| Financial State of the AI Market | |
| 📊 2024 – 2026 YTD | |
| Scope | Generative-AI economics: revenue, spend, debt, investment |
|---|---|
| As of | 29 May 2026 |
| Combined 2026 capex (top builders) | ~$660–690 bn (guidance) |
| OpenAI run-rate | ~$24 bn; 2025 revenue $13 bn |
| Anthropic run-rate | ~$30–45 bn (Apr–May 2026) |
| NVIDIA data-centre (FY25) | $115.2 bn (+142% YoY) |
| Mag-7 share of S&P 500 | ~34.8% |
| Central tension | ~$1 tn+ in commitments & capex vs ~$150 bn AI revenue |
A data-led tracker of the money behind the generative-AI boom — who is earning, who is spending, who is borrowing, and what the numbers imply for the months to Christmas 2026. This page deliberately sets the bullish industry narrative against the bearish “bubble” case (most prominently argued by writer Ed Zitron) and tries to let the figures, drawn from company filings, Bloomberg, Reuters, CNBC and others, do the arguing.
Between 2024 and mid-2026 the generative-AI sector executed one of the fastest capital build-outs in corporate history. Two facts sit side by side and define the whole debate. First, revenue is real and growing extraordinarily fast: OpenAI roughly quadrupled revenue (from about $3.7 billion in 2024 to $13 billion in 2025), Anthropic went from a ~$1 billion run-rate at the end of 2024 to a $30 billion-plus run-rate by April 2026, and NVIDIA’s data-centre business grew from $47.5 billion (FY24) to $115.2 billion (FY25). Second, spending and forward commitments are far larger than current revenue: the five biggest builders have guided to roughly $660–690 billion of capital expenditure in 2026, while AI-attributable cloud revenue is estimated at only ~$150 billion annualised, and OpenAI alone signed (and later partly walked back) infrastructure commitments once totalling ~$1.4 trillion.
The bull case holds that demand is genuine, productivity gains are showing up, and capacity is the only constraint. The bear case — argued most forcefully by Ed Zitron — holds that the leading model labs cannot pay for what they have promised, that much of the “demand” is recycled vendor money, and that depreciation accounting is flattering Big Tech’s profits. Both camps are working from largely the same numbers; they disagree about what those numbers mean. This page lays them out.
To stay objective it helps to state each side at its strongest.
| The bear case (Zitron et al.) | The bull case (industry / many analysts) |
|---|---|
| The labs are structurally unprofitable: OpenAI posted a ~122% negative non-GAAP operating margin in Q1 2026 and burned ~$17B in 2025. They have promised ~$1 trillion+ in compute they cannot fund from cash flow. | Revenue is compounding faster than almost any software category in history; gross margins on inference are improving as chips and models get more efficient; losses are a deliberate land-grab. |
| Demand is partly an illusion created by circular financing — NVIDIA funds OpenAI, which pays Oracle/CoreWeave, which buy NVIDIA chips. Revenue is being “round-tripped”. | Vendor investment is normal in capital-intensive industries; end-user usage (tokens, paid seats, Claude Code, Copilot) is large and growing independently of those deals. |
| MIT found 95% of enterprise GenAI pilots show no measurable P&L impact; depreciation on fast-ageing GPUs is understated, flattering hyperscaler profits by 20%+ (Michael Burry). | PwC/Reuters surveys show 80%+ of adopters reporting revenue impact; the 5% that integrate well capture outsized value, and chips remain useful for inference well past their depreciation schedule. |
| The whole complex now drives a third of the S&P 500 and a large share of US GDP growth — a correction would hit the real economy (power, construction, credit), not just tech stocks. | The build-out is a genuine general-purpose-technology cycle (like rail, electricity, fibre); over-building is the historical norm and leaves durable infrastructure behind. |
2024 set the template: explosive top-line growth, equally explosive losses, and the first signs that capital expenditure would dwarf revenue. OpenAI reported roughly $3.7 billion in revenue and a net loss around $5 billion; Anthropic exited the year at roughly a $1 billion run-rate (from under $100 million in January). NVIDIA was the clearest winner: data-centre revenue more than doubled to $115.2 billion for fiscal 2025 (which covers most of calendar 2024), up 142%. Alphabet’s capex rose to about $52 billion, and combined big-builder capex climbed from roughly $160 billion (2023) toward ~$230 billion. In September 2024 Ed Zitron published “The Subprime AI Crisis”, the essay that framed the bear case for the next two years.
