ResearchNo. 02
AI Infrastructure Capital Flows and Risk Transmission
Debt-funded investment faces a 41% drop in effective token prices.
- Research no.
- 02
- Published
- 2026-10-26
1. Executive Summary
Accounting profits enter balance sheets through equity revaluations. In contrast, physical capital expenditures are financed with debt, with the incremental debt share of capex expanding to 32%.
Contractual risk is transferred down the supply chain to model companies. Two model companies carry roughly $1.2 trillion in committed spending. The key variable setting market prices is the delivery timeline.
End demand volume grows while token prices fall. According to Ramp data, the effective price paid by enterprises per 1 million tokens dropped 41% over six months. If this gap narrows, contractual losses propagate in order through neoclouds, model companies, guarantors, and hyperscaler earnings per share.
2. Questions and Methodology
This report tracks the funding sources and flows of capital deployed to build AI infrastructure in 2026. It identifies whose balance sheets record accounting profits under which line items, and establishes the contractual order of loss transmission if end-user software demand falls short of market expectations. Furthermore, it outlines the trajectory along which the allocation of capital and risk reorganizes through 2030.
The analysis divides capital flows into three streams. The first is reported profit, verified by separating operating income from non-operating income on income statements. The second is actual cash flow, aggregated from operating cash flow, capital expenditures (Capex), equity and bond issuances, and uncommenced lease commitments. The third is the contractual risk transmission channel established through take-or-pay terms, termination rights, debt guarantees, and residual value guarantees. Primary sources include public regulatory filings (10-Q, 10-K, 8-K, S-1) and official disclosures from Amazon, Alphabet, Microsoft, Meta, Nvidia, Oracle, CoreWeave, Nscale, and SpaceX. Financial figures for private companies Anthropic and OpenAI rely on press reports (WSJ, Reuters, FT, The Information) citing internal documents, explicitly identified as reported figures in the text. The report then reviews the constraints setting infrastructure rental prices alongside trends in token demand, revenue, and return on invested capital (ROI) to summarize conditions for sustaining this capital loop, deriving the layers exposed to losses across three scenarios through 2030.
Clear assumptions and limitations apply to the data. All performance metrics for private model companies depend on media reports. Annualized run-rates calculated by multiplying a single month’s revenue by 12 diverge from final annual revenue. For instance, OpenAI’s actual 2025 revenue was $13.07 billion, but its year-end annualized run-rate was reported at $21.4 billion. Recognition methods for cloud marketplace sales also differ. Anthropic records sales on a gross basis, while OpenAI records them on a net basis, imposing structural limits on direct aggregation. With the exception of Meta, major hyperscalers have third-quarter 2026 earnings releases scheduled for October 28–30, and are not included as of the report date of October 26, 2026. Meta figures reflect its third-quarter 10-Q filed in October. Third-quarter 2026 financial results released after this report was written are recorded in Section 11.
3. Equity Valuation Gains and the Accounting Loop
3.1 Valuation Gains Outpacing Core Operating Income
Income statements of Big Tech firms and semiconductor manufacturers show non-operating gains from revaluing equity stakes in investees exceeding operating income earned from core businesses, such as cloud and chip sales.
| Company | Quarter | Equity Valuation Gain | Core Business Income for Same Period | Ratio (author’s calculation) | Source |
|---|---|---|---|---|---|
| Amazon | Q2 2026 | Non-operating pre-tax other income of $53.4 billion (primarily Anthropic investment), including ~ $50.5 billion upward adjustment of Anthropic non-voting preferred stock | AWS operating income of $16.6 billion (revenue $42.2 billion, operating margin 39.4%); total company operating income of $27.5 billion | 3.2x AWS operating income, 1.9x total operating income | 1,2 |
| Alphabet | Q2 2026 | Other income of $98.0 billion (primarily unrealized valuation gains on equity securities), net income contribution of $77.1 billion after tax, $6.26 per share | Diluted EPS of $9.11 | 69% of EPS | 3,4 |
| Microsoft | FY2026 (ended 2026-06-30) | Net gain on OpenAI investment of +$5.0 billion (-$3.6 billion in prior year), largely dilution gain from recapitalization | — | — | 5 |
| Nvidia | H1 FY2027 (ended 2026-07-26) | Net gain on equity securities of $23.71 billion (including $12.5 billion from publicly traded shares); carrying value of non-marketable securities increased from $22.25 billion to $51.16 billion | Half-year net income of $118.01 billion (after tax) | Pre-tax valuation gains represent 20% of net income | 6 |
Amazon’s non-operating pre-tax other income of $53.4 billion in Q2 2026 was, by the author’s calculation, 3.2 times AWS operating income of $16.6 billion and 1.9 times consolidated operating income of $27.5 billion for the same period1,2. Similarly, out of $98.0 billion recorded as other income by Alphabet in Q2 2026, $77.1 billion contributed to net income. This accounted for $6.26 of its total diluted EPS of $9.11, representing 69% by the author’s calculation3,4.
These equity valuation gains link directly to large-scale capital partnerships and compute supply contracts with investee model companies. As of June 30, 2026, Amazon carried securities related to Anthropic valued at $92.5 billion in non-voting preferred stock and $97.9 billion in convertible notes2. Amazon executed $5.0 billion in a Series G round and $5.0 billion in a Series H round during the second quarter, leaving $21.3 billion in remaining investment commitments2. This structure stems from an April 2026 agreement where Anthropic committed to spend over $100 billion on AWS infrastructure over 10 years, and Amazon agreed to invest a baseline of $5.0 billion with up to $20.0 billion in additional capital7.
Alphabet’s carrying value of non-marketable equity securities measured under the measurement alternative expanded from $64.1 billion at year-end 2025 to $124.3 billion as of June 30, 2026. In Q2 disclosures, $87.9 billion of this reflected fair value remeasurements8. Alphabet’s remaining funding commitments decreased from $30.0 billion to $20.0 billion during the quarter. While the filings do not name the counterparty, media reports identify this as a commitment to Anthropic4. Alphabet’s initial investment in Anthropic in April 2023 was $300 million4. Microsoft acquired an approximate 27% stake in OpenAI Group PBC following a corporate recapitalization executed in October 2025. In connection with this, OpenAI agreed to purchase an additional $250 billion in Azure cloud infrastructure9.
3.2 The Multiple-Booking Structure of Backlog and Valuation Gains
Behind reported valuation gains lies a structure where capital tied to a single compute agreement cycles across multiple financial statements, amplifying metrics.
Tracing this accounting path through transactions between Amazon and Anthropic reveals three stages (author’s calculation). First, when Anthropic commits to spend over $100 billion over 10 years for AWS infrastructure, that entire sum enters AWS’s remaining performance obligations (RPO). AWS backlog reached $496.0 billion in Q2 2026, roughly 2.5 times higher year-over-year according to reports citing company statements10. Second, 47% of Anthropic’s revenue ($2.16 billion in 2025) is generated through AWS and Google marketplaces, and Anthropic books these transactions on a gross basis in its revenue. In this process, platform fees of roughly $351 million (~16 cents per dollar of revenue) are recorded as marketing expenses. Amazon and Google held 60% of Anthropic’s $909 million in accounts receivable as of year-end 202511. Third, as Anthropic’s valuation rises on the back of this agreement, Amazon books a $50.5 billion fair value upward adjustment on its equity holdings as non-operating income2. A single compute purchase agreement converts simultaneously into three financial metrics: order backlog, gross revenue, and equity valuation gains.
Combined order backlogs across major tech firms expanded rapidly. Microsoft’s commercial RPO grew 25% excluding OpenAI, but total RPO increased 84% to $678.0 billion, with roughly 30% recognized as revenue over the next 12 months12. Google Cloud’s backlog stood at $514.0 billion, including TPU hardware supply contracts recognized starting in Q1 202613. Simple addition of AWS’s $496.0 billion10, Oracle’s $664.0 billion (up $209.0 billion year-over-year)14, and CoreWeave’s $103.7 billion15 yields approximately $2.46 trillion (author’s calculation). However, multi-year commitments from OpenAI and Anthropic are booked across multiple hyperscaler balance sheets. Goldman Sachs estimates net backlog, removing overlap, at over $1.7 trillion16. BofA assesses that over half of Oracle’s backlog is concentrated in OpenAI alone17.
On the other side, model developers record large losses mirroring these valuation gains. Draft IPO filings reported in the media show that out of Anthropic’s $42.0 billion net loss in 2025, roughly $34.0 billion stemmed from non-cash fair value remeasurement charges on issued financial instruments, as did roughly $72.0 billion of its $80.7 billion net loss in H1 202618,19. Leaked audited financial statements reported for OpenAI’s 2025 performance showed an operating loss of $20.92 billion, which, combined with $41.55 billion in valuation losses on convertible equity and warrants, yielded a consolidated net loss of roughly $60.4 billion and a net loss attributable of $38.5 billion20,21. OpenAI paid Microsoft $10.59 billion in compute fees out of $19.18 billion in 2025 research and development expenses, equivalent to 81% of its $13.07 billion revenue that year22.
