When Will the AI Bubble Burst? The Evidence For and Against (October 2026)

The gap between what hyperscalers are installing and what AI is currently earning is no longer a 2024 thought experiment. As of mid-October 2026, the five named buyers are guiding to roughly $800 billion of 2026 capex on Goldman’s synthesis, while the cleanest “above-trend” cloud revenue print is only about $70 billion annualized. Lab run rates have exploded, but much of that money is the same dollar counted twice—once at Anthropic or OpenAI, once at AWS, Azure, or Google Cloud.

Q3 2026 earnings for Meta (Oct 21), Microsoft and Alphabet (Oct 28), and Amazon (~Oct 29) had not yet landed as of Oct 20, so full-year figures below are company guidance through the July earnings cycle, not post-Q3 actuals.

2026 capex: guidance kept rising, except where accounting moved the number

The binding constraint is no longer “will they spend?” It is which definition of spend you use. Amazon’s cash capex, Alphabet’s property-and-equipment, Microsoft’s calendar-year figure including finance leases, Meta’s capex plus finance-lease principal, and Oracle’s gross versus customer-funded net are not the same instrument. Memory prices alone forced Amazon up $20 billion in late July; Microsoft’s number moved the other way because it lengthened assumed data-center lives and reclassified some leases as operating.

Latest company guidance for 2026 (calendar year unless noted):

• Amazon: about $220 billion cash capex, raised July 30 from about $200 billion because of higher memory costs. CEO Andy Jassy has said capacity will still fall short of demand through 2027. Q2 cash capex was $53–54 billion.
• Alphabet: $195–205 billion, raised July 22 from $180–190 billion because capacity is being delivered faster, not because the plan got more ambitious in scope. Q2 capex was $44.9 billion. Trailing free cash flow turned negative in the quarter.
• Microsoft: about $175 billion for calendar 2026 including finance leases, revised down from roughly $190 billion after useful lives on data centers and buildings were extended (some leases left finance-lease capex). Fiscal Q4 (ended June 30) capex was about $41 billion including leases. Azure demand still exceeds capacity.
• Meta: $130–145 billion including principal payments on finance leases, narrowed in late July from a wider range that had started the year at $115–135 billion. Q2 capex was about $31 billion; quarterly free cash flow collapsed.
• Oracle: fiscal 2027 (June 2026–May 2027) gross capex $90–95 billion, with management capping net cash capex at about $70 billion via customer prepayments, supplier financing, and “bring your own hardware.” Fiscal Q1 (ended Aug 31) capex was $28.5 billion. This is not a calendar-2026 number and is only partly comparable to the others.

Big-four midpoint guidance clusters around $720–745 billion. Adding Oracle on a cash or AI-infrastructure basis, and treating a slice of non-AI Amazon spend (fulfillment, satellites, robotics) as outside the AI stack, is why bank totals land near $750–800 billion for 2026 rather than a simple sum of headlines. Goldman’s September baseline is about $800 billion of AI infrastructure for the five in 2026, then $1.2 trillion in 2027 and about $1.4 trillion in 2028—above consensus, with growth decelerating after 2026. Quarterly cash prints already show the run rate: the five spent $148.4 billion in Jan–Mar 2026 and $181.5 billion in Apr–Jun 2026.

Not all of Amazon’s $220 billion is AI compute. Treating the headline as pure AI overstates the GPU-and-power bill and understates how much of the “AI trade” is still a general data-center and logistics cycle.

What is actually being earned

Reported AI revenue is real and, in absolute dollars, historically fast. It is also defined so differently that adding the headlines produces a false total.

Labs (the demand that shows up as someone else’s capex):
- OpenAI told investors annualized revenue was approaching $50 billion at the end of September 2026 (its own, net-of-partner basis). A ~$70 billion figure circulating a week earlier was investors grossing OpenAI up onto Anthropic’s accounting. Booked full-year 2026 revenue is expected around $35–36 billion. Q2 quarterly revenue was about $6.7 billion.
- Anthropic reported $16.8 billion of Q3 2026 revenue (about $67 billion annualized), up 45% from Q2’s roughly $11.5–11.6 billion, and a $65 billion+ run rate at the end of July on a gross basis that includes sales billed by Amazon and Google. Full-year 2025 revenue was $4.6 billion. It has reported positive adjusted free cash flow recently and is preparing an IPO. Partner sales were a large share of 2025 usage revenue, so a big slice of “Anthropic revenue” is also AWS or Google Cloud revenue.

Cloud AI and adjacent disclosures:
- Microsoft: AI business annual revenue run rate surpassed $37 billion as of the April 2026 call (+123% year over year), covering Azure AI, M365 Copilot, GitHub Copilot, and related products. No later run-rate update appeared in the July materials reviewed here. M365 Copilot exceeded 30 million paid seats by July (third-party trackers put seats near 36 million by September). Azure revenue surpassed $100 billion in fiscal 2026; June-quarter Azure growth was 43%, and fiscal Q1 guidance was 44–45% constant currency. Commercial RPO was $678 billion at fiscal year-end.
- Amazon / AWS: both the AI business and the custom-chip business (Trainium, Graviton, Nitro) each crossed a $25 billion annual run rate in Q2 2026, triple-digit growth. Earlier, the April shareholder letter put AI services above $15 billion annualized on Q1. AWS segment sales were $42.2 billion in Q2 (+37%), a ~$169 billion run rate, with backlog about $496 billion. Capacity on Trainium is described as sold out for the next year.
- Alphabet / Google Cloud: $24.8 billion in Q2 (+82%), operating margin 35.6%, operating income more than tripled to $8.8 billion. Cloud backlog $514 billion ($513.9 billion in the 10-Q), with just over half expected to be recognized within 24 months. Gemini app at 950 million monthly users. This is total Cloud, not a pure “AI revenue” line; management attributes the acceleration primarily to GCP and enterprise AI.
- Oracle: cloud infrastructure revenue $7.39 billion in fiscal Q1 2027 (+121%). Remaining performance obligations $664 billion. Only about 13% of backlog is expected to convert within 12 months; power and construction, not orders, are the constraint.
- Meta: almost no external AI revenue line. The return is supposed to show up as higher ad yield and engagement, not as a cloud invoice. That makes Meta the hardest name in any payback table and the one most often omitted from “AI revenue” sums.

Goldman estimates hyperscaler cloud revenue was running about $70 billion above the pre-AI trend as of Q2 2026. That incremental figure—not the sum of lab run rates plus total cloud—is the right numerator if the question is “has new AI demand appeared inside the clouds?”

How the famous frameworks differ

These estimates answer different questions. Stacking them as if they were one forecast is the main source of confusion.

Sequoia / David Cahn — lifetime end-market revenue, not this year’s P&L. The original “$200B question” (2023) and “$600B question” (June 2024) used the same identity: Nvidia data-center revenue × 2 (GPUs are about half of data-center cost) × 2 (a 50% gross margin for the cloud buyer and the application company). He updated that to about $840 billion in June 2025 and to AI’s $1.5 trillion question in July 2026. Two derivations give the same order of magnitude: Nvidia’s projected data-center run rate × 4, or ~$750 billion of 2026 hyperscaler data-center capex × 2. A separate cumulative figure—on the order of $3 trillion of revenue since ChatGPT to justify the whole build—is what some write-ups call the “$3 trillion question.” This is lifetime revenue attached to a year of spend, not an annual run rate the industry must hit in 2026. Hardware lives of three to six years turn $1.5 trillion lifetime into something like $250–500 billion a year, which is why Cahn’s number can look either catastrophic or merely demanding depending on the amortization.

Goldman Sachs (Ryan Hammond, September 2025 note, still the reference in mid-October coverage) — annual depreciation-and-opex breakeven for the hyperscalers. On average annual 2026–27 AI capex, the hyperscalers need about $308 billion a year in AI revenue to break even (depreciation plus operating cost), and $417–526 billion for a 10–20% return on capital. The simpler headline used in the same workstream is roughly $300 billion to stop losing money on an ~$800 billion 2026 spend, against ~$70 billion of cloud revenue above the pre-AI trend—a ~$230 billion annual gap. For “solid” returns to both infrastructure owners and the application layer, Goldman wants end users spending about $1 trillion a year on AI applications. Scope is hyperscaler P&L, not the full Nvidia-to-app margin stack, and the horizon is the next few years, not 2031.

Bain (September 2026 Global Technology Report) — a 2031 industry-revenue identity. Hyperscaler capex (Microsoft, Google, Amazon, Meta, Oracle) is put at about $780 billion this year. Annual AI infrastructure spending could reach $1.5 trillion by 2031. If capex stays about 25% of industry revenue (Bain’s “ambitious but reasonable” cloud-like ratio), that requires an AI market approaching $6 trillion a year by 2031. Existing consumer and enterprise AI might supply $1.2–1.8 trillion; the other ~$4.2 trillion has to come from new categories (ads inside assistants, “autonomous everything,” physical AI/robotics, and similar). This is not a 2026 payback test. It is a claim that today’s build only pencils if AI becomes a multi-trillion-dollar slice of GDP growth, on the order of +1 percentage point, by the early 2030s.

JPMorgan is primarily a capex and financing forecast, not a revenue-gap identity in the sources available. Mid-2026 materials put global AI and data-center spending at at least $5–5.5 trillion through 2030 (ceiling cited near $7 trillion in some notes), with hyperscaler capex around $650 billion in 2026 and above $1.1 trillion in 2027, and operating cash flow still projected above $900 billion by 2027. The bank’s emphasis is that the cycle can be financed—investment-grade bonds and leveraged loans—while free cash flow at the hyperscalers compresses, not that a specific ARR hurdle has been missed.

A useful cross-check from independent compilation work (not a bank): counting only 2026–27 capex implies on the order of $400–500 billion a year of cloud AI revenue; counting spend since 2023 pushes the hurdle toward $600–750 billion; folding in 2028 capex pushes it toward $900 billion+. Those bands sit between Goldman’s breakeven and Bain’s 2031 fantasy, and they are why “the” required-revenue number moves by hundreds of billions depending on the start year.

Capex, visible AI revenue, and what payback assumes

Figures are the latest public or widely cited estimates as of October 20, 2026. AI revenue is not additive across rows.

Company	2026 capex (latest guidance)	Best public AI / cloud revenue marker	What a simple payback assumes
Amazon	~$220B cash capex (raised Jul 30)	AWS AI run rate >$25B; chips >$25B; AWS segment ~$169B annualized, +37% in Q2	Most 2026 capex monetized in 2027–28; Jassy has said a large share already has customer commitments. Company-wide number includes non-AI.
Alphabet	$195–205B	Google Cloud $24.8B in Q2 (+82%), ~$99B annualized; backlog $514B, >50% in 24 months	Backlog is the payback claim. Supply, not orders, is the constraint. FCF already negative in Q2.
Microsoft	~$175B calendar 2026 incl. finance leases	AI business run rate >$37B (Apr 2026); Azure >$100B in FY26; Copilot >30M paid seats	April AI run rate is stale relative to 40%+ Azure growth. RPO $678B, but only ~25% expected to be recognized in the current year.
Meta	$130–145B incl. finance-lease principal	No external AI revenue line; return is internal ads/engagement	Payback is an ROI-on-ads assumption, not a cloud invoice. Least comparable in this table.
Oracle	FY27 gross $90–95B; net cash capex capped ~$70B	OCI $7.4B in quarter ended Aug 31 (+121%); RPO $664B	Only ~13% of backlog in the next 12 months. Customer funding lowers cash capex but does not remove execution and power risk.
Five combined	~$750–800B AI-infrastructure synthesis (big four alone $720–745B)	Incremental cloud vs pre-AI trend ~$70B; lab run rates ~$50B OpenAI net + ~$65–70B Anthropic gross, heavily overlapping the clouds	Goldman: ~$300–308B/year to break even; $417–526B for 10–20% ROC. Cahn: ~$1.5T lifetime end revenue per year of ~$750B DC spend.

How estimates differ, in one place:
- Stock vs flow. Cahn’s $1.5 trillion (and the older $600 billion) is lifetime end-customer revenue required to carry a year of infrastructure cost through a 50% margin stack. Goldman’s $300–308 billion is an annual hyperscaler revenue level that covers depreciation and opex. Bain’s $6 trillion is a 2031 industry revenue level implied by treating $1.5 trillion of annual infra spend as 25% of sales.
- Whose revenue. Goldman and the “above-trend $70 billion” measure money the clouds keep. Cahn measures money end customers must eventually pay, after cloud and app margins. Adding OpenAI + Anthropic + Azure AI + AWS AI + Google Cloud double-counts the same training and inference invoices.
- Gross vs net at the labs. Anthropic books gross partner sales; OpenAI books its net share. The $20 billion gap between “~$70 billion” and “approaching $50 billion” for OpenAI in early October is mostly that definition, not a sudden demand collapse.
- What is in capex. Amazon’s $220 billion is company-wide cash capex. Microsoft’s $175 billion moved because of lease accounting, not because it installed fewer GPUs. Oracle’s $90–95 billion gross is not $90–95 billion of Oracle cash. Goldman’s $800 billion is an AI-infrastructure cut of the five, not the sum of press-release midpoints.
- Useful life. Extending server and building lives (Microsoft did this explicitly) lowers near-term depreciation and flatters current ROIC while pushing the true payback further out. Cahn has also noted that memory and power bottlenecks have raised required revenue per gigawatt, so the old “GPUs are half of cost” rule is getting less generous.

What this means if you are underwriting the cycle

The bull case that is actually in the filings is not “AI revenue will be $6 trillion.” It is that signed backlog already exceeds a trillion dollars at Microsoft, Google, Amazon, and Oracle combined, GPU utilization at Oracle is described near 98%, Trainium is sold out, and cloud growth has reaccelerated (AWS +37%, Google Cloud +82%, Azure low-40s, OCI +121%) while the industry is still capacity-constrained. That is demand. It is not yet a return.

The bear case that survives contact with the same filings is cash-flow timing. Aggregate hyperscaler capex is on track to exceed operating cash flow around this quarter on simple trend extensions; Amazon’s trailing free cash flow was already about –$7.6 billion at June 30, Alphabet’s quarterly FCF was negative, and Oracle’s was negative despite record operating cash flow. Debt and equity issuance are filling the gap. If power, transformers, or memory stretch delivery so that 2026–27 capex is recognized as depreciation before the backlog converts—Oracle expects only ~13% of RPO inside a year, Microsoft about 25%—the ROIC math Goldman sketches ($300 billion to break even, half a trillion for a mid-teens return) does not get a free pass just because orders exist.