2025 was when the numbers turned vertical and the warnings turned official. OpenAI revenue roughly tripled to $13 billion (H1 revenue $4.3 billion against a $7.8 billion operating loss), while cash burn reached ~$17 billion. Anthropic climbed from ~$1B to a ~$9 billion run-rate. NVIDIA’s data-centre revenue kept setting records, reaching $51.2 billion in the quarter ended October 2025 alone. The defining events were a wave of enormous, interlocking deals: the $500 billion Stargate programme (January), NVIDIA’s up-to-$100 billion investment in OpenAI (September), OpenAI’s $300 billion Oracle cloud commitment, and chip pacts with AMD (6 GW, plus warrants for ~10% of AMD) and Broadcom (10 GW). In August, MIT’s “GenAI Divide” report found 95% of enterprise pilots delivering no P&L impact. By October the IMF and Bank of England were warning of bubble risk; in November Michael Burry disclosed a ~$1.1 billion bet against NVIDIA and Palantir and accused Big Tech of understating GPU depreciation.
The first five months of 2026 brought both vindication and doubt. Funding hit new extremes: Anthropic raised a $30 billion Series G at $380 billion (February) and was reportedly in talks for a further round near a $900 billion valuation by May, its run-rate vaulting past OpenAI’s to ~$30–45 billion. OpenAI closed a record $122 billion round at an $852 billion valuation (late March). But the same months exposed the strain: in February OpenAI quietly reset its compute target from ~$1.4 trillion to ~$600 billion through 2030, and by May was reported to be renting capacity it had vowed to build. Q1 2026 hyperscaler capex alone reached a combined $112 billion in a single quarter, with full-year guidance lifted to $650–700 billion. CoreWeave secured investment-grade-rated infrastructure debt (an $8.5 billion facility in March, a $3.1 billion facility in May). The bull and bear cases both got louder — and the gap between commitments and revenue is now the single most important number in the market.
The revenue story is genuinely remarkable and is the strongest plank of the bull case. Two trajectories stand out: OpenAI’s steady climb and Anthropic’s near-vertical 2026, which saw it overtake OpenAI on a run-rate basis around April 2026 — largely on the back of enterprise API usage and Claude Code.
Profitability is the counterpoint. OpenAI’s losses grew alongside revenue; the company has told investors it expects cumulative losses of roughly $115 billion through 2029 before turning profitable, with a $14 billion loss pencilled in for 2026. Anthropic, notably, is reported to spend roughly four times less on training than OpenAI for comparable revenue — one reason its unit economics look better to bulls.
If revenue is the bull case, spend is where the bear case lives. Capital expenditure by the largest cloud and AI builders has roughly quadrupled since GPT-4, from ~$160 billion in 2023 to guidance of ~$660–690 billion for 2026. The 2026 figures: Amazon ~$200 billion, Alphabet $175–185 billion, Meta $115–135 billion, Microsoft $120 billion+, Oracle ~$50 billion.
The clearest beneficiary remains NVIDIA, whose data-centre revenue is the most reliable public proxy for real spend hitting the ground.