International financial institutions issued warnings regarding these interconnected structures. In its Global Financial Stability Report (2026-10-13), the IMF reviewed 14 AI transactions announced since 2023, observing that roughly $300 billion in equity and $650 billion in debt support roughly $1.4 trillion in future revenue commitments. It identified the mechanism where declines in underlying asset values extinguish linked future revenue as “wrong-way risk”23. The BIS Annual Economic Report (2026-06-28) also pointed out that circular financing terms—where chipmakers and hyperscalers acquire stakes in model developers while developers commit to long-term compute purchases—suffer from opaque disclosure, citing the risk of single assets being pledged as collateral across multiple balance sheets24.
Book profits and backlogs expand through equity revaluations and multi-year commitments. In contrast, actual cash financing for physical capital expenditures follows a different trajectory.
4. Capex Financing Channels and Debt Structures
4.1 Capex Outrunning Operating Cash Flow Generation
Actual cash spent on expanding infrastructure outpaces operating cash inflows.
| Company | Period | Operating Cash Flow | Capex | Free Cash Flow | Source |
|---|---|---|---|---|---|
| Amazon | Trailing 12 months to Q2 2026 | $161.4 billion (+33%) | Property and equipment purchases of $173.0 billion (net $169.0 billion) | -$7.6 billion | 1 |
| Alphabet | Q2 2026 / FY2025 | Not disclosed for quarter / FY2025 $164.7 billion | Quarterly $44.9 billion / FY2025 $91.4 billion; 2026 guidance $195.0–$205.0 billion (raised twice from $175.0–$185.0 billion in Feb) | Quarterly -$5.9 billion | 3,25,26 |
| Meta | Jan–Sep 2026 | $96.87 billion | Property and equipment purchases of $83.06 billion + principal payments on finance leases of $2.77 billion; full-year 2026 guidance $125.0–$145.0 billion | — | 27 |
| Oracle | Q1 FY2027 (ended 2026-08-31) | $23.1 billion ($11.4 billion from customer advances with significant financing components) | $28.5 billion ($8.5 billion in prior-year period) | -$5.4 billion; sold $20.0 billion common stock at-the-market during quarter | 14 |
| Top 5 Aggregate | Trend | +23% YoY | +70% YoY | Aggregate zero-crossing point around Q3 2026 (Oracle already entered, Amazon entered, Alphabet expected Q1 2027, Meta Q3 2027, Microsoft Q3 2028) | 28 |
Hyperscaler capital expenditures outpace cash generation. Oracle’s Q1 FY2027 capex of $28.5 billion exceeded its quarterly revenue of $19.3 billion. Relative to operating cash flow of $11.7 billion excluding customer advances of $11.4 billion, capex reached 2.4 times pure operating cash flow (author’s calculation)14. S&P downgraded Oracle’s credit rating to BBB- on July 9, 2026, estimating its FY2027 free operating cash flow deficit at roughly $42.0 billion29. FactSet data shows that aggregate free cash flow for the five hyperscalers in FY2026 approaches zero or turns negative when excluding Alphabet and Microsoft30. Epoch analysis indicates that while combined operating cash flow for the five companies grew 23% annualized, capex grew 70% annualized, putting combined free cash flow on a trajectory to turn negative around Q3 202628.
4.2 Debt Expansion and Equity Offerings
To cover shortfalls in operating cash flow, leading companies increased their reliance on debt and equity issuances.
According to FactSet, the share of capex funded by incremental debt rose from 9% in FY2024 to 32% during the 12 months ending June 202630. Annual corporate bond issuance by hyperscalers reached roughly $121.0 billion in 2025—about four times the 2020–2024 average—before climbing to $159.0–$170.0 billion in H1 2026 alone31. JPMorgan projects full-year 2026 issuance to reach $317.0 billion32. Amazon executed its largest bond sale on record in March 2026, raising $37.0 billion and €14.5 billion (roughly $53.8 billion total)33. Alphabet raised roughly $32.0 billion in a single day in February, including a £1.0 billion 100-year bond (coupon 6.125%), the first century bond from a tech company since 199734. Meta issued $25.0 billion across six tranches in May 2026, bringing its outstanding bond principal to $84.0 billion27. Oracle attracted $129.0 billion in orders when issuing $25.0 billion across eight tranches in February35.
Direct capital raises through equity markets ran parallel to debt issuance. On June 3, 2026, Alphabet executed an equity offering totaling $84.75 billion. The financing consisted of $18.0 billion in common stock, $16.75 billion in mandatory convertible preferred depositary shares, a $40.0 billion at-the-market (ATM) program, and a $10.0 billion private placement to Berkshire Hathaway. This marked the first large-scale equity offering by a hyperscaler during this AI infrastructure investment cycle36.
Nvidia, at the apex of the supply chain, also expanded debt issuance. During H1 FY2027, Nvidia generated $74.4 billion in operating cash flow while deploying $39.0 billion for share repurchases and $6.3 billion for dividends. Alongside this, it issued $24.9 billion in bonds, lifting senior notes outstanding to $33.5 billion6. In late September 2026, its board increased its share repurchase authorization by $150.0 billion to a remaining unused capacity of $235.0 billion, planned for execution through January 2028, according to reports37. A semiconductor designer borrows to repurchase shares while reinvesting funds into customers and infrastructure partners.
4.3 Large-Scale Off-Balance-Sheet Lease Commitments
Beyond public bond and equity issuance, large-scale infrastructure lease commitments unrecorded on official balance sheets are expanding.
| Company | Uncommenced Lease Commitments | Comparison | Source |
|---|---|---|---|
| Meta | Increased from $103.8 billion as of Dec 31, 2025 to $355.7 billion as of Sep 30, 2026 (data centers, energy infrastructure, colocation, covering 2026–2036) | 4.2x outstanding bond principal of $84.0 billion; non-cancellable contractual obligations of $454.5 billion ($45.8 billion payable in 2026, $121.4 billion payable in 2027) | 38,27 |
| Microsoft | Increased from $196.6 billion as of Jun 30, 2025 to $329.1 billion as of Jun 30, 2026 (primarily data centers, commencing FY2027–FY2033) | 4.9x finance lease liabilities of $66.6 billion | 5 |
| Alphabet | As of Dec 31, 2025: long-term $52.7 billion and power purchase agreements of $9.9 billion for 2027–2047; non-cancellable purchase commitments of $149.1 billion | Total purchase commitments and contractual obligations of $811.0 billion as of Jun 30, 2026 ($200.7 billion short-term); fixed commitments over 1 year for long-term supply, energy, and content of $707.0 billion (mostly tech infrastructure supply contracts through 2030, energy contracts 2–26 years through 2054 with take-or-pay terms) | 25,8 |
| Top 5 Total | $662.0 billion as of year-end 2025 | 113% of adjusted gross debt; undiscounted gross lease commitments of $969.0 billion, defined by Moody’s as implicit debt | 39 |
Under GAAP accounting, uncommenced leases are not recognized as balance sheet liabilities until facilities are completed and operational. Moody’s calculated that uncommenced lease commitments of $662.0 billion across five hyperscalers at year-end 2025 represented 113% of adjusted gross debt, characterizing them as “implicit debt”39. Meta’s uncommenced lease commitments expanded from $103.8 billion at year-end 2025 to $355.7 billion by late September 2026, reaching 4.2 times its $84.0 billion in public debt outstanding38,27.
Changes in accounting estimates alter reported figures. Starting in FY2027, Microsoft extended the estimated useful lives of its data center buildings from 15 years to 25 years. Consequently, many upcoming data center lease agreements were reclassified from finance leases (included in capex) to operating leases (omitted from capex). As a result, reported capex for 2026 decreased by roughly $15.0 billion, from roughly $190.0 billion to roughly $175.0 billion, though the CFO stated that “underlying expectations for investment in infrastructure remain unchanged”40.
Morgan Stanley termed data center leases omitted from cash flow capex as “synthetic capex,” adding them to capex totals (~$195.0 billion in 2027, $246.0 billion in 2028)170,171. Hyperscalers’ disclosed uncommenced lease commitments surpassed $1.1 trillion, representing an increase from Moody’s year-end 2025 figure of $662.0 billion170,171. Hyperscalers issued roughly $255.0 billion in senior unsecured debt in 2026, funding roughly 30% of estimated capex171. The year 2027 is projected to show the widest cash flow deficit across the forecast horizon171.