For anyone competing or allocating against this: the scarce asset is no longer a model API. It is energized, liquid-cooled megawatts with a customer who has already signed, on a depreciation schedule the equity market will tolerate while free cash flow is negative. The $600 billion question did not get answered. It was marked up to $1.5 trillion of lifetime revenue, while the annual test the hyperscalers themselves face is closer to Goldman’s $300 billion—and the cleanest current evidence of new cloud AI revenue is still about a quarter of that.

---

Recent Findings Supplement (October 2026)

Hyperscaler 2026 capex guidance has been revised upward in mid-2026 earnings calls (primarily July), with aggregates now commonly cited at $700–780B+ for the major players (Microsoft, Alphabet, Amazon, Meta, Oracle), driven overwhelmingly by AI infrastructure.[1][2]

This reflects faster capacity delivery, higher memory/component costs, and sustained demand. Q2 2026 cash capex for the four largest (MSFT, GOOG, META, AMZN) totaled $165.1B, with year-over-year growth of 68–110%.[3] Forward projections from analysts (e.g., Goldman, TrendForce, Vanguard) point to $800B–$957B+ ranges when including broader sets or leases, with 2027 forecasts climbing to $1.1–1.58T.[4][5]

• Amazon: Raised to ~$220B cash capex for calendar 2026 (July 30, 2026 call; cited higher memory costs and AWS/AI demand). Q2 cash capex $53.1B.[1]
• Alphabet: Raised to $195–205B (July 22, 2026; from prior $180–190B). Q2 capex $44.9B; expects significant further increase in 2027.[1]
• Meta: Narrowed to $130–145B (July update). Q2 ~$31B (incl. leases).[6]
• Microsoft: ~$175B calendar 2026 (incl. finance leases; July 29, 2026; prior ~$190B reclassified). FY Q4 capex $41B (incl. leases). Strong Azure AI demand noted.[1]
• Oracle: FY2026 (ended ~May 2026) $55.7B (above prior ~$50B target). FY2027 guidance $90–95B gross / ~$70B net cash outlay (customer prepayments offset some). Q1 FY2027 capex $28.5B.[7]

Implications for competitors/entrants: The scale creates barriers via power, land, and supply-chain access but also opportunities in financing models (prepayments, leases, “bring-your-own-hardware”), specialized cooling/power, or niche capacity. Negative or sharply lower free cash flow is now expected for several into 2027, pressuring balance sheets and buybacks.[2]

AI-specific revenue run-rates for hyperscalers and labs show strong growth but remain far below the annual figures implied by sustained capex levels. Microsoft’s AI business reached a $37B+ annualized run-rate by Q3 FY2026 (ended ~March/April 2026; +123% YoY), including Azure AI, Copilot, etc.[8] Microsoft 365 Copilot paid seats reached ~30M+ by July 2026 and ~36.2M by September 2026.[9]

OpenAI’s annualized revenue run-rate was reported near $50B at end-September 2026 (with expectations of $70B+ by year-end, driven by enterprise); Anthropic reached $65B by end-July 2026 (accounting differences: Anthropic includes gross partner/cloud sales; OpenAI reports its net share).[10][11]

• Cloud revenue growth for hyperscalers has accelerated (e.g., Microsoft Azure ~40%+ YoY in recent quarters; overall cloud ~48% in Q2 2026 per some reports), running ~$70B above pre-AI trends in aggregate analyses.[12]
• Labs remain concentrated: OpenAI and Anthropic dominate reported figures, with enterprise and inference usage as key drivers.

Implications: Early monetization traction (especially Copilot seats and Azure AI) validates demand but covers only a small fraction of the infrastructure spend. Application-layer and new use-case revenue will need to scale dramatically to close gaps.

Sequoia’s updated “$600B question” (now often framed as $1.5T+), Bain’s $6T requirement by 2031, and JPMorgan/Goldman-style estimates highlight a widening revenue gap relative to capex.[13]

Sequoia partner David Cahn’s framework (capex ÷ (1 – 50% margin)) applied to ~$825B 2026 hyperscaler guidance implies ~$1.65T in annual end-customer revenue needed if spending recurs yearly; earlier versions were ~$600B.[13]

Bain (Sep 29, 2026 Global Technology Report): Annual AI infrastructure spend could reach $1.5T by 2031; at ~25% capex-to-revenue ratio, this requires ~$6T annual AI market revenue. Existing consumer/enterprise products max out at $1.2–1.8T, leaving a ~$4.2T gap that demands entirely new markets (e.g., physical AI/robotics ~$900B potential, autonomous systems, ads in AI).[14]

JPMorgan/Dimon and Goldman estimates: 2026 hyperscaler/ecosystem capex ~$700–800B, rising to ~$1–1.5T+ in 2027; required annual revenue figures cluster around $400–700B+ depending on scope (recent cohort vs. cumulative investment since 2023), assumed returns (8–10% ROIC or breakeven), and depreciation. Goldman models breakeven ~$300B and 10% ROIC ~$417B from infra alone, with apps needing ~$1T.[12][15]

Implications: These frameworks assume different time horizons, margins, and whether capex is one-time or recurring. Payback is most sensitive to sustained utilization, pricing power, and new revenue categories beyond productivity gains. Decelerating capex growth could ease pressure if revenues accelerate.

A consolidated view of 2026 capex guidance, illustrative AI revenue estimates, and payback implications (mid-2026 to Oct 2026 data; figures are approximate aggregates/guidance and vary by source methodology):

Company/Aggregate	2026 Capex Guidance (latest)	Est. AI/Related Revenue Run-Rate (recent)	Implied Payback/Notes (framework lens)
Amazon	~$220B cash	AWS AI/cloud contribution (part of broader growth)	Contributes to aggregate ~2x multiplier needs
Alphabet	$195–205B	Google Cloud AI growth	Significant 2027 ramp expected
Microsoft	~$175B (calendar, incl. leases)	AI business ~$37B+ ARR (Q3 FY26); Copilot ~30–36M seats	Strongest near-term monetization via Copilot/Azure
Meta	$130–145B	AI-driven ad/recommendation uplift	Efficiency focus noted
Oracle	FY26 $55.7B; FY27 ~$90–95B gross/~$70B net	Cloud infra growth (121%+ YoY in Q1 FY27)	Heavy reliance on prepayments/backlog ($664B RPO)
Big 4–5 Aggregate	~$700–780B (Vanguard/others); up to ~$825B midpoints	MSFT AI + labs (OpenAI ~$50–70B, Anthropic ~$65B+) + cloud AI uplift	Sequoia-style: $1.65T+ annual rev needed if recurring; Bain: path to $6T by 2031 with large gap; JPM/GS: $400–700B+ annual threshold depending on returns/scope

Sources for table elements: Company earnings/guidance via analyst summaries (July–Oct 2026); aggregates from Vanguard, Goldman, etc.; revenue from earnings calls and reports (e.g., MSFT Q3 FY26, OpenAI/Anthropic updates Oct 2026).[1][2]

Key differences in estimates: Cash vs. including leases/finance; gross partner sales (Anthropic-style) vs. net share (OpenAI); one-year cohort capex vs. cumulative/multi-year; assumed gross margins (commonly 50% in Sequoia-style math) and depreciation (influences annual “need”); whether revenue must cover only infra or broader ecosystem. Single-source financials are common for company guidance but aggregates benefit from cross-verification. Q3 2026 earnings (late Oct 2026) may provide further updates. Data is post-April 2026 and prioritized from recent publications.

The filings and Burry’s numbers are the core of the debate. I’m pulling the company-level accounting changes and analyst earnings sensitivities next.
The fight is not whether a GPU still powers on in year six. It is whether straight-line depreciation over five to six years matches the period the chip earns an attractive return for its owner. Those clocks have diverged, and the gap is now large enough to move reported earnings by billions of dollars a year.

Five clocks, one argument

GAAP does not prescribe a number of years. It requires management to depreciate an asset over the period it is expected to provide service to that owner. Auditors test the estimate; a revision is applied prospectively, not by restating the past. At least five different durations are being collapsed into “useful life”: physical life until failure, rental life at any price, service life to the first owner, the book life management picks, and the contract term that actually pays for the box. A chip can be fully utilized in year seven and still have lost most of its economic value by year three, because newer silicon does the same work at a fraction of the power and token cost. Straight-line accounting ignores that shape. It charges the same dollar every year while the cash the card generates is heavily front-loaded.

---

Recent Findings Supplement (October 2026)

Recent developments (primarily July–October 2026 filings, disclosures, pricing data, and commentary) show the GPU depreciation debate intensifying around real-world utilization and pricing evidence rather than purely theoretical obsolescence. Older NVIDIA GPUs (e.g., A100 from 2020 and H100 from 2022–2023) continue generating revenue through long-term contracts and rising rental rates, while hyperscalers show mixed accounting adjustments. Michael Burry maintains his critique, but counter-evidence from contracts, price hikes, and utilization data has prompted rebuttals.[1][2]

Michael Burry continues pressing the case for 2–3 year economic lives versus 5–6 year book lives, estimating ~$176 billion in understated industry depreciation from 2026–2028, with specific calls on companies like Oracle (~27% earnings overstatement) and Meta (~21%). He has linked recent commentary to historical parallels (e.g., Cisco-era dynamics) and highlighted risks around financing structures and potential insurance for neocloud defaults tied to resale values.[3][4]

• As of mid-October 2026, Burry remains active on the topic via X and Substack references, framing it as a question of whether accounting lives match economic reality amid rapid new-chip efficiency gains.[4]
• Steve Eisman rebutted the thesis in early October 2026, calling it “overly academic” and arguing strong demand (including for older chips) means depreciation period shifts (e.g., 3–4 to 5–6 years) would not materially harm outcomes if AI growth continues.[5]
• NVIDIA has directly engaged with a multi-page analyst memo and public data points (e.g., A100 utilization and contracts) disputing the 2–3 year claim.[6]

Hyperscalers show divergent 2025 accounting tweaks disclosed in 2026 filings, with Amazon taking a more conservative step on AI-driven tech pace while Meta extended lives. These remain the most recent quantified changes.[7][8]

• Effective January 1, 2025, Amazon shortened useful lives for a subset of servers and networking equipment from 6 to 5 years (citing faster AI/ML development), increasing 2025 depreciation by ~$1.4 billion and reducing net income by ~$1.0 billion (~$0.10 per diluted share).[8][9]
• Effective the same date, Meta extended most servers and network assets to 5.5 years, cutting 2025 depreciation by $2.92 billion and boosting net income by $2.59 billion (~$1.00 per diluted share).[10][8]
• Microsoft (FY2026 10-K, ended June 2026) and Alphabet report no further changes to their 2–6 year or generally 6-year server/network policies (prior extensions from ~4 years occurred earlier). CoreWeave holds at 6 years.[11][7]

Rental prices for H100/H200 (and older chips) have risen in mid-to-late 2026, and long-term contracts for legacy GPUs extend well beyond typical depreciation schedules, providing concrete counter-data to rapid-obsolescence claims.[2][12]

• H100 on-demand median prices reached ~$3.81/GPU-hour by mid-October 2026 (up 21% YoY and +9% in the prior four weeks across tracked providers); specific provider actions include CoreWeave ~25% across SKUs in July 2026, Nebius +17% on H100 (Oct 1, second hike in three months), and AWS reserved +20% in July (after +15% in January). H100 (2022-era) on-demand rose 16% vs. a year prior.[2][12]
• CoreWeave disclosed (Q2 2026 earnings and August updates) A100 contracts running into 2029 (nine years post-2020 launch) at attractive pricing; CEO noted prior-gen pricing “at or above where it was years ago,” with strong backlog. Similar extensions noted for other operators.[10][13]
• Independent estimates (October 2026) put certain B300 servers at ~33% of purchase price after two years on resale; older A100s retain meaningful projected rental income value.[11]

NVIDIA’s data-center product cadence remains rapid (annual rhythm referenced from GTC 2025), but evidence of multi-year utilization for prior generations (A100 through 2029, H100 price resilience) and consumer-gaming delays (RTX 60 series reportedly pushed to 2028 due to AI focus) highlight a split between inference/training workloads and gaming.[14][15]

• No major new cadence announcements in the latest period, but utilization data (e.g., seven-to-nine-year-old custom chips at high utilization, A100 fleets still mission-capable) and rental trends are cited as evidence against strict 2–3 year economic lives.[10][13]

Analyst and modeled sensitivities quantify material EPS effects from shorter lives, though actual impacts depend on capex cohorts and revenue realization. Recent modeling (referencing 2026 data) shows the stakes clearly.[16]

• JPMorgan sensitivity (early 2026 framing, still referenced): Shifting recent GPU/networking capex to a 3-year life (vs. 6) would reduce 2025 EPS by ~6–7% at Amazon/Meta/Microsoft, ~6% at Alphabet, and ~17% at Oracle.[16]
• On the 2026 hyperscaler server cohort alone (~$425 billion capex), a 3-year life implies ~$142 billion annual depreciation versus ~$71 billion at 6 years.[16]
• Extension benefits already realized (e.g., Meta’s 2025 change worth ~$2.59 billion net income) illustrate the reverse sensitivity; prior extensions lifted net income by ~4–4.6% in the affected years for several firms.[17]

For competitors or entrants, the debate underscores that accounting assumptions are management estimates under GAAP principles (service period to the owner), not fixed standards; sustained rental demand, long contracts, and price resilience for older silicon currently support longer effective lives in practice, but any slowdown in AI workloads could accelerate write-down risks or force faster cycles. Recent price and contract data provide the strongest empirical pushback to shorter-life claims. No significant regulatory updates on depreciation policies appear in the period reviewed.

The AI buildout has flipped from a cash-flow story into a credit story. Stress is already concentrated in the weakest links—Oracle, CoreWeave, and project bonds that only look safe because a hyperscaler name is on the lease—while the strongest balance sheets are carrying far more contingent leverage than ratings currently reflect.

The issuance wave, and the market’s first refusal

Hyperscalers stopped funding the buildout from free cash flow and started competing with governments for long-dated capital. The five large US cloud names (Alphabet, Amazon, Meta, Microsoft, Oracle) sold roughly $274 billion of investment-grade bonds in 2026 through late September, versus an average of about $30 billion a year in 2020–24. [1] The Economist puts borrowings by the four larger hyperscalers at $214 billion in 2026 alone, up from $83 billion in all of 2025, and puts total AI-adjacent issuance—including utilities and equipment suppliers—near $850 billion year-to-date, about four times the 2023 pace for the same group. [2] Broader AI-linked debt is reported around $450–466 billion in the first nine months of 2026, more than double the prior year. [3] [4]

The market absorbed that supply until it didn’t. Morgan Stanley data cited by the Financial Times shows global AI-related issuance collapsing from a record $113 billion in June to $23 billion in September. [4] When Amazon sold $15 billion of notes, investors pulled about 40% of orders once final pricing failed to clear a sufficient concession. [5] Long-dated hyperscaler paper has cleared with coupons above 6% as the 10-year Treasury moved above 5%. [6] S&P expects Amazon, Microsoft, Alphabet, Oracle, Meta, and SpaceX to spend more than $7 trillion on data centers and related AI capex from 2025 to 2030, with negative free cash flow at least through next year. [7] Gartner’s gap is starker: infrastructure spend through 2030 of about $7 trillion against data-center-attributable AI revenue of no more than $3.6 trillion, leaving multi-trillion-dollar asset values exposed to write-downs if utilization disappoints. [4]

What this means for anyone financing or competing in the stack: the investment-grade window is still open for AA names, but the concession has risen and the calendar is lumpy. Weaker credits are already being rationed.