A defining and contested feature of 2025–26 is the web of deals in which the same dollars appear to circulate. NVIDIA invests in OpenAI; OpenAI commits to renting cloud capacity from Oracle and CoreWeave; those providers buy NVIDIA GPUs. Critics call this “circular financing” or “round-tripping” that inflates apparent demand; NVIDIA counters that its stake is equity and the funds are not earmarked to buy its own chips. The truth is somewhere between: end-user demand is real, but the deals do make the ecosystem’s growth look more self-sustaining than independent demand alone would support.
| Deal | Headline size | Announced | Why it matters |
|---|---|---|---|
| Stargate (OpenAI, SoftBank, Oracle) | $500B | Jan 2025 | Flagship US data-centre programme; later partly deferred/rented |
| NVIDIA → OpenAI | up to $100B | Sep 2025 | Equity for ~10 GW of NVIDIA systems; lightning rod for circularity claims |
| OpenAI → Oracle (cloud) | $300B | Sep 2025 | ~$60B/yr 2027–31; the payment OpenAI must fund first |
| OpenAI ↔ AMD | 6 GW | Oct 2025 | Warrants for ~160M AMD shares (~10%) at $0.01 |
| OpenAI ↔ Broadcom | 10 GW | 2025 | Custom accelerators by 2029 |
| Meta → CoreWeave | ~$21B | Apr 2026 | Cloud capacity through 2032 |
As equity alone could not fund the build-out, 2025–26 saw a pivot to debt and off-balance-sheet structures. CoreWeave became the emblem: a $2.6 billion term loan (July 2025) tied to an OpenAI contract, an $8.5 billion facility in March 2026 that earned the first investment-grade rating for HPC-infrastructure-backed debt (A3/A low), and a $3.1 billion publicly syndicated facility in May 2026. Hyperscalers increasingly use special-purpose vehicles and bond issuance to keep data-centre liabilities off the core balance sheet. This is the part of the market the IMF and Bank of England flagged: if AI revenue disappoints, the losses would be borne not only by tech-equity holders but by the credit markets and the firms’ lenders — the channel that makes a “subprime” analogy resonate.
This is the crux. AI infrastructure capex in 2026 is guided near $675 billion, while AI-attributable cloud revenue is estimated at only ~$150 billion annualised. For a 25% return on AI-specific capex, analysts estimate the industry would need to generate roughly $169 billion of AI revenue annually by end-2028 — achievable on current growth, but with little margin for error.
Layered on top is an accounting dispute. Michael Burry argues GPUs really last ~2–3 years but are depreciated over 5–6, understating depreciation by ~$176 billion across 2026–28 and overstating hyperscaler profits by 20%+. If he is right, reported Big-Tech earnings are flattered precisely as capex peaks; if the bulls are right, the chips keep earning inference revenue long after they are “written down”, and the conservative books understate true returns.
The financial stakes are no longer confined to a few private labs. The “Magnificent Seven” reached about 34.8% of the S&P 500 by May 2026, with NVIDIA alone at ~7.0%. AI-linked capex has become a measurable contributor to US GDP growth, and data centres consume roughly 4% of US electricity. That concentration is why institutions warn that an AI repricing would be a macro event, not a sector one.
| Calendar quarter | NVIDIA fiscal quarter | Data-centre revenue | YoY |
|---|---|---|---|
| Q4 2023 | Q4 FY24 | $18.4B | +409% |
| Q1 2024 | Q1 FY25 | $22.6B | +427% |
| Q2 2024 | Q2 FY25 | $26.3B | +154% |
| Q3 2024 | Q3 FY25 | $30.8B | +112% |
| Q4 2024 | Q4 FY25 | $35.6B | +93% |
| Q3 2025 | Q3 FY26 | $51.2B | +66% |
Full-year data-centre revenue: FY24 $47.5B; FY25 $115.2B (+142%).