4.4 Financing Spread Polarization by Guarantee Status
Debt structures raised on and off balance sheets split into two tiers with distinct yields and underwriting terms depending on whether investment-grade corporate guarantees are present.
| Borrowing Entity / Structure | Amount | Interest Rate | Guarantee / Collateral | Source |
|---|---|---|---|---|
| Anthropic Big Sky (chip leasing SPV, ~1 GW of TPUs at Fluidstack site) Senior A1/A2 | $30.0 billion | 5.75% | Broadcom guarantee | 41,42 |
| Same SPV Subordinated B | $4.4–$4.5 billion | ~8.5% | No guarantee, KBRA rating BB+ | 41 |
| Anthropic North Star (~1.5 GW of TPUs) | ~ $60.0 billion (Senior ~$42.0 billion with Broadcom guarantee + Subordinated $18.0 billion) | Blackstone proposed >10% rate on $9.0 billion subordinated debt (in negotiation as of Oct 17) | Secured by Anthropic 5-year chip lease payments | 43 |
| Meta Hyperion (Blue Owl 80%, Meta 20%, Louisiana) maturing 2049 | $27.0 billion | 6.581% (~108 bp spread over Meta’s own corporate debt) | Meta 16-year residual value guarantee, S&P rating A+ | 44,45 |
| OpenAI Ohio Campus (SB Energy, 4.25 GW) 20-year lease | — | — | Nvidia guarantee capped at $105.0 billion (payable only upon tenant default, terminates if OpenAI obtains satisfactory rating) | 6 |
| CoreWeave Unsecured Notes maturing 2030/2031 | $2.0 billion and $1.75 billion | Coupon 9.25% and 9.00%; 2032 notes (9.625%) traded around 79.5 cents on Oct 14 yielding ~15% | No guarantee | 46,47 |
| Volta Infrastructure (36,000 GPUs and Norway leased assets) | $5.012 billion | 1st lien $1.3 billion underwritten by JPMorgan/Goldman; GPU-backed $3.712 billion placed best-efforts at SOFR+625–650 bp (total 11.21–11.75%) after banks declined underwriting (re-offered after Aug–Sep SOFR+500 bp failed) | GPU collateral | 48 |
| SpaceX (xAI) Nvidia Chip Debt | ~ $40.0 billion (~$10.0 billion bank loan + ~$30.0 billion IG bonds, early talks targeting 2027 close) | — | No Nvidia guarantee (FT report) | 49 |
Corporate guarantees determine financing costs. In Anthropic’s Big Sky SPV, the spread between the Broadcom-guaranteed senior debt (5.75%) and the unguaranteed subordinated tranche (~8.5%) was 275 basis points by the author’s calculation41. Four months later, for the North Star project’s subordinated tranche, Blackstone proposed an interest rate exceeding 10%, widening the spread beyond 425 basis points43. The spread between unguaranteed, GPU-backed borrowing like Volta’s tranche (11.21–11.75%) and investment-grade guaranteed senior debt reached roughly 550–600 basis points. Meta absorbed an approximate 108-basis-point premium over its own corporate paper to keep Hyperion debt off its balance sheet48,44.
The scale of this credit enhancement appears in contingent liability footnotes. Nvidia’s maximum exposure to guarantees for land, power, and facilities reached $108.5 billion, comprising $105.0 billion for OpenAI’s Ohio lease and $3.53 billion for other AI clouds. This liability amortizes as partners pay rent over 5 to 7 years, with $712 million currently held in escrow6. Meta disclosed a maximum loss exposure of $68.40 billion across unconsolidated ventures in Louisiana ($27.0 billion) and Texas ($14.0 billion). This consists of $4.73 billion in equity, $18.48 billion in lease commitments, $3.70 billion in capital contribution obligations, and up to $41.49 billion in unbooked residual value guarantees judged remote27. Alphabet carried up to $5.7 billion in counterparty financial guarantees and up to $16.9 billion in credit derivatives25, with June 30, 2026 filings specifying up to $7.6 billion in guarantees backing power purchase agreement facility procurement8. Nvidia also extended an initial $6.3 billion agreement to directly absorb infrastructure that CoreWeave fails to sell through April 203250.
In late 2026, credit risks concentrated in this second tier of debt lacking guarantees. CoreWeave’s 5-year CDS spread jumped from 452 bp in June to 855 bp in August, and reached an all-time high of 1,194.5 bp in October. Bond markets implied a five-year cumulative default probability of roughly 60%, and its stock price fell 25% between October 9 and 1647,51. Oracle’s 5-year CDS also set record highs, moving from 144 bp in January to 261 bp in October51. Bonds of the Meta Hyperion joint venture slipped to 91 cents, yielding 7.55%52. Australia’s Firmus postponed its IPO after cutting its offer price by 25%53, and Nscale encountered obstacles in its public listing attempt54. As of October 26, 2026, no formal defaults or loan covenant violations have been documented in the neocloud sector.
The federal funds target range fell by 25 bp each in September, October, and December 2025 to 3.50–3.75%, before rising 25 bp on September 17, 2026 to 3.75–4.00%165. SOFR stood at 3.90% on October 22, 2026166. While the policy rate rose only 25 bp, CoreWeave’s 9–9.75% coupon notes trade at yields near 15%, showing that rising capital costs reflect credit spreads isolated to this second tier.
On December 31, 2025, CoreWeave amended its subsidiary’s DDTL 3.0 credit agreement, reducing the minimum liquidity requirement to $100 million for monthly payment dates between March 1 and May 1, 2026. The amendment postponed initial test dates for the debt service coverage ratio covenant to October 31, 2027, and the contract realization ratio covenant to February 28, 2026. It allowed uncapped equity cures for breaches before October 28, 2026, while limiting cures to 3 out of 4 consecutive months thereafter161. On September 22, 2026, CoreWeave issued $4.2 billion of 2.875% convertible notes due 2033, showing that equity conversion features were necessary to secure lower rates relative to existing 9.000–9.750% debt162. As of June 30, 2026, scheduled principal maturities total $4.413 billion for the remainder of 2026 and $6.184 billion in 2027, summing to roughly $10.6 billion through year-end 2027 (author’s calculation), against total debt principal of $35.551 billion163. Its share price declined 37.75% from $87.12 on September 30 to $54.24 on October 23, according to reports164.
The divergence in debt terms by guarantee status dictates who absorbs losses and performance liabilities if demand weakens.
5. Infrastructure Contract Terms and Risk Allocation
5.1 Take-or-Pay Agreements Concentrated on Model Developers
Upstream infrastructure suppliers used take-or-pay terms to shift idle capacity risks downstream.
| Tier | Contract Parties | Terms | Source |
|---|---|---|---|
| Memory | Micron to customers | 26 take-or-pay contracts secured, guaranteeing over 35% of revenue through 2030; over 75% of 2027 HBM supply contracted | 55 |
| Power / Shell | Data center leases | 60–85% of allocated power set as take-or-pay minimum baseline | 56 |
| Landlord | Hut 8 to Fluidstack (used by Anthropic) | 15-year triple net (NNN) lease of $7.0 billion, 3% annual escalator, Google guarantee; net operating income (NOI) represents 98.6% of rental revenues | 57 |
| Neocloud | Aker-Nscale to Microsoft | ~52,000 GB300 chips, 5-year take-or-pay at $6.2 billion; $2.72 per GPU-hour (author’s calculation) | 58 |
| Neocloud | CoreWeave to customers | Take-or-pay terms, weighted average life ~5 years, upfront deposits 15–25% of total contract value (TCV), contracted commitments account for >98% of 2025 revenue; debt structured as asset-level debt backed by customer contracts | 46 |
| Model Company | Anthropic to Google ($111.0B), AWS ($110.0B), Microsoft ($31.0B) | ~ $450 billion in cloud commitments (7–10 years); draft IPO reports show commitments mostly non-cancellable; total commitments exceed $600 billion including >$160 billion in TPU leases, up to $750 billion including talks | 19,59 |
| Model Company | OpenAI | Purchase commitments ~ $665.0 billion, zero debt (media reports); Altman cited commitments around $1.4 trillion (Nov 2025); planned compute spend through 2030 expanded from ~ $600 billion in Feb to ~ $750 billion in July | 60,61,62 |
Contract structures concentrate ultimate risk on frontier model developers. Anthropic generated $16.8 billion in revenue in Q3 2026, implying an annualized run-rate of $67.2 billion. OpenAI’s reported run-rate at the end of September was roughly $50.0 billion. Combined, their annualized run-rates reach roughly $120.0 billion (author’s calculation). Dividing Anthropic’s $600.0 billion in long-term cloud commitments by a 10-year term yields an average annual liability of $60.0 billion, or $86.0 billion over a 7-year term. This represents 0.9 to 1.3 times its $67.2 billion run-rate, and 1.5 to 2.2 times its estimated annual compute spending of roughly $40.0 billion (author’s calculation). European neocloud Nscale signed contracts with Microsoft and Anthropic totaling $103.4 billion in total contract value (TCV). However, its H1 2026 revenue was only $140.6 million, meaning TCV is roughly 368 times its annualized half-year revenue (author’s calculation). Nscale disclosed in its S-1 filing that it “has not secured legally binding commitments for the financing necessary to execute”63.