Oracle is the fulcrum, not a sideshow

Oracle is the only large hyperscaler whose own rating and market pricing have broken. S&P cut it to BBB− in July—the lowest investment-grade rung—citing AI spending and an “uncertain path to profitability.” A further downgrade could be triggered by sustained leverage above 4x, failure to be on a path to positive free operating cash flow by fiscal 2029, or a negative view of the AI strategy. [8] [7] S&P has flagged a free-operating-cash-flow deficit near $42 billion for fiscal 2027. [9]

Reported debt totals differ by definition, but the direction is unambiguous. The Economist and NPR both put Oracle borrowings near $169 billion. [2] [10] Company-period figures show about $125 billion of notes and borrowings plus roughly $31 billion of operating lease liabilities as of August 31. [11] Capex for that quarter was $28.5 billion, up from $8.5 billion a year earlier. [11] Remaining performance obligations hit $664 billion, and roughly half of the backlog is OpenAI—more than $300 billion of contracted revenue from a single private counterparty. [12] [13]

Credit markets are already treating the paper as closer to high yield. Five-year CDS hit a record near 288 basis points on October 15 and was quoted around 295–315 basis points later that week, up roughly 300% year over year. [8] [14] [15] A 2031 cash bond has traded near a 6.23% yield, about 180 basis points over the five-year Treasury—levels more consistent with BB than with BBB−. [16] The 2055s have traded around 77 cents. [17]

The off-balance-sheet layer is larger still. Banks arranged about $38 billion of senior secured facilities for Vantage campuses in Texas and Wisconsin tied to Stargate, plus an $18 billion package for a New Mexico campus, with Oracle expected as tenant. [5] [18] Some of that project debt has traded near 90 cents (Texas/Wisconsin) and around 80 cents (New Mexico). [19] Oracle sent a force-majeure notice on the New Mexico project reserving the right to delay payments. [17] [10]

Implication: Oracle is the transmission mechanism between OpenAI’s funding and the investment-grade market. A stumble in OpenAI’s ability to pay, or another year of negative free cash flow, is what would force IG-only holders to sell.

Meta-style SPVs: the debt is real; the balance sheet is optional

Meta’s template is now a product. In the Hyperion (Louisiana) structure, Blue Owl funds own 80% of the joint venture and Meta 20%. Beignet Investor LLC, a bankruptcy-remote SPV, issued $27.3 billion of 6.581% senior secured notes due 2049, rated A+ by S&P largely because of Meta’s residual-value guarantee and construction-overrun support. PIMCO anchored roughly $18 billion; BlackRock took more than $3 billion. [20] [21] Meta leases the campus back in four-year operating-lease blocks with renewal options, which is how the asset and the debt stay off Meta’s balance sheet under US GAAP even though rating agencies and bond buyers treat the credit as Meta’s. [22] [23]

The sequel, El Paso, used the same 80/20 joint-venture shape with BlackRock’s Sopaipilla vehicle and cleared $12.5 billion of 7.534% notes due 2048—more than 90 basis points richer than Hyperion—rated A+/AA−. [24] [25] Secondary prices have already marked the structure down: Beignet notes fell from above par in early July to a little under 91 cents; Blackstone’s QTS “Project Excalibur” bonds leased to Microsoft slipped below 87 cents; Hut 8 project bonds leased in part to Nvidia traded near 91–93 cents. [26]

The footnote is the real leverage. Meta has disclosed about $347 billion of future lease obligations on projects that have not yet commenced, against roughly $84 billion of on-balance-sheet debt at end-2025. [27] One compilation of SEC filings puts Meta’s off-balance-sheet obligations near $420 billion and Oracle’s near $273 billion; uncommenced data-center lease commitments across five large spenders have been cited around $1.12 trillion. [28] [6] Tech firms have moved more than $120 billion of AI data-center spending off books via SPVs and leases. [3] S&P’s digital-infrastructure head said it is “safe to assume” there is more AI-linked debt than the agencies can see, because unrated and unreported interconnected guarantees do not show up in hyperscaler disclosures—and that the agency is not changing ratings yet. [7]

Implication: these structures do not remove risk; they relocate it to private-credit funds, insurance balance sheets, and the residual-value guarantee. If Meta walks at the first renewal, bondholders are left with a specialized building and a guarantee whose accounting treatment and economic force may not match.

Neoclouds and GPU loans: where distress is already priced

CoreWeave is the stress test private credit did not want. LSEG puts total debt near $50 billion. [29] Principal owed was $35.6 billion at June 30, 2026, plus $16.3 billion of operating lease liabilities. [30] [11] GPU-loan commitments signed or amended run to roughly $28 billion across delayed-draw facilities, with pricing that tracks the offtaker, not the chip: early Magnetar/Blackstone paper at SOFR plus 8.75–9.62%, Microsoft-backed draws much tighter, an $8.5 billion March 2026 facility at SOFR plus 2.25% anchored by Blackstone Credit & Insurance and rated A3/A(low), and later unrated-customer draws back out to SOFR plus 5.50%. [29] [30] [31]

The public bonds have broken. Five-year CDS blew past 1,000–1,200 basis points in mid-October, implying a roughly 60% five-year default probability on standard models, up from about 30% a year earlier. [15] [29] [32] The 9.625% 2032s and nearby unsecured notes traded in the mid-70s to low-80s, with yields of 15–16%. [14] [26] One compilation puts about $4.4 billion due by the end of 2026 and another $6.2 billion in 2027. [33] Later facilities add cash sweeps and contract-value traps; the parent guarantee is the real credit. [30]

The pattern is sector-wide, not idiosyncratic. Of ten listed neoclouds, only two are profitable; combined borrowings of about $100 billion are triple annual revenues, and $57 billion of that debt was issued this year. [2] Nebius equity fell in line with CoreWeave despite lighter reported leverage; it carries roughly $10 billion of convertible principal, $12.1 billion of future lease payments, a $775 million GPU loan, and 2026 capex guidance of $20–25 billion against expected revenue of $3–3.4 billion. [34] Lambda has stacked GPU-backed loans this year, including a $1 billion fixed-rate facility in October tied to Microsoft offtake. [35] A public tracker counts about $197 billion across 110 GPU, lease, and structured financings since 2023, 92% of it in 2025–26, with JPMorgan, Morgan Stanley, Goldman, and MUFG leading the bank side. [36] Gartner’s model says more than 80% of neoclouds could cease independent operations by 2030. [37] [26]

The collateral problem is structural. Loans are secured by GPUs whose resale value falls every product cycle, and by contracts with the same handful of offtakers. If several neoclouds need to sell hardware at once, the buyers are other borrowers of the same lenders. Residual-value insurance does not diversify that risk; it concentrates it. [38] [39]

Circular deals: the supplier is also the lender

The financing loop is no longer a metaphor. Nvidia invested $30 billion in OpenAI’s February 2026 round (Amazon committed $50 billion in the same round, of which $15 billion was upfront and $35 billion conditional) after a September 2025 letter of intent for up to $100 billion that management later walked back. [40] [41] [42] An August 2026 8-K discloses a residual-value guaranty of up to $105 billion backing the lessor of an OpenAI Ohio data center if OpenAI defaults on the lease; OpenAI’s reimbursement obligation to Nvidia is only valuable in the state of the world where the guaranty is not called. [43] [40] Nvidia also holds about 7% of CoreWeave and has agreed to buy at least $6.3 billion of CoreWeave cloud capacity through 2032. [41] [42] Private-company stakes on Nvidia’s balance sheet rose from $3.8 billion to $47.9 billion in a year. [43]

AMD granted OpenAI a warrant for up to about 10% of AMD, vesting against multi-gigawatt purchases. [44] [41] OpenAI’s Oracle cloud commitment has been described at more than $300 billion over roughly five years—the same concentration that sits inside Oracle’s RPO. [45] [13] Broadcom has committed to lend Anthropic up to $42 billion via convertible notes to finance part of a $125 billion five-year TPU lease, and is syndicating a wider package that IFR sizes near $61.5 billion, with senior tranches at SOFR plus 150–187.5 basis points precisely because Broadcom backstops them. [46] [47] Separate talks have covered more than $50 billion for OpenAI custom-chip financing. [48]

SoftBank completed another $30 billion OpenAI investment on October 1 after raising $11 billion of debt at rates near 10%. Its five-year CDS widened to 424 basis points. [19] [9]

The mechanism: vendor financing, equity injections, capacity backstops, and residual guarantees convert future chip purchases into current revenue, while the credit risk lands on the supplier’s balance sheet, on banks syndicating the guaranteed tranche, or on project lenders who believed the offtaker was independent. Nvidia CDS has itself widened to record levels near the high 80s—still investment-grade, but no longer a free option on the loop. [49] [50]

Who holds the risk, and where it is already visible

Already distressed or stressed

• CoreWeave unsecured bonds and CDS. Private-credit anchors—Blackstone, Magnetar, Carlyle, BlackRock—plus bank leads Morgan Stanley, MUFG, JPMorgan, and Goldman are the direct lenders. [29] [32]
• Oracle cash bonds and CDS, one notch above junk, with an explicit S&P downgrade path tied to leverage and free cash flow. [8]
• Stargate project loans (New Mexico, Texas, Wisconsin) trading at discounts, with a force-majeure reservation already sent. Bank syndicates of roughly 20 lenders on the New Mexico deal are the exposed parties; Oracle is the tenant, not the borrower. [17] [19]
• SPV project bonds: Beignet/Hyperion near 91 cents, Excalibur below 87, Hut 8 Nvidia-leased paper in the low 90s. Holders include PIMCO and BlackRock on Hyperion. [26] [20]
• SoftBank CDS at 424 basis points after high-coupon borrowing to fund OpenAI. [9]

Large but not yet broken

• Meta, Amazon, Alphabet, and Microsoft retain AA-tier ratings. Morningstar DBRS sees weaker metrics but limited near-term downside absent a break in the AI earnings cycle. [51] Their CDS have hit records, but from very tight levels and far more slowly than Oracle or CoreWeave. [50] The hidden risk is the lease and guarantee stack, not the rated bonds. Microsoft has been the relative abstainer on long bonds, funding more from commercial paper and cash, but it is the offtaker behind multiple neocloud GPU facilities and the Excalibur lease. [52] [26]
• Private credit as an asset class. Carlyle estimates the AI-compute buildout could require about $1 trillion of private credit—more than half of current industry AUM—and has warned that loans are being underwritten to the hyperscaler’s rating rather than project economics, echoing pre-2008 sponsor-based underwriting. It has urged a roughly 15% concentration cap. [53] [54] The OECD, via the Bank of England, estimated private credit’s share of AI investment financing rose from 9% in 2024 to 34% in 2025. [55] Blue Owl is equity in the prototype SPV and has also led GPU facilities (including part of an IREN package). [36]
• Nvidia, via equity stakes, the $6.3 billion CoreWeave capacity purchase, and the $105 billion OpenAI residual-value guaranty. The loss given a customer default is not the chip margin; it is the guaranty and the stranded capacity. [43]
• OpenAI itself, as the single name inside half of Oracle’s backlog, multiple neocloud borrowing bases, and Nvidia’s guaranty. That concentration is the correlation the sector screen misses.

Covenant and refinancing warnings already on the tape

• S&P’s Oracle triggers: leverage above 4x, no path to positive free operating cash flow by fiscal 2029, or a negative AI-strategy view. [8]
• CoreWeave’s near-term maturity stack and borrowing-base mechanics on GPU facilities. [33] [30]
• September’s issuance collapse and the Amazon order-book withdrawal, which show the IG bid is price-sensitive, not unlimited. [4] [5]
• Project-level delays: at least 47 large US data-center projects delayed, withdrawn, or blocked this year on one regulatory tally, and specific Oracle-linked campuses slipping toward 2029—close to when some of the construction debt matures. [56] [19]

The practical map for a lender or competitor is a three-tier stack. AA hyperscalers can still issue, but every dollar of SPV and uncommenced lease is contingent leverage that rating agencies have not fully folded in. Oracle is the IG name already trading like a stressed credit, with OpenAI as the single point of failure. Neoclouds and GPU loans are where private credit, bank syndicates, and Nvidia’s backstops meet hardware that depreciates. Stress is not a forecast at the bottom of that stack. It is in the prices.

