Somewhere between $6.7 trillion and $7.6 trillion in AI infrastructure capital is projected to be deployed by the early 2030s. Not venture capital. Announced corporate capital expenditure from companies already guiding to $725 billion for a single year.
The question is not whether the number is real. It is. The question is whether it can actually pay off.
This article is part of the AI infrastructure arms race — a cluster of analyses examining the forces reshaping enterprise AI from the infrastructure layer up. What Q1 2026 earnings revealed about the current capex run rate is the starting point: $725 billion in announced 2026 capital expenditure, with roughly 75% directed at AI-specific infrastructure. Here the goal is narrower: stress-test the return assumptions and reframe the investor’s question into the operational one. How long does the compute-cost advantage last, and what does GPU depreciation math actually imply for cloud pricing in 2028–2029?
Where does the $7 trillion AI infrastructure figure come from — and what does it require to be true?
Goldman Sachs Global Institute’s “Tracking Trillions” report from May 2026 puts the number at $7.6 trillion in cumulative AI capital expenditure between 2026 and 2031. The $7 trillion figure you see attributed to McKinsey and the World Economic Forum circulates via secondary attribution — no primary report has been retrieved. Treat Goldman Sachs as the primary analytic frame here.
The current run rate is not in doubt. Amazon, Alphabet, Meta, Microsoft, and Oracle have collectively guided to approximately $725 billion in 2026 capital expenditure, with roughly 75% directed at AI-specific infrastructure. Meta is tracking capex equal to 54% of sales. Microsoft is at 47%.
For the projection to generate adequate returns, three conditions need to hold simultaneously: revenue growth from AI workloads that outpaces depreciation burn; enterprise adoption curves that absorb new compute supply; and token pricing that stabilises at levels sufficient to support the underlying revenue model. Each condition is genuine — and each is genuinely uncertain.
What is the bull case for AI infrastructure returns — and why does Goldman Sachs say the cycle may have further to run?
The primary bull-case frame is Goldman Sachs’s GDP comparison. AI capex sits at approximately 0.8% of global GDP. The 1990s telecom buildout peaked above 1.5% of GDP before correcting — that gap is Goldman’s primary argument that headroom remains.
Goldman analyst Ryan Hammond has documented a consistent pattern: consensus implied capex growth of roughly 20% at the start of both 2024 and 2025; reality exceeded 50% in both years. Jefferies analyst Brent Thill put the bull position plainly: “The bear thesis is garbage.”
The structural underpinning rests on two forces. The first is elastic demand — Jevons Paradox applied to AI compute.
💡 Jevons Paradox is the economic principle that falling resource costs often increase total consumption rather than reduce it, because lower prices unlock use cases that were previously uneconomical.
As token prices fall, the addressable set of AI workloads expands faster than per-unit revenue contracts. OpenRouter’s 2025 State of AI documents this directly.
The second force is the inference era shift. Inference now accounts for 60–70% of total AI compute demand, up from roughly 40% in 2024. Token prices fell from approximately $60 per million output tokens in early 2023 to less than $1.50 by early 2025 — a 40x reduction — and usage expanded to fill the capacity made available.
GDP headroom plus elastic demand plus inference deepening. A cycle still in a self-reinforcing growth phase, by that reading.
Why does GPU depreciation make the bear case so much more severe than the dot-com comparison suggests?
The bear case has an analytical anchor the dot-com comparison completely misses: GPU depreciation math.
“Short-lived assets” is Microsoft’s SEC disclosure term for technology hardware — primarily GPUs and CPUs — with a useful economic life of roughly 2.5 to 3 years. Approximately two-thirds of Microsoft’s quarterly capex sits in this category. That means roughly $25 billion of their most recent $37.5 billion quarterly capex must earn its return within three years.
The math in plain terms: if a hyperscaler deploys $100 billion in GPU infrastructure, roughly $67 billion must be fully depreciated within three years. To generate a 15% return on capital, that asset base needs to earn approximately $10 billion per year in gross profit before it gets replaced.
That is structurally more severe than the 1990s fibre overbuilding. When Worldcom laid excess fibre in 1999, those cables lasted 20–30 years — the assets eventually carried traffic. GPU infrastructure that is not earning revenue in year one is burning through its economically useful life in real time.