5.2 Asymmetric Contract Termination Rights Across the Supply Chain
Contract terms restrict model developers from terminating upstream hyperscaler contracts, while granting them rights to terminate contracts with downstream neoclouds.
Draft IPO filings reported in the media indicate that Anthropic’s cloud commitments with hyperscalers are non-cancellable19. In contrast, Nscale’s S-1 filing reveals that if data center developers fail to meet agreed delivery schedules, its customer, Anthropic, may “terminate the tranche without liability.” In H1 2026, 52% of Nscale’s total revenue was tied to this single customer63. Microsoft cancelled its contract for Nscale’s UK Loughton site (50 MW, 23,000 GPUs) on October 18, 2026, citing grid connection delays54. Microsoft also declined to exercise an approximate $12.0 billion option with CoreWeave in 2025, which OpenAI subsequently assumed under an $11.9 billion contract64,46.
Asymmetries appear in transactions between Nvidia and AI clouds as well. Regarding $36.0 billion in capacity Nvidia committed to purchase from AI clouds, its filings state that “AI cloud partners may unilaterally stop providing capacity to us and resell it to third parties under more favorable terms”6.
Hyperscalers have also attempted to transfer contract risks back to landlords. On September 24, 2026, Oracle issued a force majeure notice to infrastructure fund Blue Owl regarding Project Jupiter (a 2.45 GW Stargate campus in New Mexico). Citing gas pipeline completion delays pushed to February 1, 2027, Oracle requested a rent payment deferral of up to 3 years. This marked the first case of a hyperscaler attempting to pass external infrastructure delay risks onto facility owners and bondholders65.
If demand slows, default liabilities fall first on model developers bound by non-cancellable terms. Downstream neoclouds exposed to contract cancellations take the next hit. Hyperscalers bear liabilities as guarantors rather than direct tenants.
5.3 Hardware Obsolescence Risk and Tiered Returns
Profitability across the infrastructure value chain diverges based on the absorption of depreciation and hardware obsolescence risks.
| Tier | Company | Profitability | GPU Obsolescence Risk | Source |
|---|---|---|---|---|
| Powered Shell Landlord | Riot (Rockdale site, 191 MW, 20-year lease, $9.1B) | NOI of $365–$411 million per year against invested capital of $2.1–$2.3 billion ($11–$12 million/MW) → 15.9–19.5% return on cost (author’s calculation) | None absorbed | 66 |
| Powered Shell Landlord | Hut 8 (River Bend site, 245 MW, 15-year lease, $7.0B) | NOI of $6.9 billion (98.6% of rent); Needham estimates cost at $9–$11 million/MW → 16.8–20.6% return on cost (author’s calculation) | None absorbed | 57 |
| GPU Cloud | CoreWeave (Q2 2026) | Revenue $2.575 billion, depreciation $1.393 billion (54% of revenue), adjusted operating income $128 million (5% margin), GAAP operating loss $49 million, interest expense $640 million (TTM interest $1.87 billion vs operating loss $231 million); debt ~ $35.1 billion (total debt including leases ~ $51.6 billion); Q3 interest guidance $860–$940 million vs adjusted operating income guidance $200–$260 million (interest is 3.3–4.7x operating income) | Fully absorbed | 15,67 |
| GPU Rental (Hyperscaler) | Oracle Nvidia GPU Rental (quarter ended Aug 2025) | Revenue ~ $900 million, gross profit $125 million (14% margin); Blackwell rentals operate at a loss | Fully absorbed | 68 |
| Landlord Guarantor | Google’s Fluidstack lease guarantees | TeraWulf increased from $1.8B to $3.2B (~14% warrant acquired), Cipher $1.4B (5.4% warrant acquired) | Conditionally absorbed | 69,70 |
Powered shell landlords controlling grid interconnections (Riot, Hut 8) generate cash yields of 16% to 20% on invested cost (author’s calculation). They bear no chip depreciation or obsolescence risks, securing returns through 15- to 20-year leases and conditional corporate guarantees66,57. Google secured warrants representing an approximate 14% stake in TeraWulf and 5.4% in Cipher in exchange for guaranteeing Fluidstack’s leases69,70.
In contrast, neoclouds holding GPUs directly on their balance sheets absorb depreciation and interest burdens. CoreWeave’s Q2 2026 depreciation expense reached $1.393 billion, accounting for 54% of its $2.575 billion quarterly revenue15. Its adjusted operating margin stood at 5%, with a GAAP operating loss of $49 million. Trailing twelve-month interest expenses of $1.87 billion, factoring in $640 million for the quarter, exceeded its operating loss of $231 million over the same period15,67. With roughly $35.1 billion in debt (total debt including leases at roughly $51.6 billion), third-quarter interest guidance of $860–$940 million represents 3.3 to 4.7 times expected adjusted operating income of $200–$260 million15. CoreWeave calculates depreciation using server useful lives extended from 5 years to 6 years in 202346. Shortening useful lives to 3 years to account for technical obsolescence would double depreciation, widening its GAAP operating margin from -2% to roughly -56% (author’s calculation).
An analysis posted on X (2026-10-19) noted that out of eight neoclouds (CoreWeave, Nebius, Cipher, IREN, WYFI, Applied Digital, Hut 8, TeraWulf), only CoreWeave and Nebius reported positive EBITDA in recent quarters, with debt-to-equity ratios including leases standing at 10.3x for CoreWeave, 36x for TeraWulf, and 1.0x for Nebius169. Financing methods within the tier differ. Nebius raised convertible debt of roughly $4.3 billion in March 2026 (1.250% due 2031, 2.625% due 2033)167 and roughly $5.75 billion in August (0.50% due 2030, 4.50% due 2034)168, using conversion options to lower coupons, whereas CoreWeave relied primarily on standard notes carrying 9.000–9.750% coupons. This breakdown represents an individual analyst’s calculations, with only Nebius figures verified by corporate disclosures.
5.4 High-Cost Short-Term Leases and Model Developer Cash Burn
Model developers blend cancellable short-term leases to retain operational flexibility.
According to SpaceX’s amended S-1/A filing, Anthropic contracted for roughly 325,000 GPUs at the Colossus I and II sites, agreeing to pay $1.25 billion monthly through May 2029. After an initial 3-month term, either party can cancel the arrangement with 90 days’ written notice71. Investor presentations reported in the media indicate that Anthropic generated $16.8 billion in Q3 2026 revenue while capturing $1.70 in revenue per $1 of compute spent. This implies quarterly compute spending of roughly $9.9 billion (author’s calculation). The $3.75 billion quarterly lease paid to SpaceX ($1.25 billion over 3 months) represents roughly 38% of Anthropic’s quarterly compute expenditures and roughly 22% of its quarterly revenue (author’s calculation)72. This translates into an annualized spend of $15.0 billion, exceeding Anthropic’s reported annual inference budget of roughly $7.0 billion from January 202673.
Reports show Anthropic generated $16.3 billion in H1 2026 revenue against roughly $9.0 billion in compute and infrastructure outlays ($7.33 billion for full-year 2025). It reportedly posted an adjusted operating margin above 11% and positive free cash flow in Q319,74. Conversely, FT reports project OpenAI’s cumulative free cash flow deficit from 2026 through 2030 at -$278.0 billion, against a 2030 revenue target of $350.0 billion. Cash reserves of $122.0 billion raised in March 2026 are projected to be exhausted during 202875.
Model developers carry long-term non-cancellable liabilities upstream with hyperscalers, retain cancellation rights downstream with neoclouds on delivery delays, and rely on short-term infrastructure terminable within 90 days. If demand growth decelerates, developers can exit short-term contracts with neoclouds or SpaceX, leaving hundreds of billions of dollars in obligations to hyperscalers intact. Hyperscalers and chipmakers that extended guarantees would face simultaneous subrogation liabilities and equity valuation write-downs.