---

Recent Findings Supplement (October 2026)

Recent market pricing and analyst commentary (primarily October 2026) show growing caution toward AI infrastructure debt, with visible stress in neocloud and project-specific bonds, heavy hyperscaler issuance continuing amid negative free cash flow projections, and S&P Global highlighting under-appreciated interconnected liabilities.[1][2]

Hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle, and others) have dramatically scaled bond issuance to fund AI capex, with 2026 volumes already far exceeding prior years and forecasts pointing to further acceleration, though September saw a sharp pullback in new supply.[3][4]

• AI-linked debt issuance reached an estimated $466 billion in 2026 year-to-date (or ranges of $450–570 billion for the full year across bonds, loans, and private placements), more than quadrupling the prior year’s pace in some tallies; hyperscalers alone accounted for a large share, including ~$274 billion in investment-grade bonds through late September (vs. ~$30 billion annual average 2020–2024).[3][4]
• Issuance dipped to $23 billion in September from a June record of $113 billion; Morgan Stanley and others still project a Q4 rebound, with Goldman Sachs forecasting hyperscaler AI-related debt potentially approaching or exceeding $1 trillion cumulatively in 2027.[3][5]
• Specific deals include Amazon’s $54 billion March 2026 offering and Oracle’s $18 billion (September 2025) plus $25 billion (February 2026) multi-tranche sales; the top six hyperscalers are projected to spend >$7 trillion on AI-related capex through 2030 with negative free cash flow persisting.[1][5]
• Implication: Heavy long-dated issuance is competing with sovereign borrowers for capital and pressuring secondary markets, though hyperscalers retain a yield advantage (e.g., ~29–42 bp discount per a CEPR-linked analysis).[6]

Oracle has rapidly increased leverage to support its AI cloud buildout, with total debt reaching ~$167–169 billion (nearly doubled in two years), negative free cash flow of $23.7 billion in FY2026, and FY2027 capex guidance of $90–95 billion; specific project debt is showing stress tied to delays and heavy OpenAI concentration.[7][8]

• S&P downgraded Oracle in July (now barely investment-grade territory); the company has signaled potential repayment delays on its New Mexico “Project Jupiter” data center via force majeure notice amid power permitting issues, with ~$18 billion in associated loans moving into stressed territory.[7][2]
• Michigan AI campus bonds have sold off to ~88 cents on the dollar; Oracle’s backlog includes major OpenAI commitments (analysts like Steve Eisman have flagged concentration risk, with OpenAI reportedly promising ~$300 billion over five years).[2][8]
• Q1 FY2027 capex hit $28.5 billion (vs. $8.5 billion prior year), driving negative FCF despite strong operating cash flow; the firm plans ~$40 billion in debt/equity raises for the year (including a completed $20 billion ATM equity program).[9]
• Implication: Oracle is among the most directly exposed hyperscalers due to execution risks on large projects and customer concentration; rising CDS spreads and bond yields reflect this, even as cloud backlog grows to ~$664 billion.[10]

Meta/Blue Owl-style SPVs and other off-balance-sheet structures have enabled large-scale data center financing, but secondary market prices for these project bonds have declined sharply in recent weeks, signaling investor skepticism.[2][11]

• Beignet Investor LLC (Blue Owl/Meta JV SPV for the Hyperion Louisiana campus) bonds have fallen to ~91 cents on the dollar (from above par in early July); the ~$27 billion private placement carries a 6.58% coupon and benefits from Meta backstops but trades at a premium to Meta’s own bonds.[2][12]
• Other data center project bonds (e.g., Blackstone/QTS “Project Excalibur” leased to Microsoft at <87 cents; Hut 8 projects leased to Nvidia at 91–93 cents) have also sold off.[2]
• Broader off-balance-sheet lease commitments for hyperscalers exceed $1.2 trillion (much not yet commenced); S&P notes “interconnected financing structures” where hyperscalers effectively guarantee builds by smaller entities.[1][13]
• Implication: These structures keep debt off primary balance sheets but do not fool credit markets; Blue Owl’s Stack Infrastructure is testing sentiment with a proposed $6 billion Australian loan for a Melbourne project.[14]

CoreWeave and other neoclouds have relied heavily on private credit and GPU-backed facilities, but their unsecured bonds have entered distressed territory with yields near 16% and CDS spreads at 1,000–1,200 bps.[15][2]

• CoreWeave’s total debt is ~$50 billion ($35.6 billion principal as of mid-2026), built via multiple facilities including a $2.3 billion 2023 GPU-backed loan, $7.5 billion 2024, and an $8.5 billion March 2026 delayed-draw term loan anchored by Blackstone Credit; additional DDTLs and convertibles followed.[15][16]
• Unsecured notes (e.g., 9.625% due 2032 issued at ~9.6–9.7%) have traded below 75–80 cents, with five-year CDS implying ~60% default probability; many facilities are SPV-structured against customer contracts (Microsoft, OpenAI, Meta) with parent guarantees.[2][17]
• Private credit players (Blackstone, Magnetar, Carlyle, BlackRock, etc.) dominate; Carlyle has noted private credit could finance ~$1 trillion of the AI buildout.[15]
• Implication: CoreWeave represents the clearest visible stress test for private credit’s AI exposure; tighter conditions and higher scrutiny are expected for similar borrowers.[18]

GPU-backed lending has expanded rapidly, with Nvidia partnering on platforms targeting >$500 billion in third-party capital, but lenders are demanding stronger guarantees amid uncertainty over chip collateral values.[19][20]

• Nvidia’s August 2026 agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR aim to mobilize third-party financing for customers buying Nvidia systems, with GPUs as collateral and potential Nvidia backstops or residual-value guarantees.[19][21]
• Recent examples include CoreWeave facilities (some rated investment-grade due to customer contracts) and smaller deals via platforms like USD.AI (e.g., $128.9 million facility in October); Lambda has closed multiple GPU loans.[16][22]
• Lenders express caution on long-term chip values and are pushing for more contractual or Nvidia support in pipeline deals.[20]
• Implication: This creates circular exposure (Nvidia financing purchases of its own chips) while shifting risk to private credit and structured vehicles; secondary markets for these loans are emerging but face repricing pressure.[23]

Overall, S&P and other observers flag that hyperscalers likely have more AI-linked liabilities (via guarantees and structures) than disclosed on balance sheets, with Gartner highlighting potential capex-revenue mismatches ($7 trillion spend vs. up to $3.6 trillion revenue through 2030) and risks of large write-downs.[1][3] CoreWeave shows the most acute pricing stress, Oracle faces project-specific and concentration risks, and broader data center/project bonds are repricing lower. Counterparties most exposed include neocloud operators reliant on private credit, Oracle due to its debt ramp and OpenAI ties, and investors in SPV/project debt without strong primary guarantees. No major covenant breaches or refinancings have been reported yet, but elevated yields and CDS levels signal refinancing risk ahead.

The frontier labs are no longer one story. OpenAI is still a capital-markets machine whose own plan burns cash faster than revenue can refill it. Anthropic has flipped to adjusted cash-flow positive while locking itself into mostly non-cancellable compute. xAI’s losses now sit inside a public SpaceX balance sheet and are increasingly paid by other labs renting its GPUs. Mistral is a sovereign-capital story an order of magnitude smaller. What ties them together is that suppliers, not the labs’ equity accounts, now hold much of the duration risk.

OpenAI: a $278 billion burn plan that runs out of cash before it turns

OpenAI’s July 2026 internal plan, reported by the Financial Times and confirmed by Reuters, projects negative free cash flow of $278 billion from 2026 through 2030, against $856 billion of compute and infrastructure spend and cumulative revenue of about $840 billion (from $36 billion in 2026 to $350 billion in 2030). That is a slight improvement on a May projection of roughly $305 billion of burn, but it still exhausts the cash already raised by about 2028.

The accounting of “how much revenue” is itself a market event. OpenAI told investors annualized revenue was approaching $50 billion at the end of September 2026 on its own net basis (it books only its share of partner-routed sales). A widely circulated ~$70 billion figure was an investor gross-up meant to match Anthropic’s method. The company still told fundraising counterparties it expects $70 billion or more by year-end 2026, with enterprise the main driver (Q3 annualized growth about 77 percent, enterprise about 107 percent).

• 2025 revenue was about $13.1 billion; year-end 2025 cash was about $40 billion. Q1 2026 revenue was $5.7 billion with operating costs of $3.7 billion; cash and securities ended the quarter at $73 billion, mostly the March raise, not operations.
• March 2026: up to $122 billion at an $852 billion post-money valuation. On the July plan, that stack plus year-end 2025 cash (about $162 billion in hand, in one reconstruction) is gone by 2028, leaving roughly $116 billion still to raise before the plan’s own cash-flow turn.
• A February 2026 plan (The Information) still showed cash-flow positive in 2030 (+$39 billion) after annual burns of −$25 / −$57 / −$85 / −$51 billion. Some sell-side notes read the July update as pushing that turn further out and requiring gross margins north of 70 percent to reconcile the capex with the stated burn. Reported 2025 gross margin was about 33 percent, with inference the constraint.
• Commitments that show up in counterparties’ filings include an incremental $250 billion of Azure, Amazon’s $38 billion plus a $100 billion expansion, and CoreWeave contracts in the high teens to low twenties of billions. Oracle’s roughly $300 billion / multi-year figure is a press report, not an Oracle filing that names OpenAI. Stargate remains a $500 billion investment intention, not a single signed offtake.

What this means for a competitor or a supplier: OpenAI’s near-term constraint is not demand narrative, it is the gap between signed capacity and cash on hand. A slower enterprise ramp or stickier inference costs does not just delay profit; on the company’s own arithmetic it forces another raise before 2028 or a cut in the capacity plan. [1] [2] [3] [4] [5] [6]

Anthropic: operationally ahead, contractually trapped

Anthropic’s draft IPO prospectus, reviewed by Reuters, is the first audited-style look inside a frontier lab. 2025 revenue was $4.59 billion (up from $386 million), operating loss $8.06 billion, compute and infrastructure $7.33 billion. The headline ~$42 billion net loss is mostly a ~$34 billion non-cash remeasurement of convertibles as the valuation rose. Cash and short-term investments were $20.28 billion at 31 December 2025. Two customers were nearly a quarter of 2025 revenue; sales through cloud marketplaces were about 47 percent.

2026 flipped the unit economics. Quarterly revenue ran $4.73 billion (Q1), about $11.5–11.6 billion (Q2), and $16.8 billion (Q3, up about 45 percent). The company told investors it was profitable on an adjusted basis for a second straight quarter, with adjusted operating margin around 11 percent and $1.70 of revenue per $1 of compute, versus roughly 63 cents in 2025. Simple annualization of Q3 is about $67 billion; the late-July run rate cited in the prospectus process was above $65 billion. Bankers have been pricing an IPO case on $190–200 billion of 2028 revenue.

The binding constraint is the other side of the ledger: at least $518 billion of cloud, compute, and infrastructure commitments over roughly a decade, about 80 percent non-cancellable or payable regardless of use.

• Broadcom-related equipment leases about $161.2 billion; Google at least $111.1 billion; Amazon $110 billion; Microsoft $31.4 billion; AMD capacity expected to exceed $20 billion (AMD also agreed to buy up to $5 billion of stock). The filing language on Google is explicit: if actual spend falls short, Anthropic must pay the difference.
• xAI/SpaceX capacity of up to about $84.5 billion through 2029 is the flexible piece—mostly cancellable on 90 days’ notice. A separate Colossus lease has been described at about $1.25 billion a month through May 2029.
• Equity already raised is large (Series G $30 billion in February 2026; Series H $65 billion in May at a $965 billion valuation; one tracker puts post-2025 equity near $105 billion this year alone). An IPO of up to about $100 billion, targeted for mid-November after the U.S. midterms, is still in the plan even though adjusted free cash flow is positive. Valuation talk has ranged from $1.8–2 trillion down to investor discussion around $1.5 trillion after the October investor day. New Constructs’ reverse-DCF put a low-case equity value near $144 billion.

What this means: Anthropic no longer needs outside money to “keep the lights on” on its own adjusted figures. It needs continued capital—and continued revenue growth—because the take-or-pay book is fixed while customers, the prospectus warns, often are not. A growth miss shows up as a cash shortfall against Google and Amazon minimums, not as a quiet cut in capex. [7] [3] [8] [9] [10] [11] [12]

xAI inside SpaceX: burn is now a public-company segment, partly funded by rivals

SpaceX’s IPO filings are the first hard look at xAI. The AI segment (xAI, Grok, and X, recast under common control) lost $6.4 billion from operations in 2025 on $3.2 billion of revenue, versus a $1.56 billion loss on $2.62 billion in 2024. AI capex was $12.7 billion in 2025 and $7.7 billion in Q1 2026 alone (about a $31 billion annualized run rate). Q2 2026 AI revenue was $2.56 billion, operating loss $1.26 billion, but segment adjusted EBITDA turned positive at $1.15 billion—while capex hit $15.8 billion, about 6 times that quarter’s revenue.

The mechanism that changed the story is not Grok subscriptions. Product revenue is still small relative to the build (2025 “AI solutions and infrastructure” included only $365 million of X and Grok subscriptions). The segment is increasingly a landlord: Anthropic agreed to pay on the order of $1.25 billion a month for Colossus capacity through May 2029, with additional reported leases (Google and others) in the high hundreds of millions per month. Colossus and Colossus II were described as about 1 gigawatt combined at filing; later commentary put capacity nearer 1.4 gigawatt. Most of the Anthropic offtake can be cancelled on short notice, so this revenue is real and also reversible.

xAI raised a $20 billion round in January 2026 and was valued around $250 billion at the February combination with SpaceX. The parent’s cash pile (reports around $100 billion of cash and marketable securities after the IPO) is what makes a $1 billion-a-month style infrastructure burn financeable without a standalone xAI raise. Use of IPO proceeds explicitly includes more AI compute, including a path toward “multiple trillions of parameters.”

What this means: xAI’s dependence on continued private raises is lower than OpenAI’s because SpaceX equity and cash now intermediate it—but the equity story only works if GPU rental demand from Anthropic, Google, and others stays tight. If those labs slow training or exercise cancellation rights, SpaceX is left with a capex line that already dominates group spending. [13] [14] [15]

Mistral: sovereign capital at a $24 billion mark, still pre-scale

Mistral is not in the same cash-burn regime, and it does not publish burn, margins, or audited revenue. What is public is a run-rate and a capital stack built for European sovereignty rather than a global training race.

CEO Arthur Mensch told the Financial Times the annualized run rate was above $400 million by February 2026, up from about $20 million a year earlier. CFO Johan Bergqvist told Reuters on 8 September 2026 the company was on track to pass $1 billion of ARR by year-end. As of early October, no report confirmed that mark had been crossed. About 60 percent of revenue has been described as European.

• 8 September 2026: €3 billion Series D (about $3.5 billion) at a post-money valuation above €21 billion (about $24 billion), led by Samsung, the EU-backed Scaleup Europe Fund (EQT-managed), and PSG Equity. That followed a September 2025 Series C of €1.7 billion at €11.7 billion, led by ASML.
• March 2026: $830 million of bank debt to buy 13,800 Nvidia chips for a site near Paris. Separately, about €1.2 billion has been discussed for a Swedish site, with targets of 200 megawatts by 2027 and 1 gigawatt by 2030.
• No public loss or cash-burn figure. Outside commentary treats the company as not yet profitable. At the pre-money implied by the September round, Reuters Breakingviews put the ARR multiple near 21 times, in the same neighborhood as the Americans’ last private marks—on a base two orders of magnitude smaller.

What this means for entrants: Mistral’s financing path is strategic capital (EU vehicle, ASML, Samsung, Nvidia) plus modest vendor-backed debt, not a $100 billion compute book. The risk is utilization of owned GPUs and open-weight price pressure, not a 2028 cash-out date. A competitor that cannot access sovereign or industrial anchors will not replicate this round on revenue alone. [16] [17] [18] [19]

Who actually funds the build: sovereign money, vendor paper, and other people’s debt

The labs’ equity rounds are only the visible layer. A large share of the infrastructure is financed by suppliers, project lenders, and sovereign vehicles that are paid only if the labs keep growing.

Sovereign and strategic equity. Abu Dhabi’s MGX, chaired by Sheikh Tahnoon bin Zayed, has participated in OpenAI, Anthropic (including the $65 billion Series H), and xAI, and has secured on the order of $50 billion in commitments for AI investments. OpenAI’s current talks for at least $30 billion at about a $1.4 trillion pre-money valuation include MGX and other UAE funds (a syndicate discussed at up to $10 billion) plus BlackRock. SoftBank completed a further $30 billion into OpenAI around 1 October 2026—after selling about $11 billion of high-yield bonds near 10 percent—and has been reported in talks to raise up to $100 billion from Gulf investors for a broader AI vehicle. Stargate UAE (a planned gigawatt-scale cluster, initial 200 megawatts) ties the same capital to in-country inference, not just equity checks.