| Metric | 2024 | 2025 | 2026 YTD / run-rate |
|---|---|---|---|
| OpenAI revenue | $3.7B | $13B | ~$24B run-rate |
| OpenAI net loss / burn | ~$5B | ~$17B | $14B (proj.) |
| Anthropic run-rate (exit) | ~$1B | ~$9B | ~$30–45B |
| Anthropic valuation | — | $183B (Sep) | $380B→~$900B (talks) |
| OpenAI valuation | ~$157B | ~$500B | $852B (Mar) |
| Year | Combined capex | Notes |
|---|---|---|
| 2023 | ~$160B | Pre-boom baseline |
| 2024 | ~$230B (est.) | Alphabet alone $52B |
| 2025 | ~$300–340B | MSFT ~$80B, Alphabet $75B, Meta $60–65B, Amazon ~$105B |
| 2026E | ~$660–690B | Guidance; Q1 alone ~$112B |
What can the debt, revenue, spend and investment data plausibly tell us about where the market sits at the end of 2026? Below are three scenarios. They are conditional — deliberately neither a copy of the bear thesis nor uncritical optimism — and each names the indicator that would confirm it. The single pivotal variable in all three is OpenAI’s ability to fund its 2027 Oracle payments and the market’s willingness to keep financing the gap.
Revenue keeps compounding: OpenAI exits 2026 near a $30 billion run-rate and Anthropic above $50 billion, with enterprise seats and coding agents (Claude Code, Copilot, Codex) proving sticky. Inference costs keep falling, lifting gross margins. Capex is high but increasingly debt-financed against real, contracted cloud demand, and the IPO window (a possible OpenAI listing) re-liquefies the sector. Markets treat the capex–revenue gap as an investment phase, not a warning. Christmas 2026: equities hold near highs, no major lab insolvency, the narrative is “build-out, not bubble”. Watch for: AI-cloud revenue clearly tracking toward the ~$169B/yr needed by 2028.
Growth stays strong but decelerates and dispersion widens: a clear tier of winners (NVIDIA, Anthropic, the big clouds) pulls away while weaker AI start-ups fold and some pilots are quietly cancelled. OpenAI continues to defer or rent capacity it once promised to build, and at least one high-profile commitment is renegotiated. A sharp but contained correction (perhaps 15–25% in the most AI-exposed names) is plausible around an earnings disappointment or a financing scare, followed by stabilisation. Debt markets grow more selective; investment-grade names (CoreWeave-style) keep raising, speculative ones do not. Christmas 2026: the sector is intact and still growing, but visibly more disciplined and more concentrated, with “show me the ROI” the dominant theme. Watch for: capex guidance being trimmed rather than raised, and widening credit spreads on data-centre debt.
The Zitron-style failure mode: a funding round stalls or a large customer cannot meet a payment, puncturing confidence. Because demand was partly sustained by circular deals, a single failure at a key node (e.g. OpenAI’s ability to begin paying Oracle, or a deferred Stargate phase) cascades. Hyperscalers cut capex guidance, NVIDIA orders soften, GPU depreciation write-downs hit earnings, and the concentrated index drags the broad market down with knock-on effects in power, construction and the credit that funded the build-out. Christmas 2026: a meaningful drawdown in AI-linked equities, at least one distressed restructuring among the model labs or neoclouds, and a freeze in speculative AI debt. Watch for: a failed or down-round raise, a missed contractual payment, or a hyperscaler explicitly pausing data-centre projects.
On the present data, the middle scenario is the most defensible: revenue is real and large enough to make a total collapse unlikely in 2026, but the commitment-to-revenue gap, the reliance on circular and debt financing, and the index concentration are large enough that a sharp, possibly violent repricing is well within range. The optimists are right that the technology and demand are real; the bears are right that the financing structure is fragile. Both can be true at once — which is exactly why the next two quarters of funding and capex announcements matter more than any benchmark score.
This page synthesises reporting and data from, among others:
Figures are drawn from public reporting and company disclosures as of 29 May 2026. Private-company revenue, losses and valuations are press/investor estimates, not audited. Run-rates, projections and capex guidance are explicitly labelled. Charts are illustrative; where a data point is interpolated or projected it is noted in the caption.