6. Real-Time Availability Setting Infrastructure Prices
6.1 Delivery Timeline Premiums on Operational Capacity
Price premiums for power and compute capacity in infrastructure transactions reflect operational delivery schedules rather than semiconductor silicon generations.
| Contract | Hardware / Delivery | Term & Cancellation | Per 1 MW / Year (author’s calc.) | Per GPU / Hour (author’s calc.) | Source |
|---|---|---|---|---|---|
| Anthropic at SpaceX Colossus | Unspecified generation (Colossus 1 Hopper-heavy), operational now | Through May 2029, 90-day cancellation | $34–$50 million (<$50M above 300 MW; ~$34M based on 325k GPUs and 733 GPUs/MW) | $5.27 (325k GPUs) to $7.78 (220k GPUs) | 71,76 |
| Google at SpaceX Colossus | ~110,000 Nvidia GPUs, Oct 2026–Jun 2029 | 90-day cancellation after Dec 31, 2026 | — | $11.46 | 77 |
| Anthropic at Nscale (Monarch 460 MW) / Lambda (Beacon Point 350 MW) / Volta (121 MW IT) | Vera Rubin, delivery 2027 | 6-year contract | $16.3 million / $16.7 million / $13.8 million | — | 78,79,80 |
| OpenAI at Oracle | ~4.5 GW, starting 2027 | 5-year contract, ~$300 billion | $13.3 million | — | 81 |
| Microsoft at IREN (200 MW) / Aker-Nscale (~100.6 MW) | GB300, signed 2025 | 5-year take-or-pay | $9.7 million / $12.3 million | ~ $2.91 / $2.72 | 82,58 |
| Anthropic on Google TPU | ~5 GW, starting 2027, ~$200 billion | 5-year contract | ~ $8.0 million | — | 83 |
| Anthropic on Amazon (Trainium) | Up to 5 GW, 10-year contract, >$100 billion | — | >$2.0 million (floor) | — | 7 |
Comparing annual costs per megawatt, active GPUs running today at SpaceX Colossus ($34–$50 million/MW) cost 2 to 3.6 times more than next-generation Vera Rubin chips scheduled for delivery in 2027 ($13.8–$16.7 million/MW), and 2.8 to 5 times more than 5-year GB300 contracts signed in 2025 ($9.7–$12.3 million/MW)71,76,78,79,80,82,58. Immediate availability commands higher rates than future silicon architectures. SpaceX (xAI), providing immediate capacity, saw adjusted EBITDA turn from -$609 million in Q1 2026 to +$1.1 billion in Q2, with its CFO noting an infrastructure payback period under one year84. Oracle renewed expiring AI contracts at 20% higher rates, maintaining a 97.9% utilization rate in Q1 FY202714. Market rental rates on 1-year contracts for H100 chips launched in 2022 rose from $1.70 in October 2025 to $2.60 in August 202685.
Amazon and Google provide capacity at lower prices by pairing sales with equity investments. Amazon secured over $100 billion in purchase commitments while investing an initial $5.0 billion and committing up to $20.0 billion (25% of contract value). Google matched an approximate $200 billion commitment by investing an initial $10.0 billion and committing up to $30.0 billion (20% of contract value) (author’s calculation)7,86. Filings do not explicitly define a legal tie between equity investments and discounted pricing.
6.2 Data Center Re-leasing and Price Stacking
Tracing how infrastructure is subleased clarifies where margins accumulate.
At the Beacon Point site in Texas, developer Hut 8 leased two 352 MW buildings (704 MW total) to Nvidia over 15 years for $19.6 billion, or roughly $1.86 million per MW per year. Specialized cloud provider Lambda then contracted to supply Anthropic with roughly 350 MW of GPU cloud capacity over 6 years for roughly $35.0 billion, pricing the capacity at roughly $16.7 million per MW per year. Across this subleasing chain, the price of the same megawatt increased roughly ninefold (author’s calculation)79. At the Tydal site in Norway, Bitdeer leased colocation space to Volta at $202 per kW per month ($2.42 million per MW per year), while Volta supplied Anthropic with 121 MW of IT capacity over 6 years for $10.0 billion ($13.8 million per MW per year), creating a 5.7x price differential (author’s calculation)80.
Breaking down the annual economics of operating a 1 MW installation yields the following components (author’s calculation). While final GPU cloud revenues generate $9.7–$16.7 million, annual shell facility rent consumes $1.86–$2.42 million. Electricity costs total $690,000–$990,000, assuming 1 MW IT capacity consumes 9.9 GWh annually at wholesale power rates of $70–$100 per MWh (Meta’s nuclear contract with Constellation was estimated at roughly $70 by Jefferies, with broader market estimates at $70–$95)87,88. Power accounts for 4% to 10% of final GPU revenues, while facility leases represent 11% to 25%.
Wholesale rents for powered shells rose from $1.80 per watt annually in summer 2025 to $2.30–$2.40 per watt in 2026 ($2.38 per watt in Morgan Stanley models, with 85% margins). This produces a net operating income of roughly $490,000 per MW, representing 3% to 5% of GPU cloud revenues (author’s calculation)89. Installing on-site gas generation to avoid grid connection delays adds a premium equal to 2% to 6% of gross data center revenue90. Of 66 GW of data centers under construction across North America, 95% is pre-leased prior to delivery. Rental rates rose roughly 70% since 2020, with annual escalators of 3% or higher standard in leases91. North American vacancy averages 1.4%, with Northern Virginia at 0.2%56. Power represents a small share of operating costs, but grid interconnections act as a bottleneck that channels margins to early-delivery providers.
Supply constraints underpin early-delivery capacity pricing. Janus Henderson projects that out of 157.4 GW in announced US data center capacity, 54% (roughly 84.7 GW) will be operational by 203092. Sightline data shows that of 16 GW planned for 2026 delivery, only about 5 GW broke ground as of March93. Morgan Stanley projected a 57 GW US power deficit over 2026–2028, estimating that even with on-site generation, supply will fall short by 32 GW, or 34% of demand94. GE Vernova’s gas turbine backlog and slot reservations reached 116 GW, exceeding its annual manufacturing capacity of 20 GW (targeting 30 GW by 2030)95,96. Lead times for high-voltage step-up transformers exceed 160 weeks97. LBNL data shows the median wait time from interconnection request to commercial operation for US power projects exceeds 5 years98.
6.3 China’s Hardware-Constrained Infrastructure Despite Power Surpluses
China illustrates that cheap, abundant electricity does not lower infrastructure price premiums when critical hardware components face supply restrictions.
According to China’s National Bureau of Statistics, domestic power generation from large-scale utilities reached 9,716 TWh in 2025, 2.2 times the US level of 4,425 TWh. Total generating capacity stood at 4,103 GW as of August 2026 compared to 1,319 GW in the US, with average thermal plant utilization hours falling 9.1% year-over-year in 202599,100,101,102. Industrial power rates in western and northern data center hubs run between 0.33 and 0.36 RMB per kWh ($49–$53 per MWh), roughly half the US industrial average of $91 per MWh in 2026103,88.
Yet operational data center capacity stands above 24 GW, less than half the US base of roughly 56 GW projected for year-end 2026104. US export controls created a compute capacity deficit. Zhipu AI curtailed service availability to 20% of previous levels and raised prices. Alibaba paused dedicated coding plan offerings, while ByteDance, Tencent, and Baidu instituted compute rationing systems105. Epoch estimates Huawei’s 2026 Ascend shipments at roughly 1.5 million units, representing under 4% of Nvidia’s aggregate compute capacity. Half of these chips used mobile LPDDR rather than HBM due to packaging bottlenecks. Pre-sanction stockpiles of roughly 13 million HBM stacks were exhausted during 2026106. The FT reported that domestic AI chip demand exceeded 5 million units in 2026 against a domestic supply of roughly 3.5 million units107. Nvidia’s compliance-tailored H200 shipments saw minimal volume under export licenses, accounting for less than 1% of its data center revenue6.
Despite energy availability, compute hardware constraints sustain premium infrastructure prices. Under these constraints, Chinese model developers pursue aggressive token pricing. Published pricing per million tokens across models like DeepSeek V4 Pro ($1.32 input / $3.96 output, with 50% night discounts), Qwen3.8-Max ($2 / $6), GLM-5.3 ($1.40 / $4.40), and Kimi K3 ($3 / $15) run at 10% to 75% of Opus 5.5’s standard output price of $20108,109,110,111. Daily token volume reached over 500 trillion in August 2026 reports, up from 140 trillion in March 2026 data from the National Data Administration112,113.
6.4 Fixed-Price HBM Contracts and Upstream Component Inventory
The high-bandwidth memory (HBM) sector upstream shows distinct pricing models and margin behavior.
Micron reported an 86.8% company-wide gross margin in Q4 FY2026 (ended 2026-09-03), while its Cloud Memory business, which houses HBM, generated $16.28 billion in revenue with an 83% gross margin114,115. However, starting in Q1 2026, standard server DDR5 generated higher profit per wafer than HBM. Because HBM contracts were locked into annual fixed-price agreements, they lagged price spikes in commodity server DRAM. TrendForce noted that while HBM consumed 22% of total DRAM wafer production in 2026, it accounted for only 9% of bit output116,117. Memory suppliers derived a share of profits in 2025–2026 from standard DRAM price increases driven by HBM wafer reallocation.
Price resets are scheduled for 2027. Blended HBM average selling prices (ASPs) are projected to rise 121% as HBM4 and HBM4e scale, with over 75% of 2027 supply already contracted118,55. Driven by an approximate 270% increase in 2026 server DRAM contract prices, the share of DRAM and NAND in cloud hyperscaler capex is projected to rise from 47% in 2026 to 68% in 2027119. HBM reached 63% of total component bill-of-materials costs in AI accelerators by Q4 2025120. This cost increases pressure on accelerator designers like Nvidia, whose long-term supply and memory commitments grew from $119.0 billion to $279.0 billion in one quarter6. Nvidia guided Q3 gross margins down slightly to 74% from 75.0% in Q2121. Its CFO stated on earnings calls that gross margins would trough around 71–72% in Q4 FY2027 before recovering to 72–73% in FY2028 through price adjustments122.