Vendor financing. Broadcom has offered Anthropic up to $42 billion of convertible notes toward a large TPU lease book. AMD’s stock-purchase commitment is equity-for-capacity. Nvidia is simultaneously supplier, CoreWeave shareholder ($2 billion of stock in January 2026), and a backstop buyer of unsold CoreWeave capacity (an order form with an initial value of $6.3 billion through 2032). Amazon and Microsoft are investors, distributors, and the counterparties on take-or-pay cloud contracts. OpenAI’s own credit lines are small relative to the plan (about $4.7 billion, described as undrawn). Anthropic’s revolver has been described as moving from $2.5 billion toward $15 billion.

Partner and project debt is where the slowdown hits first. By late 2025, nearly $100 billion of borrowing had been lined up against OpenAI-linked sites (Abilene, New Mexico, Vantage and others). SoftBank-controlled SB Energy has been reported with $174 billion of capital needs against $430 billion of contracted future revenue, largely OpenAI, and has delayed a listing. Oracle carries a $664 billion remaining-performance-obligation book (August 2026) and about $125 billion of borrowings, with some project debt (including an ~$18 billion New Mexico facility) trading at stressed levels after permitting delays. CoreWeave’s bonds sold off into distressed territory in mid-October; the New York Times reported $72 billion of liabilities across 51 data centers and investor pricing that implied a better than 60 percent chance of missing obligations. Anthropic can walk away from a reported $45 billion Nscale deal if financing or delivery timelines slip.

Hyperscalers with multi-tenant clouds (Microsoft, Amazon, Google) can re-rent a freed rack. Single-tenant neoclouds and project SPVs cannot. The Information’s industry contacts expect hundreds of neoclouds to consolidate to roughly 7–10. Gartner has said data-center owners would need cumulative revenue of $8.5 trillion through 2030 to cover capex, financing, and a normal return, against attributable AI revenue it forecasts at no more than $3.6 trillion—leaving more than $3.5 trillion of spending at write-down risk in Bloomberg’s summary of the same work. Bain’s September technology report says the industry needs about $6 trillion of annual AI revenue by 2031 to justify a build whose annual infrastructure spend could reach $1.5 trillion, with existing consumer and enterprise AI products covering at most $1.2–1.8 trillion.

Dependence, ranked. OpenAI is the most raise-dependent of the four on its own plan (cash out in 2028, IPO pushed to 2027 after a confidential filing, Altman citing safety). Anthropic is operationally less dependent and contractually more dependent; the November IPO is as much a commitment-funding event as a liquidity event. xAI’s burn is now a SpaceX problem, buffered by public-market cash and rental income that Anthropic can cancel. Mistral is dependent on the next sovereign or strategic round, not on a $100 billion take-or-pay book. Google, Microsoft, Amazon, and Meta remain the contrast case: they fund training and inference from operating cash flow and can slow orders without a down round. [20] [21] [22] [23] [24] [24] [25] [26]

Break-even timing and what actually moves it

Reported timelines, not a single consensus.

Lab	What has been said	Sensitivity
OpenAI	Feb 2026 internal plan: cash-flow positive in 2030. July plan: $278 billion negative FCF over 2026–30, cash from the March round gone by 2028. Some analysts read positive FCF as pushed past 2030. Older Reuters reporting had cash-flow positive only in 2029.	22V estimates gross margin must clear ~70 percent to reconcile the capex path with the stated burn. 2025 gross margin was ~33 percent.
Anthropic	Adjusted operating profit and adjusted free cash flow positive by Q3 2026 (company to investors). IPO case assumes $190–200 billion revenue in 2028. 2025 operating margin was deeply negative (−175 percent on New Constructs’ reading of the draft).	Take-or-pay means “profitable” and “able to fund commitments” are different tests. A growth fade against Google/Amazon minimums recreates a cash need even if adjusted margins stay positive.
xAI / SpaceX AI	No stated break-even year. Q2 2026 adjusted EBITDA positive; GAAP operating loss and capex still several times revenue.	Rental mix vs. owned training. Cancellation of the Anthropic/Google leases would drop revenue faster than capex.
Mistral	No public target. Still described externally as loss-making at a few hundred million of ARR.	Owned-GPU utilization. Idle sovereign capacity is a cash drain at a scale equity can still cover; it is not yet a systemic supplier problem.

One independent model (06.ai, 19 October 2026) has a base case of about $72 billion OpenAI and $100 billion Anthropic annualized revenue by December 2026, rising to about $230 billion and $265 billion by December 2027, with wide ranges. That is a third-party path, not company guidance. Anthropic has beaten its own longer-range projections since the start of 2025; OpenAI’s September print came in $20 billion below the gross figure investors had been using.

The three sensitivities that matter, in order.

1. Inference cost per dollar of revenue, not headline user growth. Anthropic’s compute ratio moved from spending $1.60 to earn $1 in 2025 to earning $1.70 per compute dollar in Q3 2026. That is the entire “already profitable” claim. OpenAI’s 2025 inference bill was large enough to hold gross margin near 33 percent even as revenue tripled. Price cuts (Anthropic’s Haiku 5.5 at about 75 percent below its predecessor; ongoing frontier token deflation) help adoption and hurt the ratio unless hardware efficiency or mix shifts faster.

2. Subscription growth is a weak profit lever and can be a compute sink. SemiAnalysis estimates Anthropic subscriptions are about 10 percent of revenue but can absorb over 40 percent of inference compute, lowering blended revenue per megawatt by about $36 million. A maxed-out premium plan on the expensive model can be deeply negative gross margin; at realistic utilization the same plan is fine because API list prices embed very high token margins. OpenAI’s response has been to cut token allowances on the $200 plan rather than subsidize heavier use. More ChatGPT Plus users, without a shift to enterprise or ads, does not close a $278 billion burn plan.

3. Enterprise and consumption revenue is the lever that can. OpenAI says enterprise has overtaken consumer and is the reason it still guides to $70 billion-plus by year-end despite the $50 billion September net print. Anthropic’s 2025 revenue was already mostly usage-based (~$3.8 billion consumption vs. $789 million subscription), which scales with customer workloads rather than seat counts—but also concentrates risk (two customers, nearly a quarter of sales; many large customers with no long-term lock-in). Ads are still small (OpenAI’s ads business was reported around $1 billion annualized by August 2026). Bain’s point is the structural one: even if both labs hit their 2027–28 paths, existing consumer and enterprise AI products are not, on Bain’s and Gartner’s math, large enough to underwrite the industry-wide data-center book. The residual has to come from new categories (agents that replace labor, robotics, industry-specific automation) or the capacity does not earn its cost of capital.

If funding slows, the sequence is visible now, not hypothetical. First, project debt and neocloud equity reprice—that already happened in the week of 13 October, led by CoreWeave and, more mildly, Oracle and SoftBank credit. Second, flexible offtake is cut (Anthropic’s 90-day xAI option; Nscale walk-away rights; chip orders that are not yet leases). Third, labs that still need equity (OpenAI before 2028 on its plan; anyone whose adjusted profit does not cover minimums) either raise at a lower mark, delay capacity, or both. Microsoft, Amazon, and Google absorb a slowdown as a utilization problem inside a broader cloud P&L. CoreWeave, SB Energy, and single-tenant Oracle campuses absorb it as a refinancing problem. Nvidia absorbs it last, as orders, unless it has already financed the buyer. [3] [27] [28] [29] [30] [31] [32]

Confidence and gaps. The OpenAI burn and revenue-plan figures are internal presentations reported by the FT, not audited statements. Anthropic’s 2025 figures and the $518 billion book come from a confidential draft prospectus seen by Reuters; Q3 2026 profitability is company-reported and adjusted (stock-based pay and, in some accounts, training costs sit outside the headline margin). xAI figures are SpaceX SEC disclosures and are the hardest of the set. Mistral has no public burn. Oracle’s $300 billion OpenAI figure remains a press report. Treat run-rate “annualized revenue” as a growth signal, not GAAP, and do not compare OpenAI’s net figure to Anthropic’s gross figure without the partner-share adjustment.

---

Recent Findings Supplement (October 2026)

OpenAI’s July 2026 internal plan projects $278 billion in negative free cash flow from 2026–2030 (down from a May forecast of $305 billion), driven by $856 billion in planned compute and infrastructure spending while revenue scales from $36 billion in 2026 to $350 billion in 2030 (cumulative $840 billion).[1][2]

• As of end-September 2026, the company reported an annualized revenue run rate approaching $50 billion (net of partner shares), with expectations of $70 billion+ by year-end; earlier $70 billion estimates reflected investor application of gross accounting methods similar to Anthropic’s.[3][4]
• The March 2026 $122 billion raise (at $852 billion post-money valuation) is projected to be exhausted by 2028; the company is accelerating talks for a new ~$30 billion round at a target valuation above $1.2–1.4 trillion ahead of Anthropic’s planned IPO.[5][6]
• Confidential IPO filing occurred in June 2026, but CEO Sam Altman stated in September that no public listing will occur in 2026 due to AI safety concerns.[7]

This underscores heavy dependence on continued large-scale raises to bridge the multi-year cash gap, with revenue growth (including enterprise and ads) needing to outpace inference and training costs for break-even scenarios pushed toward or beyond 2030.

Anthropic reported Q3 2026 revenue of $16.8 billion (up ~46% sequentially from $11.5 billion in Q2), achieving adjusted profitability and positive free cash flow for the second straight quarter, with $1.70 in revenue per dollar of compute spend.[2][8]

• Its confidential IPO prospectus (reviewed by Reuters in late September) details at least $518 billion in compute/cloud/infrastructure commitments over roughly a decade across six partners (~80% non-cancellable or payable regardless of usage), including Broadcom (~$161 billion in leases), Google (≥$111 billion), Amazon ($110 billion), Microsoft ($31.4 billion), plus a more flexible up-to-$84.5 billion xAI-linked deal.[9][10]
• 2025 full-year figures showed $4.59 billion revenue (12x prior year) against an $8.06 billion operating loss and $7.33 billion in compute spend; the company has raised over $100 billion in equity since 2025 and plans a November 2026 IPO (potentially targeting or discussed around $1.5–2 trillion valuation, with some investor tempering noted after sequential growth slowdown).[11][12]
• Positive cash flow reduces near-term equity dependence for operations, but locked-in commitments create ongoing financing needs and exposure to revenue sensitivity or cost pressures.

xAI was acquired by SpaceX in an all-stock deal effective February 2026 (implying ~$250 billion standalone / ~$1.25 trillion combined valuation), following its January 2026 upsized $20 billion Series E at a $230 billion standalone valuation.[13]

• SpaceX filings disclose an AI segment with $3.2 billion revenue and $6.4 billion operating loss in 2025 (standalone core Grok/API estimates around $500 million ARR); capex is ramping sharply.[14]
• The merger integrates xAI’s Colossus infrastructure with SpaceX resources, shifting funding dynamics toward the parent’s balance sheet and upcoming IPO plans.

Mistral closed a €3 billion (~$3.5 billion) Series D in September 2026 at a >€21 billion (~$24 billion) post-money valuation (led by Samsung, with PSG Equity and Scaleup Europe Fund), its largest round to date.[15]

• ARR exceeded $400 million by early 2026 (targeting >$1 billion by year-end, described as on track); revenue mix includes API usage, sovereign/enterprise deployments, and consumer tiers.[16]
• Additional March 2026 debt financing (~$830 million) supported data-center expansion.

Hyperscalers and compute vendors (Microsoft, Google, Amazon, Oracle, Broadcom, Nvidia) hold hundreds of billions in backlog from these labs, with debt issuance and capex guidance rising sharply to fund shared infrastructure (e.g., OpenAI-linked Stargate elements, Anthropic multi-GW deals).[2][17]

• Non-cancellable or take-or-pay elements in Anthropic’s contracts (and similar OpenAI commitments) transfer some risk to suppliers via minimum spend obligations or financing arrangements (e.g., Broadcom convertibles or partner-raised debt), but slowdowns in lab funding or utilization could pressure vendor revenues, utilization rates, and their own debt service.
• Sovereign and strategic vendor participation (e.g., MGX, Amazon, Nvidia investments) provides layered financing but creates complex incentive alignments.

Break-even and sensitivity scenarios remain tied to execution on revenue growth versus inference/compute costs, with limited new public analyst modeling post-April 2026. OpenAI’s updated plan shows deeper near-term burns offset by aggressive 2030 targets; Anthropic’s adjusted profitability highlights efficiency gains (revenue per compute dollar) but underscores that fixed obligations amplify downside risks from subscription/enterprise adoption shortfalls or price competition. Suppliers face asymmetric exposure if capital markets tighten or AI demand growth moderates.[18]

These developments (primarily September–October 2026 reporting) highlight accelerating capital intensity, accounting nuances in reported run rates, and the shift toward public-market scrutiny for the leaders, with xAI’s integration and Mistral’s European-scale raise adding diversity to the funding landscape.

The market is simultaneously scarce at the frontier and underused inside ordinary enterprises. What looks like a shortage or a glut depends on which utilization definition, which silicon generation, and which buyer you measure. Public price and survey series can be checked; fleet-level GPU utilization and the training-versus-inference split mostly cannot.

Utilization is not one number, and the “100%” and “5%” headlines describe different machines

Bank of America argued on October 16, 2026 that deployed AI infrastructure is running near 100% utilization, and that rental rates on even older A100, H100, and B200 generations are up 12–37% year to date—evidence of shortage, not the idle fiber of 2000. That claim is about capacity that has already been sold into training clusters and long-term cloud contracts, where a kernel is almost always running and the customer is paying whether or not the job is efficient.

Inside enterprises that own or provision their own GPUs, the picture is the opposite. A VentureBeat Research survey of 573 technical leaders found 86% of operators reporting utilization of 50% or less. Cast.ai’s 2026 Kubernetes report, covering 23,000 clusters through April 2026, put average GPU utilization at about 5% (Azure 2%, AWS 5%, Google Cloud 6%). Those are not contradictory once the metric is unpacked. NVIDIA’s own NVML “GPU utilization” only records whether at least one kernel ran during a sample interval; it is not a share of peak FLOPS. A small job can pin the gauge at 90%+ while using a fraction of the silicon. Older academic traces make the same split: Microsoft’s Philly study found hardware utilization of about 52% on GPUs that were in use; Alibaba’s PAI traces showed high allocation with much lower actual activity; dedicated LLM training clusters have reported median utilization of 97–99%.