The utilisation rate imperative follows directly. At 90% utilisation, a GPU earns roughly $31,500 annually; at 50%, approximately $11,000. A 2.5-year depreciation schedule implies a capital cost of roughly $5,000 per GPU per year — feasible at high utilisation, problematic at low. Idle GPUs do not get a depreciation holiday.
It is a binary dilemma: raise utilisation rates by capturing more enterprise workloads, or raise cloud prices to earn adequate returns per unit deployed. Either path has direct implications for enterprise AI compute costs in 2028–2029 — which connects directly back to the $725 billion AI capex picture.
What do the FCF stress signals from Amazon, Meta, and CoreWeave tell us?
These are bear-case data points, not proof of a bubble.
💡 Free cash flow (FCF) is operating cash flow minus capital expenditure — the cash a business actually generates after paying for the assets it needs.
Amazon’s FCF collapsed from $26 billion in Q1 2025 to $1.2 billion in Q1 2026. JP Morgan projects it could turn negative by up to $28 billion for full-year 2026. Amazon CEO Andy Jassy’s response was that “most of the new supplies are already spoken for.” The question is whether revenue materialises before the GPU depreciation clock expires.
Meta’s free cash flow is projected to drop approximately 90% in 2026. When Meta first signalled a major capex increase, the stock dropped 11%. Investors ultimately rewarded the commitment — which makes Meta a complicated bear example.
CoreWeave‘s Q2 2026 guidance miss is the most direct real-time stress indicator. Q2 guidance of $2.45–2.6 billion fell short of analyst consensus at $2.69 billion — the stock down 4% despite a $99.4 billion revenue backlog.
Capital is being deployed well ahead of revenue materialisation. Whether that resolves within the GPU depreciation window is the key question. For the systemic financing risk that sits underneath all of this, see our analysis of the BIS warning about off-balance-sheet AI financing risk.
Is the AI infrastructure buildout a repeat of the dot-com bubble — what is the same and what is different?
The similarities are structural. Infrastructure investment is running well ahead of proven near-term demand. Equity markets are rewarding capex commitment before revenue materialises. And circular financing dynamics have emerged — NVIDIA has invested $100 billion in OpenAI, which buys NVIDIA chips; Microsoft has invested in OpenAI, which pays Microsoft for Azure.
The differences are also structural. Data centre vacancy rates stand at a record low of approximately 1.6%, with three-quarters of capacity under construction already pre-leased. AI applications are generating measurable revenue today. JP Morgan notes that “whereas early internet firms built first and monetized later, AI is monetizing as it builds.”
GPU depreciation also creates faster feedback loops than anyone got in the 1990s. Those fibre cables lasted decades. GPU infrastructure that is not earning adequate returns becomes apparent within 2–3 years.
JP Morgan captures the synthesis well: “We think the risk that a bubble will form in the future is greater than the risk that we may be at the height of one right now.” Both sides are probably right about different parts of the picture.
What does the AI infrastructure returns debate mean for enterprise AI planning — specifically cloud compute costs in 2028–2029?
The investor’s question is: “Will hyperscalers get their money back?” The board’s question is different: “How long does the compute-cost advantage last?” These are not the same question, and they have different answers.
The 40x decline in token prices over the past two years has driven genuine enterprise AI adoption. But the Arxiv AI Token Futures analysis argues this reflects a supply-driven buyer’s market — providers subsidising inference below marginal cost to acquire workloads. When the application layer explodes, that dynamic will reverse.
As the capital recovery window compresses, the downward pricing trajectory may flatten or reverse in the 2027–2029 period. Total AI budgets may still grow even if per-unit token costs stabilise — cheaper models across more workloads drive expanding consumption. The compute bill keeps climbing even when unit prices plateau.
So, three practical decisions you can make now:
1. Negotiate multi-year committed-use pricing now. The current pricing environment may be the most favourable before GPU depreciation pressure affects cloud cost trajectories.
2. Treat the current compute-cost decline as a finite window, not a structural baseline. Budget decisions based on continued cost deflation should be stress-tested against a scenario in which pricing flattens or reverses by 2028.
3. Build architectural flexibility across cloud providers. If pricing diverges between AWS, Azure, and Google Cloud as the depreciation cycle matures, workload portability becomes a direct cost lever. Lock-in to a single hyperscaler removes it entirely.