Total HBM sales to accelerator manufacturers across memory suppliers run at roughly $30.0 billion for Micron, $57.0 billion for Samsung, and $49.0 billion for SK hynix170,172. Combined, these sales total roughly $136.0 billion (author’s calculation), representing the HBM component cost embedded in 2027 AI hardware.
When data center delivery stalls due to power bottlenecks, accelerator memory accumulates as inventory. Nvidia halved CPU memory (SOCAMM) configurations on its Rubin platform and tested Rubin Ultra with 192 GB or 256 GB configurations instead of the planned 1 TB123,124. Morgan Stanley observed that prolonged grid delays would push inventory corrections onto memory and optical component suppliers ahead of Nvidia and Broadcom125. Micron expects DRAM industry bit growth to remain in the low 20% range through 2027–2028, projecting supply to remain tight55. Downstream, fixed-margin GPU operators face margin erosion. Oracle’s Nvidia GPU rental gross margins stood at 14% for the quarter ending August 2025, with newer Blackwell instances running at negative gross margins, showing the vulnerability of fixed-margin operators to rising component costs68.
With power and component bottlenecks keeping infrastructure prices elevated, end-user token demand and spending support the stability of the system.
7. Sustainability of End-User Demand and Financial Returns
7.1 Token Demand and Price Deflation
User token volumes are expanding while token prices fall.
| Metric | Value | Source |
|---|---|---|
| Google Customer API Tokens (per minute) | Rose from 7 billion in Q3 2025 to >10 billion in Q4, >16 billion in Q1 2026, and ~22 billion in Q2 2026; up 3.1x over three quarters, annualized ~4.6x; quarterly growth rates of 43%, 60%, and 37.5% | 126,127,128,129 |
| Google Total Tokens (including free tiers, monthly) | Rose 6.7x from 480 trillion in May 2025 to 3,200 trillion in May 2026; 22 billion/min equals ~950 trillion/month, implying paid API represents ~30% of total (author’s calculation) | 130,131 |
| Google Cloud Revenue | From >$20.0 billion (+63%) in Q1 to $24.8 billion (+82% YoY, +24% QoQ) in Q2 | 3,132 |
| Enterprise Effective Token Price (Ramp data, per 1M tokens) | Fell -41% from $1.15 in March 2026 to $0.68 in September; OpenAI at $0.48, Anthropic at $0.90; frontier model share contracted from 53% to 45% | 133,134 |
| Microsoft Operational Metrics | Azure FY2026 revenue exceeded $100.0 billion (+41%); Foundry customers consuming >1 trillion annualized tokens grew 4x YoY; reported demand exceeds available capacity | 12 |
Google’s paid customer API volume grew 3.1-fold across three quarters, from 7 billion tokens per minute in Q3 2025 to 22 billion in Q2 2026 (quarterly growth rates: 43% → 60% → 37.5%). Over the same six-month period, Ramp data showed the effective price paid by enterprise customers dropped 41%, from $1.15 to $0.68 per million tokens between March and September 2026133,134. Across the same enterprise cohort, average monthly AI spending per employee rose 11.7% month-over-month in September 2026, and the share of companies paying for AI reached 55.9% in September, up from 50.6% in April141. Epoch estimates that the price of tokens at a fixed capability level falls 47% per quarter.
In the Ramp dataset, token volume growth outpaced price declines, leading to net spending increases. Gartner terms this the “inference paradox.” While autonomous agent systems expand the number of inference tokens required per task through 2028, per-token prices fall roughly 95% by 2030. Combining these estimates indicates an approximate 100-fold increase in token consumption per business task (author’s calculation; forecast horizons differ)135. OpenRouter statistics show average prompt lengths increased from 1,500 to over 6,000 tokens, with reasoning models accounting for over 50% of volume136.
7.2 Performance-Adjusted Deflation and Downward Migration
The model ecosystem shows a divergence between price drops for fixed performance tiers and list price increases on flagship models.
Epoch reports that the cost to access a given benchmark intelligence level declines by 47% per quarter, an annual reduction factor of 13. Price premiums persist briefly after a model claims the top benchmark ranking137.
Simultaneously, listed rates for flagship tiers rose with new model generations. OpenAI’s list prices per million tokens (input / output) moved from $1.25 / $10 on GPT-5 (2025-08) to $5 / $30 on GPT-5.5 (2026-04), and $10 / $50 on GPT-6 Astra (2026-09)138. Anthropic shifted from $5 / $25 on Opus 4.5 (2025-11) to $10 / $50 on Fable and Mythos (2026-06 onward)139. Google maintained Gemini 4 Argon pricing at $2 / $10 through November 30, 2026, while announcing a price doubling to $4 / $20 effective December 1, alongside a 100% price increase on its Flash tier scheduled for January 1, 2027140. Intermediate updates within model families launched at lower prices (Opus 5.5 at $4 / $20, GPT-6.1 Sol at $2 / $10).
Enterprise buyers are migrating downward to lower-cost tiers. Ramp data shows frontier models’ share of enterprise token consumption dropped from 53% to 45%134. While 55.9% of Ramp’s corporate clients paid for AI in September, the top 1% of enterprises represented roughly 80% of total dollar spend across OpenAI and Anthropic. In September, 8% of enterprises switched primary AI vendors141. As outlined in Section 6.3, Chinese frontier models offer output pricing at 10% to 75% of Western equivalents, limiting developers’ pricing power.
7.3 Narrow End-Buyer Bases and Self-Referential Demand
Capital deployed in infrastructure relies on corporate budgets and a consumer base concentrated in narrow segments.
Surveys yield divergent assessments of enterprise ROI. McKinsey’s 2026 study of 1,719 companies found that 37% reported EBIT contributions from AI adoption, flat year-over-year, with top performers at 6%. Large enterprise deployment of autonomous agents increased from 27% to 40%142. A BCG survey of over 1,300 companies (2026-09-30) found that roughly 50% realized meaningful financial value, up from 5% in 2025, with AI budgets averaging 3.3% of revenue143. Conversely, Gartner’s October survey reported that only 13% of CIOs realized transformational value, and 44% of CEOs struggled to prove financial returns144,145. The US Census Bureau BTOS survey, cited by the Federal Reserve, showed 18% to 20% of all businesses and 37% of businesses with over 250 employees using AI (up from roughly 10% prior to an adjustment in questions in November 2025)146. An August 2025 MIT NANDA report claiming a 95% failure rate across 150 enterprise pilots drew criticism for narrow criteria and small sample size147.
Consumer tiers face paid conversion plateaus and downward pressure on average revenue per user. ChatGPT reported over 50 million paid subscribers out of 900 million weekly active users, an implied conversion rate of 5.6% (author’s calculation)148. Leaked internal projections reported by The Information indicate that out of 122 million projected consumer subscribers in 2026, roughly 112 million are expected to select an $8 per month ‘Go’ tier. Higher-priced $20 per month ‘Plus’ subscribers are projected to decline 80% to roughly 9 million149. Advertising revenues launched in February 2026 are projected by Truist to remain below $1.0 billion for the year150. Google does not disclose paid subscribers for Gemini, which reached 1.0 billion monthly users (up from 950 million)13. Microsoft 365 Copilot paid seats rose from 15 million in January to 30 million in July 202612.
Software spending remains concentrated within the top tier of customers and the AI industry itself. In Ramp data, average monthly spend per employee among the top 1% of enterprises fell 3.1% to $9,568 in September from its July peak of $9,873141. Industry estimates indicate that roughly half of frontier token revenue comes from internal AI development demand: AI research accounts for roughly 20%, software development and startups represent roughly 15%, and quantitative finance represents roughly 15%151. Over 80% of Anthropic’s model usage is concentrated in programming tasks136.
7.4 The Gap Between Capex and Disclosed AI Revenues
A gap persists between hyperscalers’ disclosed AI revenues and their capital expenditures.