Power waste sits on top of that. NVIDIA’s Rod Evans said in mid-October 2026 that only about 60–70% of electricity drawn from the grid ends up doing AI work: roughly 20% lost to cooling, conversion, and distribution, and about 10% to misconfigured racks, failures, and restarts. A contracted 100 MW site may deliver closer to 93 MW, and less at peak hours.

What this means for anyone entering the market: quoting “utilization” without saying whether you mean reserved, billed, kernel-busy, or useful throughput is not analysis. The scarce asset is liquid, interconnect-dense, power-backed cluster time. The surplus is poorly scheduled enterprise GPUs and Kubernetes overprovisioning. A neocloud that only reports occupancy, and an enterprise that only reports nvidia-smi, are talking past each other.

• BofA: near-100% on deployed AI infrastructure; A100/H100/B200 rents up 12–37% YTD. [1] [2]
• Enterprise self-report: 86% at ≤50% utilization (n=573). [3]
• Kubernetes fleets: ~5% average GPU utilization across 23,000 clusters. [4]
• NVML utilization is kernel presence, not peak performance or profit. [5] [6]
• Grid-to-useful-compute: ~60–70%. [7]

Price signals split: GPU hours have firmed in 2026; tokens have collapsed

Independent rate panels as of October 18–19, 2026 show a wide but coherent ladder. Compute Credit Index Research’s guaranteed on-demand panel (U.S. and Europe) printed B200 at $14.24 per GPU-hour on hyperscalers, $8.11 on neoclouds, and $8.36 on marketplaces; H200 at $10.30 / $4.50 / $5.31; H100 at $11.25 / $3.34 / $3.07; A100 at $4.45 / $1.82 / $1.58. Interruptible grades are far cheaper—H100 interruptible at $3.30 hyperscaler and $2.15 neocloud—because the operator can reclaim the box. Other public medians land in the same neighborhood: cloud-gpus.com on October 19 had B200 on-demand at $8.34, H200 SXM at $5.53, and H100 SXM at $3.99, with floors near $2 on marketplaces and hyperscalers often 2–3× the specialist floor.

Direction matters more than the level. getdeploying’s median on-demand series was up 21% over the year to the week of October 19, 2026, and up 9% over the prior four weeks. That is the opposite of a glut tape. Earlier in the cycle, H100 lead times had collapsed from about 52 weeks to under 8 by March 2026, and spot H100 prices had fallen sharply from the 2023–24 peak. The 2026 rebound is selective: newest rack-scale Blackwell remains allocated, while older generations are no longer collapsing in price because inference and batch work still pay for them.

Token prices moved the other way. BenchLM’s frontier token-price index stood at 15.5 on October 19, 2026, 84.5% below its March 2023 base. OpenAI cut GPT-6 Sol to $2 / $10 per million input/output tokens and Luna to $0.10 / $0.50, roughly half prior promotional rates. Anthropic held an Opus-class line for several releases and then cut it; open-weight models run a large share of volume at a steep discount to closed frontier SKUs. Ramp’s economist Ara Kharazian reported the effective blended price down 41% to $0.68 per million tokens from a 2026 peak of $1.15 in March, with cheaper standard models driving volume and frontier models’ token share falling from 53% to 45%.

The trap is equating cheaper tokens with cheaper work. One 2026 cost series found list prices down about 91% from Q1 2023 to Q3 2026, but real cost per completed task down only about 63%, and up roughly 80% from a Q4 2024 trough, because reasoning models burn more tokens per task. Falling API prices can raise, not lower, the GPU bill if agent loops multiply calls.

What this means: a buyer who shops only hyperscaler list rates overpays; a seller who underwrites residual value off 2024 H100 peaks will be wrong in both directions. The public, checkable series are list and marketplace rates. Contracted one-year deals and take-or-pay prices are only partly visible.

• CCIR guaranteed ladder, October 18, 2026. [8] [9]
• Median on-demand snapshot, October 19, 2026. [10]
• On-demand medians up 21% year over year. [11]
• Frontier token index −84.5% from March 2023. [12]
• Effective price $0.68/M tokens; frontier share 45%. [13]
• List prices down far more than cost per completed task. [14]

Power, not wafers, is the binding constraint on new capacity

Morgan Stanley’s October 2026 work puts the U.S. data-center power deficit at about 32 GW through 2028, roughly 34% of projected requirements, even after on-site generation and fuel cells. The bank says NVIDIA has confirmed the thesis: more semiconductors are being built than there are places to plug them in. Large, creditworthy players are expected to take about 94% of global data-center capacity additions in 2026, falling to 64% in 2027 and 57% in 2028—a squeeze on smaller developers, not a chip shortage in the 2023 sense. A related Morgan Stanley note cut the expected U.S. share of global compute from 60% to 55% because overseas power cannot absorb that 32 GW gap, and flagged a possible widening toward 68 GW by 2029 in one framing.

Hyperscalers are still describing a demand-constrained tape, not an inventory build. Combined 2026 capex guidance for the four largest was cited in mid-October at $720–745 billion, roughly double the 2025 pace. Amazon has said a roughly $220 billion plan will not cover 2026 demand and that it expects to stay constrained into 2027, monetizing capacity as fast as it is installed. Microsoft has said Azure demand continues to surpass supply (Azure growth was cited at 43% in the latest quarter in that note). Alphabet has raised its capex range to deliver capacity faster. TSMC has said packaging is tight enough to limit customer growth. Those are management statements, not audited utilization, but they are the opposite of the fiber-era pattern in which suppliers cut prices into unsold inventory.

The forward revenue math is the bear case, and it is a forecast, not a measurement. Gartner’s October 14, 2026 release said AI providers are expected to spend nearly $1 trillion on infrastructure in 2026 and another $6 trillion from 2027 to 2030. To cover that spend, financing, and a 12% return, data-center owners would need about $8.5 trillion of cumulative revenue through 2030. Gartner forecasts no more than $3.6 trillion of enterprise and consumer AI revenue attributable to those data centers, and says more than $3.5 trillion of spending could be exposed to impairment by the end of 2030. It also expects more than 80% of neoclouds to cease independent operations by 2030 through acquisition, partnership, or failure—not necessarily bankruptcy. Hyperscalers can absorb a miss because AI is not their only P&L. A single-purpose GPU lessor cannot.

What this means: the scarce input for a new entrant is an interconnection and a credit rating, not a purchase order for H100s. Projects that cannot energize on the hyperscaler timetable will slip even if the chips arrive. The Gartner gap is a returns problem, not proof that today’s racks are empty.

• 32 GW U.S. shortfall through 2028, ~34% of need. [15] [15] [16]
• Hyperscaler guidance and “demand exceeds supply” statements. [17] [17]
• Gartner: $8.5 trillion required vs ≤$3.6 trillion attributable revenue; >$3.5 trillion impairment exposure; >80% of neoclouds exit independence by 2030. [18] [19]

Enterprise adoption is real, narrow, and still mostly a productivity story rather than an EBIT story

The “MIT 95% of pilots fail” line is a misread of Project NANDA’s GenAI Divide (fieldwork January–June 2025, published mid-2025). The funnel in the report was: about 60% of organizations investigated custom or task-specific tools, 20% reached a pilot, and 5% of all organizations had a successful production deployment with measurable impact. Among the minority that actually piloted, roughly one in four cleared a high bar—marked P&L or productivity effect within six months. The executive summary’s “95% of organizations getting zero return” is a different claim from “95% of pilots fail.” The more useful finding was operational: partner-built tools reached deployment far more often than internal builds (on the order of two-to-one in the coverage of the report), and generic chat tools were widely used by individuals while workflow-embedded tools stalled.

McKinsey’s 2026 State of AI survey (1,719 respondents, 97 countries, May 4–June 8, 2026) shows the same split a year later. About 80% of respondents say AI improved their own productivity. Only 37% attribute any EBIT impact to AI, flat versus the prior year. About 6% qualify as high performers (at least 5% of EBIT from AI, and described as significant). Scaling is uneven: 44% say AI is scaling across the enterprise, up from 38%; 40% of companies with $1 billion-plus revenue say they are scaling agents, up from 27%, while smaller firms are flat. About one in five organizations already cites AI operating costs as a constraint. McKinsey’s separate cost work says moving a tool from pilot to production nearly quadruples spend, because unguarded model access, cloud compute, and integration dominate the license fee. These are self-reported perceptions, not audited earnings.

The Census Bureau’s Business Trends and Outlook Survey is the cleanest public adoption series, and the question wording is the story. Through late 2025 the survey asked whether the firm used AI in producing goods or services, and the rate sat near 10%. In November 2025 the question became “any business function,” and measured use nearly doubled. The 2026 AI-supplement working paper (reference period November 2025–January 2026) found 18% of firms using AI in at least one function, 32% on an employment-weighted basis, with 22% expected within six months. Very large firms in information, professional services, and finance ran 50–60% (60–70% employment-weighted). Among users, 57% used AI in three or fewer functions, most often sales and marketing. Worker task use showed up in 23% of firms (41% employment-weighted), mostly writing, document analysis, and search. AI-related employment decreases were reported in only 2% of firms. A later biweekly wave (September 21–October 4, 2026) put the share at 23.8%.

Spend data sees a different population. Ramp, tracking card and vendor payments at tens of thousands of U.S. businesses, has published paid-AI adoption well above Census—56.1% as of August 2026 in one write-up, versus Census near 22%—because a $16–$20 monthly ChatGPT seat counts as adoption in transaction data and may not count as “using AI in a business function” on a Census form. Ramp’s October 2026 index also showed the tension in the price war: token volumes rose even as spend at the heaviest users softened from a July high (top 1% at $9,568 per employee per month in September, down 3.1% from $9,873), before early-October spend ticked to a new peak. Median firm spend is tiny; the bill is concentrated.

What this means: competing on “enterprises are not adopting” is wrong. Competing on “enterprises have proven scaled ROI” is also wrong. The wedge is narrow workflows with a P&L owner, usually bought rather than built, in large knowledge-sector firms. Broad agent rollouts that quadruple the bill before the process is redesigned are how the 37% EBIT figure stays flat.

• MIT NANDA funnel and the 95% misquote. [20] [21] [22]
• McKinsey 2026: 80% individual productivity, 37% any EBIT, ~6% high performers. [23] [24] [25]
• Pilot-to-production spend nearly quadruples. [26]
• Census: 18% / 32% employment-weighted (Nov 2025–Jan 2026); question change doubled the series; 23.8% in early October 2026. [27] [28] [29]
• Ramp vs Census and the volume-versus-spend split. [30] [31] [32]

Inference has likely passed training on spend, not yet as a public census of watts

Gartner’s AI-optimized infrastructure-as-a-service forecast for 2026 is about $42.3 billion, nearly double 2025, with inference at $23.3 billion and training at $19 billion—the first year inference is the larger line, at 55%, and a sign that buyers are serving models in production rather than only training them. Deloitte’s compute-share estimates put inference near one-third of AI compute in 2023, about half in 2025, and roughly two-thirds in 2026. Futurum’s platforms sizing had managed inference already ahead in calendar 2025 ($23.1 billion vs $16.3 billion for training) and at 71.7% of that infrastructure slice by 2030 in its base case.

Power models are more cautious and should not be averaged with dollar models. A McKinsey data-center estimate cited in 2026 coverage had 2025 training demand at 23.1 GW and inference at 20.9 GW—training still slightly larger on watts—then inference growing faster (about 35% a year vs 22%) to 93.3 GW versus 62.2 GW by 2030. Lifetime compute for a deployed model is even more inference-heavy, because the training run happens once and serving runs for years. None of these splits is a regulatory filing. They are vendor and analyst constructs with different denominators (IaaS dollars, managed-platform dollars, IT power, GPU-hours).

The mechanism that links this to prices: inference is bursty, latency-sensitive, and willing to run on last-generation memory-rich GPUs, which is why H100 and H200 rents can rise even as frontier training clusters hoard Blackwell. Training still sets the peak rack density—GB200-class halls around 130 kW per rack, with roadmaps far higher—which is exactly what makes the power queue the constraint.

What this means: underwriting a data center as if 2023’s training mix still dominates will mis-size both the interconnect and the customer. Inference customers care about cost per million tokens and time-to-first-token, not cluster scale. That is a different sales motion, and a different residual-value story for older GPUs.

• Gartner: $23.3 billion inference vs $19 billion training in 2026 IaaS. [33] [34]
• Deloitte compute share and McKinsey GW paths. [35]
• Futurum inversion already in CY2025. [36]

Circular financing is large enough to contaminate “demand,” but it is not the whole tape

The circular loop is now in filings, not just diagrams. NVIDIA has taken equity in customers, committed to buy cloud services from them, and extended guarantees tied to sites that must run NVIDIA hardware. Coverage of its disclosures cites on the order of $36 billion in cloud-service commitments and guarantees capped around $105 billion for an OpenAI-leased campus, plus a $2 billion CoreWeave stake and a residual-capacity purchase arrangement reported at $6.3 billion through 2032. Microsoft’s relationship with OpenAI includes a very large Azure purchase commitment (reported around $250 billion incremental) that shows up as Microsoft cloud revenue. Amazon, Google, and AMD have analogous stake-plus-compute structures with Anthropic and OpenAI. Proponents call this a flywheel that accelerates scarce supply. Critics note that supplier, investor, landlord, and customer are the same set of firms, so one dollar can be counted as capex, revenue, and backlog more than once.

CoreWeave is the stress case that shows both sides. At the end of Q2 2026 it reported a $104.2 billion revenue backlog with names such as Microsoft, OpenAI, Meta, and Anthropic, and management said that backlog more than supports its debt. It also carries on the order of $50 billion of debt, guided 2026 capex at $35–39 billion, and saw five-year default protection widen to roughly 1,000 basis points in mid-October 2026 reporting. A take-or-pay contract is real revenue even if the customer’s end users never show up. It is not the same thing as a marketplace clearing price.

How to separate the two, using what is actually public:

• Closer to end demand. Census BTOS (representative, repeated, question-sensitive). Ramp and similar card/API spend (selected sample of firms that already use a spend platform, but it is money leaving a bank account to OpenAI or Anthropic, not a supplier investing in a customer). Marketplace GPU rents that rise when labs are not the only bidder. Hyperscaler statements that incremental capacity is sold as it is energized—still management commentary, but paired with Azure/AWS growth rates in filings.
• Contaminated or subsidized. Equity investments that round-trip into chip or cloud purchases. Residual-capacity backstops. Guarantees that only pay if a named lab leases the site. Backlog concentrated in three or four labs that are themselves financed by the same vendors. Internal research clusters that are “100% utilized” and produce no external rental revenue.
• Ambiguous. Token-volume explosions during a price war. Volume can be real usage and still fail to cover depreciation if the price per token falls faster than volume rises—the pattern Ramp and Bank of America both flagged in late summer and early fall 2026.