The $7.6 trillion Goldman Sachs projects through 2031 contains the conditions for both a transformative general-purpose technology buildout and a significant capital misallocation episode. Distinguishing between those outcomes in advance is harder than either side acknowledges. Build accordingly.
For a complete overview of the infrastructure forces shaping this picture — from neocloud financing structures to sovereign compute geography — see our comprehensive AI infrastructure arms race resource.
Frequently Asked Questions
What is the $7 trillion AI infrastructure projection by 2030?
Goldman Sachs put the number at $7.6 trillion cumulative through 2031 via its May 2026 “Tracking Trillions” analysis. The widely-cited $7 trillion figure attributed to McKinsey and WEF circulates via secondary attribution only — treat Goldman Sachs as the primary analytic source.
Who is Ryan Hammond and why does his analysis matter for AI infrastructure ROI?
He is a Goldman Sachs Research analyst who documented that consensus estimated 20% capex growth for both 2024 and 2025; reality exceeded 50% in both years. That systematic underestimation underpins Goldman’s entire bull-case framing.
What does “short-lived assets” mean on a hyperscaler balance sheet?
It is Microsoft’s SEC disclosure term for technology hardware — primarily GPUs and CPUs — with useful economic lives of roughly 2.5–3 years. Approximately two-thirds of Microsoft’s quarterly capex falls into this category, meaning the capital recovery clock runs far faster than for data centre buildings, which have useful lives of 25–40 years.
Why does a 40x drop in token prices make the AI ROI picture harder to read?
If the cost of an AI inference token falls 40x, the revenue each unit of compute generates also falls 40x — unless usage volume expands proportionally. The model commoditisation risk is that infrastructure was scaled to generate revenue at 2023 token prices that no longer exist. The bull case (Jevons Paradox) is that cheaper tokens drive proportionally more usage, keeping total revenue stable or growing.
What is Jevons Paradox and why does it matter for AI infrastructure?
💡 Jevons Paradox is the economic principle that falling resource costs often increase total consumption rather than reduce it, because lower prices unlock use cases that were previously uneconomical.
Applied to AI: cheaper tokens may expand the addressable market faster than prices are falling, keeping aggregate compute revenue intact even as per-unit economics compress.
What is CoreWeave’s Q2 2026 guidance miss and why does it matter?
CoreWeave is the leading publicly listed GPU cloud operator — a direct proxy for AI compute demand. Its Q2 2026 guidance fell short of analyst consensus, sending the stock down 4% despite a $99.4 billion revenue backlog. It is a leading indicator of demand-supply mismatch, not definitive proof of a bubble.
How does Amazon’s negative free cash flow projection affect the AI infrastructure investment thesis?
JP Morgan projects Amazon’s free cash flow could turn negative by up to $28 billion in 2026. AWS revenue continues to grow, but the FCF pressure reflects capital deployed ahead of revenue materialisation. Whether it resolves before the GPU depreciation clock expires is the key question.
Is the AI infrastructure build-out a bubble like the dot-com era?
Structurally similar in some ways: capital ahead of near-term demand, equity markets rewarding capex commitment, circular financing dynamics. Structurally different in others: AI applications generate measurable revenue today; data centre vacancy rates are at a record low 1.6%; GPU assets depreciate in 2.5–3 years, creating faster feedback loops than 20-year fibre.
How does AI capex as a percentage of GDP compare to the 1990s telecom buildout?
Goldman Sachs estimates current AI capital expenditure at approximately 0.8% of global GDP. The 1990s telecom buildout peaked above 1.5% of GDP before correcting — that is Goldman’s primary argument that the bull case has remaining headroom.
What should a company do differently with its AI compute budget given this uncertainty?
Three things: negotiate multi-year committed-use pricing while the current environment holds; treat the current compute-cost decline as a finite window, not a permanent condition; and build workload architecture that can shift between AWS, Azure, and Google Cloud to preserve flexibility if pricing diverges.
Why might cloud AI compute prices rise in 2028–2029 even if they are falling now?
Current token pricing reflects a supply-driven buyer’s market — providers subsidising inference below marginal cost to acquire enterprise workloads. If hyperscalers face sustained FCF pressure and GPU assets must be depreciated within 2.5–3 years, they will need to either raise utilisation rates or raise prices to recover capital before assets become obsolete.
