architecture. Their Harness system decouples the “brain” (Claude’s reasoning) from the “hands” (sandboxed execution environments), enabling long-horizon agent operation with clean security boundaries. Anthropic also identified that agents sometimes wrap up tasks prematurely as they sense their context limit approaching, a behaviour they termed “context anxiety,” and addressed it by adding context resets.
The practical assessment is straightforward. Ask four diagnostic questions of each AI tool your team uses: does it know who the user is, what project they are working on, which documents are authoritative, and what decisions were made in the last meeting? If the answer is no across multiple tools, your environment is context-poor. Production-ready agent platforms demonstrate persistent context across sessions, agent sandboxing separating execution from credentials, agent-to-agent handoff protocols, and governance monitoring. Prototyping-only platforms lack these and create agent sprawl, the uncontrolled proliferation of agents without unified context or governance.
This architectural layer connects directly to the psychology and regulation domains described earlier. When AI tools arrive context-poor, your team exhausts themselves feeding context. When they break, they morally disengage. Better infrastructure breaks this chain. The EU AI Act provides the external pressure. Enterprise graphs and managed agents provide the mechanism.
The three domains, psychology, regulation, and architecture, are not a menu of options. They are interlocking necessities, tying psychology, regulation, and architecture into a coherent response to the botsitting crisis. Moral disengagement explains why workers stop supervising AI. Regulation makes supervision legally mandatory. Architecture makes supervision practically feasible at scale. Remove any one prong and the response collapses: psychology without architecture burns workers out on manual context-feeding; architecture without regulation lacks the organisational urgency to justify investment; regulation without psychology creates punitive surveillance that drives AI use underground.
The 69% botshitting rate is not an individual moral failing. It is the predictable output of systems that demand supervision without providing the infrastructure to make it sustainable. Better infrastructure, architecture that frees people to do the judgment work that only people can do, is a deeply human response to the botsitting crisis. The organisations that will thrive combine architectural infrastructure that makes human oversight lightweight with psychological conditions where workers actively seek collegial scrutiny of AI-assisted work.
August 2026 is not a distant deadline. It is the present. Organisations treating the three-pronged response as sequential, psychology first, then regulation, then architecture, will find themselves out of time. The viable path is simultaneous action across all three.
Botsitting is the hidden labour of making AI useful: feeding it organisational context, debugging its outputs, verifying its claims, and cleaning up the messes it leaves behind. The Glean Work AI Index 2026 quantified this at 6.4 hours per week per knowledge worker, roughly half of all reported AI time savings. It is the unglamorous underside of the AI productivity story that no vendor slide deck mentions.
Botshitting is not an innocent error; it is the deliberate delivery of AI-generated work that the worker cannot explain or defend. The distinction lies in knowing abandonment: the worker recognises they have not properly verified the output but submits it anyway. Mistakes involve effort and oversight that fell short; botshitting involves withdrawing that oversight entirely, then deflecting accountability onto the tool.
It is primarily a system problem with psychological consequences. Workers are placed in environments where context-poor AI tools demand constant, exhausting manual feeding, surveillance-heavy oversight cultures drive AI use underground, and individual productivity metrics reward output volume over output quality. The 69% figure reflects a rational response to a broken system, not a sudden epidemic of professional negligence. Fix the architecture and the psychology follows.
Three things converge: operational rot, regulatory exposure, and talent flight. Operationally, AI-generated slop proliferates through decision-making pipelines, eroding institutional knowledge. Under the EU AI Act, organisations deploying high-risk AI systems without meaningful human oversight face enforcement action from August 2026. And workers who spend half their AI time savings on botsitting are workers who will take their skills somewhere that treats their attention as a finite resource.
Yes, if those organisations place high-risk AI systems on the EU market or their AI outputs affect people within the EU. The Act follows the same extraterritorial logic as GDPR: market access is the lever, not corporate domicile. Australian and American organisations selling AI-powered HR tools, hiring platforms, or critical infrastructure systems into Europe are squarely within scope. Compliance is not optional just because the headquarters is in Sydney or San Francisco.
The enterprise graph is the infrastructure (Glean’s connected data model linking people, projects, documents, and organisational knowledge). Context engineering is the practice (Anthropic’s term for curating the optimal information presented to an AI for inference). The enterprise graph provides the raw connective tissue; context engineering determines which threads the AI pulls on for a given task. Both are necessary; neither is sufficient alone.
Managed Agents reduce botsitting by removing the need for workers to manually shovel context into AI tools before every interaction. Because Managed Agents maintain persistent awareness across scheduled deployments (who the user is, which documents are authoritative, what decisions were made in the last meeting), the worker shifts from context-feeder to judgment-reviewer. The drudgery is automated; the oversight that remains is the oversight that matters.
Yes, and the gap is substantial. AI-native organisations operate with continuous monitoring, agent-aware architecture, and governance assumptions built into their tooling from day one. Legacy organisations must retrofit these capabilities into fragmented systems that were never designed for AI oversight. The compliance burden falls disproportionately on the organisations with the least AI maturity, creating what the article describes as a split change-management landscape.
Start with lightweight context engineering: create and maintain a single shared document that maps who does what, which documents are authoritative, and what decisions were recently made. Even manual context curation, applied consistently, reduces the exhausting cold-start problem where every AI interaction begins from zero. The principle matters more than the platform: persistent context trumps fragmented tooling, regardless of budget.
Name the trade-off explicitly. Frame it not as resistance to AI but as the difference between automation that elevates work and automation that hollows out the capabilities that make your role valuable. Reference the Meaning vs. Automation Trade-off: some tasks should remain stubbornly human because their performance builds skill, preserves organisational memory, and sustains the deep knowledge that AI tools consume but cannot produce. If your manager still insists, document your concerns; the EU AI Act’s human oversight mandates are on your side.
The evidence is specific. Amy Edmondson’s thirty years of research demonstrates that teams with high psychological safety outperform on precisely the behaviours that counter botshitting: admitting uncertainty, surfacing errors early, and scrutinising each other’s work collegially rather than defensively. In an AI context, this means workers volunteer that they cannot explain an AI output rather than hiding their AI use. The mechanism is not soft; it is the hard precondition for honest disclosure at scale.
Ask four diagnostic questions of each AI tool your teams use: does it know who the user is, what project they are working on, which documents are authoritative, and what decisions were made in the last meeting? If the answer is no across multiple tools and your workers are manually re-entering the same organisational context into different systems, your environment is context-poor. The consequence is predictable: 64% more digital exhaustion and 31% more botshitting than context-rich environments.