A central-bank-style claim that a majority of AI-firm investment comes from other AI firms has circulated in secondary write-ups; treat it as contested until the primary statistical release is in hand. The filing-level circular amounts above do not need that statistic to be material.

What is reliably public, and what is not

Metric	How public	Reliability caveat
API list prices	Provider rate cards, daily	Reliable for the SKU; blend and “effective” indexes depend on mix and cache assumptions
On-demand GPU rents	Marketplace listings, hyperscaler price pages, third-party panels (CCIR, cloud-gpus, getdeploying)	Reliable as offers, not as cleared volume; config, region, and interruptibility must be separated
Hyperscaler capex and cloud growth	10-Q / earnings	Reliable; “demand exceeds supply” is commentary
NVIDIA investments, guarantees, purchase commitments	10-Q footnotes	Reliable as disclosed amounts; economic substance of circularity is interpretation
Neocloud backlog and debt	Issuer filings (e.g. CoreWeave)	Contracted revenue, not collected cash; customer concentration is the risk
U.S. firm AI use	Census BTOS, public microdata and working papers	Best representative series; wording change in Nov 2025 breaks the old time series
Enterprise ROI	McKinsey, MIT NANDA, S&P, IDC	Self-report; samples and success definitions differ; not audited EBIT
Fleet GPU utilization	Almost never	Operator claims and nvidia-smi are not interchangeable with billed or productive hours
Inference vs training split	Gartner, Deloitte, McKinsey, Futurum	Analyst estimates; dollars and gigawatts disagree on the exact crossover year
Power shortfall	Sell-side (Morgan Stanley) and some grid telemetry	Model-dependent; UK distribution data shows booked vs used capacity can diverge sharply
“$8.5T vs $3.6T” returns gap	Gartner press release, Oct 14, 2026	A forecast with an assumed 12% hurdle, not an observed loss

The practical reading as of October 20, 2026: frontier cluster time and energized power are scarce, which is why rents on deployed GPUs have risen this year and hyperscalers are still raising capex. Ordinary enterprise GPUs are often idle, which is why utilization surveys look disastrous and why ROI has not followed individual productivity. Token prices are in a real price war that boosts usage faster than spend. A large slice of the most cited “demand”—backlog, chip purchase commitments, cloud revenue between related parties—is contractually real and economically circular. The metrics that can adjudicate the argument without taking a bank’s word for it are Census adoption, card-level API spend, and independent GPU rental panels. Fleet utilization and the training/inference watt split are not in that set.

---

Recent Findings Supplement (October 2026)

Recent data (primarily October 2026 publications, with some September–mid-2026 reports) highlight a widening split between high-utilization hyperscaler/neocloud deployments and widespread underutilization or inefficiency in enterprise and single-tenant settings.[1]

Sedai analysis (Oct 13) found roughly one-third of GPUs operating below 15% utilization, with even systems reporting >95% standard utilization often achieving only ~38% kernel efficiency—leading to estimates of ~$70,000 annual waste per H100-class GPU at 30% effective use. A VentureBeat Intelligence survey (referenced Oct 15) reported 81% of enterprises running AI workloads at ≤50% GPU utilization and 31% below 25%, yet 63% planned infrastructure expansion and only 10% had finance-reviewed ROI figures.[2]

In contrast, Oracle reported 97.9% GPU fleet utilization in its most recent quarter (after adding >300,000 GPUs), with renewed capacity resold at a 20% premium. Bank of America (Oct 19) cited near-100% AI infrastructure utilization and GPU rental price increases of 12–37% year-to-date as evidence of shortage rather than surplus.[3][4]

A new National Compute Grid initiative (launched Oct 13) aims to pool idle capacity, noting single-tenant data centers average <15% utilization and targeting 2 GW by 2030 from an initial 760 MW pipeline (with >$5B in reservations).[5]

This divergence suggests compute scarcity is concentrated among well-capitalized players with committed workloads, while broader buildout risks creating stranded or inefficient capacity. Enterprises face challenges in achieving billable/effective utilization versus raw allocation; new pooled marketplaces or optimization tools (e.g., Sedai) could help entrants compete on efficiency rather than raw scale. Public metrics remain fragmented—enterprise surveys show low averages, while hyperscaler self-reports and rental trackers show tightness.

GPU rental rates have risen overall amid contract lengthening and newer-generation scarcity, even as older capacity shows signs of softening and token/API prices continue sharp declines.[6]

Median on-demand H100 pricing reached ~$3.81–$4.14 per GPU-hour across providers as of mid-to-late October 2026 data (+21% year-over-year; +9% in recent weeks per one tracker), with spreads as wide as 5.9× between cheapest and priciest offers. Earlier 2026 contract examples included ~$4.30/GPU-hour for multi-year B300 clusters. Newer chips (B200/B300) command premiums; spot/interruptible rates are lower but volatile.[7][8]

Token prices have fallen rapidly (e.g., LLM usage benchmark at $0.97 per million tokens after a 29% drop in one month; broader deflation of ~47% in some gateway averages), driven by competition and cheaper open-weight models. Goldman Sachs (early September) warned this could contribute to compute oversupply if usage growth does not outpace price declines.[9]

Rising rentals signal persistent demand for scarce new capacity and long-term contracts, but falling token prices and older-GPU softening point to potential bifurcation—newer silicon remains tight while commodity tiers face pressure. Entrants should monitor generation-specific dynamics and consider flexible or pooled offerings; hyperscalers and specialists with locked-in deals hold advantages. Rental price indices from multiple independent trackers (e.g., Price of Compute, getdeploying) provide relatively reliable public data here.

Power and grid constraints have emerged as the dominant near-term bottleneck, with Nvidia highlighting operational inefficiencies and analysts projecting significant U.S. deficits favoring large players.[10]

Nvidia (Oct 14) stated AI data centers deliver useful compute on only 60–70% of grid power drawn (20% lost to cooling/distribution/conversion; up to 10%+ to failures, restarts, and misconfigurations), with potential for recovery through better design. Morgan Stanley (Oct 17) estimated a 32 GW U.S. data center power deficit through 2028 (~34% of projected needs), warning it will concentrate capacity additions among creditworthy hyperscalers and neoclouds (projected 94% of 2026 global additions, declining later).[11]

Grid interconnection delays of 5–10 years are common; events like a July 2026 PJM fault affecting 3.9 GW of load and regional pauses (e.g., ERCOT) underscore risks. Projects are shifting or using on-site generation (e.g., xAI turbines). Broader pipeline credibility is low—Bernstein estimated only ~35% of planned U.S. capacity as realistic, with dozens of large projects delayed or blocked in 2026.[12]

Power access now differentiates winners more than chip availability alone; smaller players risk being squeezed out or forced to Asia/other regions. New efficiency tools (e.g., dynamic power allocation yielding 24%+ token throughput gains in tests) and flexible-load approaches offer mitigation paths. Public data on deficits and queues (from banks, grid operators, and trackers) is increasingly cited but often modeled rather than fully realized.

McKinsey’s 2026 State of AI survey (fielded May–June, 1,719 respondents) shows enterprise scaling advancing but ROI lagging, consistent with earlier MIT NANDA findings of limited P&L impact; Census data provides a broader U.S. baseline.[13]

McKinsey: 44% of organizations scaling AI enterprise-wide (up from 38% prior year); 37% report any EBIT impact from AI (flat year-over-year); only 6% are “high performers” (≥5% EBIT contribution with significant impact). Individual productivity gains are widespread (80% report improvement), but ~20% cite operating/token costs as a constraint. Agentic AI scaling is rising among large firms.[13]

MIT Project NANDA’s 2025 “GenAI Divide” report (widely referenced in 2026 coverage) found 95% of organizations saw no measurable P&L return from GenAI deployments (based on interviews, surveys, and public cases); clarifications note this reflects limited documented impact rather than outright pilot failure rates. U.S. Census Bureau (Sep 2026): 23.2% of businesses using AI in any function.[14][15]

Adoption is real and broadening, but translating to enterprise ROI remains difficult—high performers redesign workflows; most do not. This supports distinguishing subsidized/circular demand (e.g., internal hyperscaler usage or pre-committed lab capacity) from organic enterprise pull. Surveys like McKinsey and Census offer the most reliable public benchmarks here.

Inference/token demand has surged dramatically, outpacing many forecasts and shifting spend toward inference (now majority of AI-optimized IaaS), with agentic workloads driving higher per-task consumption.[16]

OpenRouter routed ~23.8T tokens/day in October 2026 (33× year-over-year); agentic traffic reached ~7.3T tokens (7-day avg.), up 14× in six months and surpassing human prompting. Inference accounted for 55% of 2026 AI-optimized IaaS spend (~$23.3B vs. $19B training) and is projected to rise further. Open-weight models handled 71.5% of Vercel gateway tokens (23% of spend) in September.[17][18]

Gartner projects agentic workflow inference costs rising >5× through 2028 despite per-token deflation, due to 5–30× higher token use versus basic chatbots. Volume growth (e.g., Google/Microsoft hyperscaler reports of multi-fold increases) appears demand-driven rather than purely subsidized.[19]

This represents tangible usage growth (especially agentic/coding), helping differentiate real demand from circular buildout. Public trackers like OpenRouter/Vercel provide granular, timely signals; hyperscaler commentary offers directional confirmation but less granularity.

Overall, recent evidence points to persistent tightness at the frontier (high utilization, rising rentals for new capacity, exploding inference volume) coexisting with inefficiencies, project delays, and ROI challenges elsewhere. Power/grid issues and execution gaps are the clearest new constraints. Metrics from rental trackers, McKinsey/Census surveys, and grid reports are among the more reliable public sources; utilization and pipeline figures vary significantly by segment and self-reporting. For competitors, advantages accrue to those optimizing existing assets, securing power early, or proving measurable ROI in agentic/inference use cases.

The AI trade is not a pure replay of 1999 or 1845, but the financing mix is shifting toward the version of those episodes that produced systemic pain rather than merely expensive equity losses. Forward multiples on the profitable leaders look ordinary; the cyclically adjusted market, the concentration, and the scale of the build-out do not.

Where the numbers sit in October 2026

Nvidia is a $5.24 trillion company trading at a trailing P/E of about 27.5 and a forward P/E in the high teens to mid-20s, depending on the earnings year used, after revenue more than doubled and data-center sales continued to dominate. That is not the 100-plus trailing multiples that defined the Nasdaq at the 2000 peak. The index-level picture is more stretched on long-cycle measures than on next-year earnings. The S&P 500’s forward 12-month P/E is 19.0, slightly below its five- and ten-year averages, while the Shiller CAPE sits near 40.5–40.7—the second-highest reading since 1881, behind only the December 1999 peak of 44.2. Concentration is the clearest parallel to past manias: a JPMorgan 41-stock AI basket accounts for roughly 49% of S&P market cap (second only to 19th-century railroads in thematic weight), the information-technology sector is about 40%, and the top ten stocks are around 39–40%, with Nvidia alone near 7.5–8%.

Capex is the variable that has moved fastest. Epoch AI estimates AI-related computing infrastructure (above-trend computers, data-center construction, and networking) reached roughly 1.5% of U.S. GDP by early 2026, about double the 2015–2022 average. Longer projections are larger still. Columbia economist Stijn van Nieuwerburgh’s Brookings work puts cumulative AI infrastructure investment near $9–10.3 trillion from 2025 to 2032, averaging roughly 3.2–3.6% of GDP per year and, in one draft, peaking near 5%—above the multi-decade averages he assigns to U.S. railroads, highways, telecom, canals, and electrification. Goldman has cited about 1.9% of GDP for AI investment in 2026 alone. Definitions differ (cash capex versus leases, announced versus completed, U.S. versus global), so the precise share is contested; the direction is not.

What this means for anyone competing or allocating capital: the market is pricing a durable earnings boom in the chip and cloud leaders while the CAPE and concentration metrics price a once-in-a-century thematic bet. The gap between those two lenses is where the argument lives.

Three historical episodes, and what the financing structure did

British Railway Mania (peak planning 1845–46, investment peak ~1847). Railway investment reached about 7% of GDP and roughly half of all investment. Shares were often sold with only a 10% deposit, equity was the intended financing vehicle, and promoters promised 10% dividends. Reality delivered ordinary-share dividends near 1.8% by 1850. Many lines were never economic; George Hudson’s “Railway King” structure paid dividends out of fresh capital. The physical network largely survived and was later refinanced with a rising share of preference shares and debentures. Equity holders of the speculative companies were the primary losers; the infrastructure became the backbone of the Victorian economy.

U.S. railroad boom and the Panic of 1873. Track mileage roughly doubled from 1865 to 1872. Bonds financed more than three-quarters of the expansion. Jay Cooke’s failure on Northern Pacific bonds in September 1873 closed the New York Stock Exchange for ten days, bankrupted on the order of 100 railroads (counts vary from the mid-50s in the first wave to 121 in broader tallies), and helped produce a depression that ran into 1879. A third of railroad bonds stopped paying interest within two years. Again the network survived—reorganized under bondholders—and later giants (Vanderbilt interests, successors to the failed roads) compounded. Debt turned an overbuild into a banking crisis rather than a contained equity wipeout.

Dot-com and the telecom/fiber supercycle (1996–2002). The Nasdaq fell about 78% from the March 2000 peak and took roughly 15 years to reclaim it. Technology-sector forward multiples peaked near 58 times versus something closer to the mid-20s for leading tech today; large fractions of the Nasdaq-100 were either unprofitable or above 60 times earnings. Telecom carriers spent on the order of $500 billion on the false premise that internet traffic doubled every 100 days (it was closer to annually). Estimates put 85–97% of long-haul fiber dark after the bust. Telecom equipment and carrier capex peaked around 0.8–1.2% of U.S. GDP—below current AI projections on most measures. Vendor financing (Lucent, Nortel lending customers the money to buy gear) created circular revenue. WorldCom’s bankruptcy was about $107 billion; Global Crossing went from a $47 billion valuation to Chapter 11. Cisco fell roughly 86% and did not make a new high until late 2025. What survived: the fiber itself (bought out of bankruptcy by Level 3 and others, later carrying cloud and streaming traffic), Amazon, and eventually Microsoft and the search/advertising winners that were not the 1999 leaders. What failed: most pure-play carriers, the equipment vendors’ equity, and any business model that required continuous external capital to cover operating losses.

