When 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.
How do AI IPOs compare to traditional tech IPOs in terms of risk profile?
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 AI models get superseded by a single release.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.
What metrics should investors use to assess an AI company’s revenue quality versus traditional SaaS?
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.
How do AI pricing models differ from traditional SaaS 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.
What is model obsolescence risk and how does it affect AI company valuations?
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.
How do I evaluate an AI company’s compute dependency and infrastructure risk?
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.
How do I gauge the durability of an AI company’s competitive moat before investing?
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.
How do I determine whether an AI company’s governance structure protects shareholder interests?
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?
How should retail investors evaluate an AI company IPO before buying?
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.
Frequently Asked Questions
What is the AIBQ framework and how should I use it?
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.
What exactly is circular financing and why does it matter?
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.
Should I buy an AI IPO on the first day of trading or wait?
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.
How long should I expect to hold an AI company before it reaches profitability?
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.
Is it true that open-weight models make AI company moats worthless?
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.
What happens to AI company valuations if NVIDIA loses its GPU monopoly?
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.
What is the difference between investing in Core AI versus Applied AI companies?
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.
How do lock-up periods affect AI IPO investing?
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.
What regulatory risks should AI investors watch for beyond safety incidents?
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.
Is it better to invest in an AI ETF than individual AI IPOs?
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.
What questions should I ask during an AI company’s first earnings call?
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.
How do I tell whether an AI company is a platform or just a feature?
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.