Microsoft reported an annualized AI run-rate of $37.0 billion (+123% YoY) in March 2026, with no subsequent updates152. AWS announced during its Q2 2026 earnings call that its AI revenue run-rate surpassed $25.0 billion1. Combined, disclosed AI run-rate revenues for both companies total roughly $62.0 billion. Amazon reported that its silicon business, including proprietary accelerators, passed a $25.0 billion run-rate, without specifying cloud AI overlap1. Google does not break out AI-specific revenue. Goldman Sachs estimates full-year 2026 capex across the five hyperscalers at roughly $800.0 billion, projecting $1.2 trillion in 2027153. Meanwhile, Amazon’s trailing twelve-month property and equipment purchases reached $173.0 billion through Q2 20261, and Microsoft’s reported 2026 capex stood at roughly $175.0 billion40. These figures cover differing tracking windows (trailing 12 months versus calendar 2026). Comparing their combined $348.0 billion in capex against $62.0 billion in annualized AI run-rate revenues yields a capex-to-AI revenue ratio of roughly 5.6x (author’s calculation). Including Amazon’s $25.0 billion chip business ($87.0 billion total) yields a ratio of roughly 4.0x (author’s calculation). Adding model developers’ net annualized revenue of roughly $100.0 billion brings total ecosystem revenues to $160.0–$190.0 billion. This serves as an upper bound, given that model developer spending is booked inside hyperscaler AI revenue.
Morgan Stanley estimates combined capex for hyperscalers and neoclouds at roughly $980.0 billion in 2026, $1.58 trillion in 2027 (up ~60%), and $1.8 trillion in 2028171. Firm-level 2027 cash capex estimates include Google at $375.0 billion, Amazon (AWS only) at $270.0 billion, Meta at $221.0 billion, Microsoft at $209.0 billion, leases at $195.0 billion, SpaceX (including estimated Terafab capex) at $132.0 billion, Oracle at $97.0 billion, CoreWeave at $47.0 billion, and Nebius at $37.0 billion, with SpaceX, Oracle, CoreWeave, and Nebius holding large shares for the first time170,172. By procurement category, capital flows to semiconductors and servers ($672.0 billion), overseas supply chains ($511.0 billion to Korean memory, Taiwanese ODMs, etc.), US memory and storage ($217.0 billion), networking and optics ($215.0 billion), and facilities and power ($119.0 billion), with overlapping items summing to over $1.58 trillion170,172. Roughly 65% of 2027 spending (~$1.03 trillion) accrues to S&P 500 revenues, with Nvidia capturing roughly $690.0 billion (44% of total)170,171. Nvidia’s 2027 revenue carries a 91.7% exposure to AI infrastructure demand, the highest in the survey group171.
Warnings continue regarding revenue requirements for investment payback. Goldman Sachs calculates that sustaining 2026–2027 capex levels requires roughly $300.0 billion in annual pure AI revenue to break even over the long run16. Gartner projected on October 14, 2026, that cumulative revenues must reach $8.5 trillion by 2030 to justify spending, but forecast actual cumulative market revenues at $3.6 trillion (42% of required volume). It estimated that over $3.5 trillion in infrastructure assets face potential impairment, forecasting that over 80% of current neoclouds will fail to maintain independent operations through 2030154. Bain (2025-09) projected that reaching $2.0 trillion in annual revenues by 2030 would leave an approximate $800.0 billion revenue gap155. David Cahn of Sequoia estimated infrastructure revenue shortfalls at $200.0 billion in 2023, $600.0 billion in 2024, $840.0 billion in 2025, and $1.5 trillion in July 2026, implying an aggregate shortfall of roughly $3.0 trillion156.
Gross margins reported by model companies reflect real-time inference serving costs. SemiAnalysis data indicates that the share of compute allocated to inference across OpenAI and Anthropic has remained between 29% and 39% since Q1 2024, with the remainder dedicated to training new models157. Internal plans from OpenAI reported in The Information show training compute costs rising from $8.3 billion in 2025 to $32.0 billion in 2026 and $65.0 billion in 2027158. Anthropic’s reported 85% gross margin is an adjusted metric that excludes ongoing training expenditures72. Epoch calculated that GPT-5 generated roughly $6.0 billion in revenue against $4.0 billion in serving costs from August to December 2025 (a 30% serving gross margin). However, $5.0 billion in pure pre-training R&D over the preceding four months exceeded its $2.0 billion operational gross profit. Epoch characterized frontier model weights as “rapidly depreciating infrastructure”159. OpenAI’s leaked 2025 financials showed $13.07 billion in revenue and $7.50 billion in cost of sales (43% gross margin), but R&D spending of $19.18 billion represented 147% of revenue (author’s calculation)22.
This gap between revenue generation and capital spending establishes where losses transmit if shocks occur.
8. Three Scenarios Through 2030 and Loss Transmission Paths
The table below outlines transmission sequences and resilient layers across three potential paths for the infrastructure supply chain through 2030.
| Scenario | Conditions | First to Absorb Losses (Sequence) | Resilient Layers | Supporting Sections |
|---|---|---|---|---|
| ① Demand Stays Ahead | Token volume grows >4x annually; top model pricing holds (Gemini Dec increase takes hold) | No losses. 2027 HBM price increase (+121%) transfers into Nvidia costs (margins trough at 71–72%) → margin compression for Oracle/Microsoft neocloud partners on 5-year fixed GB300 contracts | Memory (contracted), shell landlords (16–20% return on cost), model companies (expanding margins), Tier-2 debt refinanced | 6.4, 7.1 |
| ② Power Delays | Only 54% of planned capacity online by 2030; transformers and turbines booked through 2031 | 1) Neoclouds with bare GPUs (accumulating unpowered GPU inventory; Google carries $122.8B in unplaced assets) 2) Memory (demand tied to server buildouts; Rubin memory downsizing) 3) Landlord creditors (Oracle Project Jupiter precedent: tenant demands rent deferral) | Operational data centers (retain delivery premium), Nvidia and Broadcom (Morgan Stanley analysis) | 6.1–6.4, 5.2 |
| ③ Demand / Pricing Disappoints | Quarterly 47% token price deflation outpaces volume, or shift to low-cost models (switching costs near zero; 8% of Ramp cohort switched vendors in Sep; Chinese models at 10–75% of Opus 5.5 prices) | 1) Neoclouds (model firms exercise delay termination clauses; interest exceeds operating income; CDS at 1,194 bp) 2) Model companies (~$1.2T in commitments vs ~$120B run-rate; Anthropic cloud contracts of $450B largely non-cancellable) 3) Guarantors (Nvidia $108.5B, Meta $68.4B, Alphabet $22.6B) 4) Hyperscaler EPS (equity gains reverse: 69% of Alphabet EPS, $53.4B Amazon other income) 5) Asset impairments (Gartner estimates $3.5T) | Landlords (protected by 15–20 year take-or-pay and Google guarantees) — assets transfer to guarantors if tenants default | 3, 4.4, 5 |
Mechanisms and figures for these scenarios are detailed below.
In Scenario ② (Power Delays), grid connection delays impact neoclouds holding chips they cannot energize. Alphabet reported $122.8 billion in construction-in-progress assets not yet operational as of June 30, 2026, up from $78.6 billion at year-end 20258. As data center delays push back chip deployments, secondary impacts hit memory suppliers. During the 2023 cyclical downturn, SK hynix posted a -24% annual operating margin, illustrating the volatility of upstream suppliers160. Concurrently, tenants requesting rent payment deferrals due to grid delays (as seen in Oracle’s Project Jupiter) transfer risk to project bondholders financing the builds65.
In Scenario ③ (Demand / Pricing Disappoints), losses hit neoclouds first. Model developers can cancel contracts by citing delivery delays. Neoclouds absorb utilization declines alongside unguaranteed debt loads and hardware obsolescence. Modeling CoreWeave with a 3-year depreciation schedule shifts GAAP operating margins to -56%46. Large non-cancellable liabilities then weigh on model developers. Anthropic faces annual cloud commitments of $60.0–$86.0 billion against an annualized revenue run-rate of $67.2 billion. If revenue declines 30% due to slowing token demand, annual payment obligations would reach 1.3 to 1.8 times revenue (author’s calculation). Losses then transfer to corporate guarantors backing infrastructure debt: Nvidia ($108.5 billion), Meta ($68.4 billion), and Alphabet ($22.6 billion)6,27,25.
Downward revisions in investee equity valuations would reduce hyperscalers’ reported net profits. As noted in Section 3.1, Amazon’s Q2 2026 pre-tax non-operating income of $53.4 billion was 1.9 times its consolidated operating income of $27.5 billion. If valuation write-downs of equal magnitude hit carrying values, Amazon’s quarterly pre-tax earnings would turn negative, assuming other operating segments remain unchanged (author’s calculation, hypothetical assumption).
In a downturn, losses hit neoclouds carrying unguaranteed debt and hardware obsolescence risks first, then move to model developers’ non-cancellable purchase obligations. Hyperscalers absorb impacts later through contingent guarantee payouts and equity mark-to-market losses.
9. Conclusion
As of autumn 2026, the AI infrastructure ecosystem distributes accounting profits, physical capital funding, and contractual risk across separate layers.