The mechanism that separates outcomes is leverage and circularity. Equity-financed manias destroy shareholder wealth and leave cheap assets for the next owner. Debt-financed manias, especially when banks or vendor finance are in the loop, transmit losses into the credit system and lengthen the bust. Today’s hyperscalers funded the first leg mostly from operating cash flow—closer to a healthy industrial build-out than to 1999 startups. That is changing. Alphabet, Amazon, Meta, and Oracle had issued more than $200 billion in bonds by mid-October 2026, more than double all of 2025 in some tallies. Morgan Stanley has described hyperscaler cash burn and a cumulative free-cash-flow shortfall on the order of $600 billion through 2028 as spending outruns operating cash flow, with lease financing layered on top. Neoclouds such as CoreWeave carry heavy near-term maturities and negative free cash flow while scaling capex into the tens of billions. Circular arrangements—chip vendors investing in customers who then buy the chips, or capacity deals between labs and clouds—echo Lucent’s vendor financing more than they echo a simple Microsoft capex cycle.

For a new entrant or a lender, the historical lesson is not “infrastructure always wins.” It is that the equity of the builders often does not, and that debt turns a utilization miss into a credit event. The fiber and the rails were valuable at a fraction of construction cost. The original capital structures were not.

What failed, what compounded, and why utilization mattered more than the narrative

In every case the technology was real and the long-run demand arrived. British and American railways cut transport costs permanently. Dark fiber became the substrate for Netflix, cloud, and mobile backhaul a decade later. The internet did reorder commerce. The investors who bought the peak equity or the marginal debt did not capture that value; the buyers of distressed assets and the later application-layer winners did.

Failure modes were consistent:
- Demand was overestimated in the near term (traffic doubling every 100 days; every authorized railway line earning 10%).
- Capacity was built ahead of paying customers, so utilization collapsed before depreciation schedules assumed.
- Financing assumed refinancing would always be available. When it was not, equity went to zero and bondholders took the assets at a discount.
- Accounting and circular revenue masked the gap (WorldCom, vendor finance, dividend-from-capital railways).

Survival modes were also consistent: fortress balance sheets (Microsoft in 2000 had profits and cash; today’s Nvidia, Microsoft, Alphabet, and Amazon still do, even as free cash flow is being consumed by capex), businesses with end-user revenue that did not depend on the next round of infrastructure spending, and acquirers who bought capacity after the bankruptcy courts reset the cost basis.

The non-obvious implication for AI is depreciation speed. Railroad track and fiber ducts last decades. GPUs have useful lives measured in a few years as architectures turn over. Burry and others have argued that hyperscalers are under-depreciating, which flatters current earnings and understates the recurring capex needed just to stand still. If that is right, the “infrastructure survives, equity doesn’t” pattern is harsher here than in rails or fiber, because the asset itself obsolesces before the debt is paid down.

The bears: strongest points and track records

Michael Burry has consolidated shorts into puts on the Nasdaq-100 and the VanEck Semiconductor ETF (SMH), expiring in 2027, and has said he expects a significant top within zero to nine months, having pulled his base case forward from 2028. His substantive points are depreciation (GPUs as a recurring expense, not a one-time outlay), concentration in the hottest semiconductor names inside SMH, and the shift from cash-funded to debt-funded GPU purchases. Track record: the 2005–2008 housing short was one of the best-timed macro bets of the century. Subsequent public calls have been early or wrong often enough that “Burry is short” is not itself a timing signal; the depreciation and financing arguments still have to be judged on the cash-flow math.

David Einhorn has focused on circular revenue. In Greenlight’s framing, a dollar of loss-making end-customer spend can cascade into several dollars of reported revenue across chipmakers, neoclouds, and clouds, so aggregate “AI revenue” overstates economic profit. He has said the spending numbers are so large that a tremendous amount of capital destruction is plausible even if the technology exceeds forecasts. Track record: exceptional from the mid-1990s through the financial crisis (including early skepticism of structured credit); Greenlight’s returns have lagged badly for much of the past decade. The circularity critique does not require him to be right on the stock market’s next 20%.

Howard Marks is the most careful of the group. In late 2025 and again in October 2026 he has argued that a classic bubble requires investor obliviousness to risk, and that widespread bubble talk is evidence that obliviousness is missing. He notes that forward P/Es on the largest tech names are around 23 times versus roughly 65 times for the top tech cohort before the 2000 peak, and that today’s leaders have real revenue and cash flow. He still flags debt-funded expansion, circular financing, and the chance of an “inflection bubble,” and has suggested it may take about a decade to know whether the infrastructure earns its cost. Track record: Oaktree’s credit record and his 1999–2000 and 2007 memos are among the better documented examples of identifying excess without pretending to time the top. His current stance is “elevated and risky, not yet a mania,” which is a narrower claim than the outright bears.

Daron Acemoglu (Nobel 2024) argues from task-level economics, not price charts. His assessment is that AI may automate only about 5% of human tasks over the next decade and add on the order of 1.5% to GDP over ten years—not a productivity revolution—and that firms using the tools are not yet showing large measured gains. The electricity analogy he and others use cuts both ways: the technology was real, and the economic payoff lagged the capital spending by decades. Track record: this is academic forecasting, not a trading record. If he is even roughly right on the GDP increment, the revenue required to justify multi-trillion-dollar capex (Goldman researchers have sketched a path that needs on the order of $1 trillion in annual AI-related revenue for acceptable returns on near-term data-center spend) does not appear.

Ed Zitron argues from the demand side of the compute market. His claim is that OpenAI and Anthropic represent the bulk of incremental AI compute demand, that much of their enterprise revenue comes from other unprofitable AI startups, and that non-lab compute purchases are far too small to absorb the capacity now under construction if the labs cannot pay. He treats data-center debt and project cancellations as the tell, and rejects the idea that idle GPUs will be productively reused the way dark fiber was. Track record: a consistent critic and journalist, not an investor. The mechanism—concentration of paying demand in two loss-making labs—is checkable as revenue disclosures improve; the “capacity is worthless” claim is stronger than the fiber precedent and depends on GPU obsolescence.

Gary Marcus locates the problem in the technology, not only the capital structure. Without reliable world models, he argues, generative systems remain too error-prone for the high-value workflows that would have to fund the build-out, so the economics never close even if capex is somehow financed. He has called 2025 a peak-bubble year and expects the recognition of inherent limits to unwind the trade. Track record: early and correct on hallucination and reliability limits; repeatedly early on commercial timelines. The technical critique explains why enterprise ROI surveys keep showing thin measurable gains; it does not by itself date a market top.

The bulls: strongest points and track records

Jensen Huang’s recent line is that the scare is temporary because the technology is transformative across industries, that frontier labs are only about a quarter to a third of Nvidia’s demand, and that multiple labs and sovereign and enterprise buyers are growing. He has told investors to stay on first principles and that “all of this is going to pass.” The strongest supporting fact is operational: Nvidia remains supply-constrained on management’s own description, with data-center revenue still growing triple digits in recent quarters and gross margins near 75%. Track record: he correctly bet the company on accelerated computing years before the revenue showed up. CEO optimism is not an independent valuation check; the order book and the breadth of buyers are.

Sam Altman has said investors are overexcited and that someone will lose a phenomenal amount of money, while also saying OpenAI will spend trillions on data centers and that the social value will be enormous. OpenAI’s signaled annualized revenue has been noisy—figures around $50 billion for a recent month were below an earlier ~$70 billion indication, with the gap partly blamed on comparison methodology—while Anthropic’s run-rate has been described as moving from single-digit billions toward tens of billions very quickly. Track record: he scaled a research lab into the demand center of the industry. The tension in his own comments is the bull case and the bear case in one sentence: the spending is necessary and someone still gets destroyed.

Dan Ives calls this the third inning of a nine-inning game, keeps Nvidia as the “bedrock” with a $300 target, and argues the constraint is supply (chips, power, memory), not demand, with a multi-trillion spending wave still ahead and secondary software winners only beginning. Track record: a high-profile tech bull who has been directionally right through the AI rerating and whose price targets are often ambitious relative to subsequent prints. Useful as a statement of the institutional bull consensus; not a source of conservative underwriting.

Goldman Sachs and Morgan Stanley are financing the build-out and describing it. Goldman wealth and research voices have argued that concentration and capex growth do not by themselves mark a peak, that earnings—not multiple expansion—have carried much of the index gain, and that hyperscalers can still fund a large share of the spend. CEO David Solomon has simultaneously warned that estimates north of $10 trillion through 2032 imply a financing test: not every project will clear the market if rates or risk premia rise, and demand for compute currently exceeds supply only if the capital is actually raised. Morgan Stanley has been more explicit about the cash-flow gap, projecting AI compute investment well above a trillion dollars in 2027 and a large cumulative free-cash-flow shortfall for the top hyperscalers, filled by bonds and leases. Track record: both firms were constructive on tech through the post-2022 rally and are now earning fees on the debt and equity that fund the next leg. Their research is most useful where it quantifies the funding gap rather than where it reiterates the demand story.

How to read the comparison without forcing a rhyme

The cleanest distinctions from 1999 are profitability and starting multiples at the top of the leaderboard. Nvidia, Microsoft, Alphabet, Amazon, and Meta have enormous earnings; Cisco in 2000 was profitable too, but the surrounding ecosystem was not, and index multiples were far higher. The cleanest similarities to the dangerous historical cases are concentration, capex as a share of GDP on the projections that matter for 2027–28, and the migration from internal cash flow to debt and circular contracts. Rails and fiber show that a real general-purpose technology can still ruin the capital that built it if utilization lags the depreciation and debt schedule. The AI-specific wrinkle is that the core asset depreciates in technology generations, not in decades, so a miss on end-user revenue (Acemoglu’s modest GDP add, Zitron’s narrow payer base, Einhorn’s circular math) does not leave a 40-year duct in the ground. It leaves a warehouse of last-generation GPUs and a stack of bonds.

A practical reading, not a forecast: the forward P/E of 19 on the S&P and the high-teens to mid-20s forward multiple on Nvidia say the market already discounts some normalization of growth. The CAPE near 41, a 49% AI thematic weight, and a debt-funded capex wave sized above prior infrastructure booms say the residual risk is not a modest multiple compression. It is a utilization or refinancing shock that reprices the builders the way 1873 and 2001 repriced railroads and carriers—while the technology, bought later at a lower cost of capital, keeps compounding for someone else.

---

Recent Findings Supplement (October 2026)

Recent October 2026 analyses highlight elevated but not extreme valuations relative to dot-com peaks, with Shiller CAPE at 40.5–40.7× (near the 44.2× 1999–2000 record) and S&P 500 top-10 concentration at a record ~42% of market cap (Nvidia + Apple + Microsoft alone ~21%). Forward P/Es for major AI names (Alphabet, Meta, Microsoft, Nvidia) sit at roughly 22–25× as of early October 2026, at or slightly above the Nasdaq-100 average (~22×), far below Cisco’s ~200× at the 2000 peak; outliers like Palantir (~85×) and AMD (~40×) appear stretched.[1][2][3]

Big Tech (Amazon, Microsoft, Meta, Alphabet, Oracle) is projected to invest ~$729 billion in AI in 2026, with global AI-related capex exceeding $1 trillion; Goldman Sachs estimates U.S. AI investment at ~1.8% of GDP in 2026, rising toward 2.5% in 2027.[4][5]

Multiple October 2026 reports draw the closest parallel to 1840s UK railway mania (peak investment 5.7–6.7% of GDP) rather than dot-com or telecom, noting that only railways exceeded the current AI buildout in scale relative to the economy; U.S. railways peaked near 4% of GDP in 1872. Dot-com telecom/fiber capex reached only ~1.2% of GDP in 2000. Infrastructure from past buildouts (rail networks, fiber) frequently survived and powered later growth even as many financing vehicles and overbuilt operators failed.[4][5][6]

Equity financing by cash-rich hyperscalers has supported the current wave with stronger balance sheets and higher RoE (~46% for Magnificent 7 vs. ~28% for late-1990s IT leaders) than debt-heavy telecom or railway episodes, where leverage amplified failures.[7]

Michael Burry (October 2026 commentary) flagged Anthropic’s $965 billion May 2026 private valuation (potential $2 trillion IPO discussions) as a bubble, arguing it could purchase 78 profitable S&P 500 companies (examples: Domino’s, Clorox); he has also questioned Microsoft’s AI spending pace and timing of any revenue shortfall. Burry’s prior calls (2008 housing) were highly accurate; recent skepticism focuses on one disappointing revenue season potentially triggering a shift in hyperscaler capex priorities.[8][9][10]

Howard Marks (October 18, 2026) argued it is not yet a classic bubble because widespread discussion of risks indicates investors are not “oblivious”; top-five tech forward P/Es average ~23× versus ~65× pre-2000 peak.[11]

Ed Zitron (October 2026 interviews/podcasts) emphasized that ~80% of OpenAI/Anthropic revenue comes from unprofitable AI startups reliant on finite VC funding, with audited figures showing heavy losses and massive future compute commitments ($518 billion cited for one); he views much reported revenue as non-recurring or resold and warns of widespread data-center underutilization post-bubble.[12][13]

Dean Baker’s CEPR AI Bubble Monitor (October 18–19, 2026) highlighted CoreWeave’s debt distress (yields >16%, CDS implying ~63% 5-year default probability) and broader neocloud/private-credit exposure.[13]

Jensen Huang (October 19, 2026 podcast) expressed confidence that demand extends beyond OpenAI/Anthropic to xAI/Grok, Meta, Google Gemini and others, with rising system pricing and content-per-gigawatt; he views the buildout as early-stage. Dan Ives (September 30, 2026) called the environment a “1997 moment, not 2000,” citing that only ~15% of a projected $4–5 trillion AI spend has occurred and noting Nvidia’s large buyback increase.[14][15]

Goldman Sachs (September 24, 2026 research) estimated hyperscalers need ~$300 billion annual AI revenue in coming years for breakeven on 2026–2027 capex (higher for strong returns); it noted AI infrastructure stocks de-rated from ~32× to ~22× median forward P/E since April 2026, with hyperscalers at decade-low premiums to the S&P 500.[16]

October 2026 developments spotlight debt stress in neoclouds: CoreWeave (~$50 billion debt, yields 15–16% on bonds, heavy private-credit facilities from Blackstone, Magnetar, Carlyle, etc.) faces sharply higher default risk despite Nvidia backstops and customer contracts (Microsoft, OpenAI, Meta); this contrasts with hyperscalers’ equity-funded model. Past episodes show debt-financed overbuilds (telecom, some railways) produced sharper failures and larger investor losses than equity-heavy ones.[17][18]

For competitors or entrants: Hyperscaler equity strength and proven earnings growth provide a buffer versus historical debt-driven busts, but neocloud leverage and reliance on a narrow set of unprofitable startup customers create concentrated downside risk if revenue disappoints. Infrastructure assets may retain long-term value even amid company failures, as in railways and fiber. Recent de-rating in multiples and vocal bear commentary suggest some froth has already been priced out, but sustained capex requires rapid monetization beyond current backlogs.