Reported earnings exceed core operating income, supported by valuation adjustments on equity stakes in model developers that are justified by multi-year compute commitments. Amazon’s non-operating income reached 3.2 times AWS operating income, while 69% of Alphabet’s EPS derived from valuation adjustments. Conversely, cash capital deployed into physical assets exceeds operating cash flow, relying instead on debt markets. The share of capex funded by incremental debt grew to 32%, and uncommenced off-balance-sheet leases across the five hyperscalers reached $662.0 billion. Meanwhile, take-or-pay volume risks reside with the two primary model developers at the base of the chain. These two firms carry roughly $1.2 trillion in commitments against combined annualized revenues of roughly $120.0 billion, with Anthropic’s $450.0 billion in cloud commitments reported as largely non-cancellable.
Credit structures diverge across debt tiers. Senior debt backed by investment-grade corporate guarantees prices at 5.75% to 6.6%. In contrast, second-tier debt backed solely by GPUs and enterprise contracts commands 8.5% to 12%. Widening CDS spreads and bond issuance hurdles in late 2026 concentrated within this second tier.
Current infrastructure market pricing centers on operational delivery timelines rather than silicon generations or power costs. Annual lease rates on operational, older-generation GPUs run 2 to 5 times higher than contracts for next-generation silicon scheduled for delivery in two to three years. Power and facility leases represent a limited portion of total operating costs. As demonstrated in China, where power generation is 2.2 times US levels and electricity costs are half, semiconductor export controls create compute shortages and premium infrastructure pricing similar to the US.
The capital loop holds as long as token consumption growth of 37.5% to 60% per quarter outpaces the quarterly 47% decline in token prices. If agentic systems fail to expand token volumes enough to offset deflation, or if software monetization bottlenecks persist, accumulated imbalances transfer losses through the contract chain: starting with neoclouds facing contract cancellations, moving through model developers, engaging corporate guarantors, and ultimately impacting hyperscaler earnings per share.
10. Signals to Monitor
Key metrics and thresholds for assessing balance and risk transfer across the infrastructure financing chain include:
| Signal | Current Level | Path Where Structure Holds | Path Where Cracks Appear |
|---|---|---|---|
| Google Customer API Token Volume (per minute) QoQ Growth | 37.5% (Q2 2026); Q3 earnings scheduled for release Oct 28 | Holds above 35% | Decelerates below 30% |
| Gemini 4 Argon Dec 1 Price Increase to $4/$20 | Announced | Successful implementation and peer price increases | Price increase cancelled or discounts offered |
| Ramp Effective Token Price / Frontier Model Share | $0.68 / 45% | Price rebounds / share stabilizes | Further price declines / share contracts |
| CoreWeave 5-year CDS / 2032 Notes Price | 1,194.5 bp / ~79 cents on the dollar | Stabilizes below 800 bp | Default event or credit covenant breach |
| Unguaranteed GPU-Backed Borrowing Costs (Volta and North Star subordinated) | 10–11.75% range | Falls below 9% | Repeated debt syndication failures |
| Nvidia Guarantee Exposure / AI Cloud Purchase Commitments | $108.5 billion / $36.0 billion | Exposure declines (OpenAI secures IG rating) | Commitment backlog increases |
| Oracle 5-year CDS / Customer Advances Share | 261 bp / 49% of operating cash flow | CDS contracts | Advances share and CDS widen further |
| Hyperscaler Uncommenced Lease Commitment Backlog | Meta $355.7 billion, Microsoft $329.1 billion | Growth rate slows | Continues growing >50% QoQ |
| Ratio of Operational GPU Rental Rates to 5-year Forward Contracts | 2–4x range | Multiple narrows (supply comes online) | Multiple persists or widens (worsening delays) |
| 2027 Blended HBM ASP Outlook | Projected +121% | Underperforms projection | Exceeds projection |
| Anthropic Revenue per $1 of Compute Spent | $1.70 (gross basis; ~$1.57 net of platform fees) | Metric rises | Metric declines |
| CoreWeave Covenant Cure Terms and Nov Earnings | Equity cure limited to 3 out of 4 consecutive months after Oct 28; Q3 results expected Nov | Meets covenants without cures | Triggers equity cures or covenant relief |
| CoreWeave Debt Principal Maturing through End of 2027 | ~ $10.6 billion | Refinanced below 10% coupon | Refinancing failure or asset sales |
11. Q3 2026 Earnings Update (2026-10-30)
This report was written before Amazon, Alphabet, and Microsoft released their third-quarter 2026 financial results on October 28–29, 2026. The existing sections are kept as written; this section records what the Q3 results show.
11.1 Equity Valuation Line
| Company | Prior basis (report body) | Q3 2026 | Source |
|---|---|---|---|
| Amazon | Q2 2026 other income of +$53.4 billion, mostly from Anthropic investment | “other income (expense), net” of −$703 million (Anthropic gains mentioned only for Q3 2025 at $9.5 billion); comprehensive income change in net unrealized gains on available-for-sale debt securities of +$35.974 billion net of tax of $12.233 billion (the 8-K does not say what drives it; 10-Q not yet filed as of 2026-10-30) | 173 |
| Alphabet | Q2 2026 other income of $98.0 billion ($6.26 of diluted EPS $9.11, or 69%); non-marketable equity securities fair-value remeasurement of $87.9 billion | Other income (expense) of −$8.221 billion; equity securities loss of $8.7 billion reduced diluted EPS by $0.55 (diluted EPS was $2.12; loss equals 26% of EPS), with unrealized losses in marketable equity securities “largely from SpaceX” ($75.2 billion carrying value in marketable securities, subject to short-term sale restrictions); non-marketable equity securities fair-value remeasurement of $8.8 billion (carrying value $128.2 billion; remaining funding commitment $20.0 billion unchanged) | 174,175 |
| Microsoft | FY2026 total net gain on OpenAI investments of +$5.0 billion | FY27 Q1 net gains from OpenAI investments in other income of $272 million (net income +$207 million, EPS +$0.03); commercial arrangements revenue with OpenAI of $10.0 billion, receivables from OpenAI of $7.2 billion, stake about 24% on an as-converted basis | 178 |
The elevated Q2 ratios discussed in Section 3.1—where Amazon’s valuation gain reached 3.2 times AWS operating income and Alphabet’s equity gains accounted for 69% of EPS—reflected a single quarter. In Q3 2026, the same accounting line flattened or swung negative across all three firms, demonstrating that reported earnings fluctuate with investee stake valuations in both directions.
Furthermore, the transmission path outlined in Section 8, where a decline in investee valuation cuts hyperscaler EPS, appeared at Alphabet during Q3. In this instance, the earnings drag originated through Alphabet’s listed stake in SpaceX rather than an AI model developer174,175.
11.2 Cash and Debt
Cash burn accelerated at Amazon and Alphabet during the third quarter, while debt levels expanded, relating directly to Section 4.1 and Section 4.3. Over the trailing twelve months to Q3 2026, Amazon generated operating cash flow of $168.3 billion (+29%) against purchases of property and equipment net of proceeds of $210.7 billion, taking free cash flow to −$42.4 billion from −$7.6 billion through Q2173. Amazon’s Q3 net cash capex was $75.9 billion; its full-year 2026 cash capex outlook of about $220 billion remained unchanged, but management stated on its earnings call that for 2027 it will “expect to spend meaningfully more”173,176. Amazon’s long-term debt reached $173.1 billion at September 30, 2026, an increase of $107.5 billion from $65.6 billion at December 31, 2025, and Q3 interest expense rose to $1.691 billion from $0.538 billion a year earlier173.
Alphabet recorded Q3 capex of $51.4 billion and operating cash flow of $42.3 billion, widening negative free cash flow to −$9.1 billion from −$5.9 billion in Q2174. Alphabet kept its 2026 capex guidance of $195–205 billion unchanged, while its CFO noted that 2027 capex will “increase significantly”, without providing a dollar figure174,177.
Microsoft maintained positive cash generation in FY27 Q1, with operating cash flow of $58.0 billion and cash paid for property and equipment of $39.4 billion, yielding free cash flow of $18.6 billion178. Capital expenditures including finance leases were $45.2 billion, lower than planned “as some third-party capacity deliveries we expected in Q1 were delayed”178,179. Microsoft projects capex above $50 billion next quarter and about $170 billion for calendar 2026179. Leases not yet commenced grew 10.8% sequentially to $364.6 billion at September 30, 2026 (from $329.1 billion at June 30), with lease start dates running from fiscal 2027 to fiscal 2034, while finance lease liabilities reached $71.0 billion178. As discussed in Section 6, Microsoft’s deferred capital outlay illustrates how the delivery timing of third-party capacity governs hyperscaler spending and near-term pricing.
11.3 Demand Indicators
Underlying demand continued to grow across all three cloud platforms, connecting to Section 3.2, Section 7.1, and Section 7.4. Google reported first-party model API volume of “nearly 32 billion tokens per minute — 40% more than last quarter”, up from approximately 22 billion in Q2129,180. However, Q2 disclosures referenced “model APIs” whereas Q3 specified “first-party model APIs”, meaning the two figures may not be measured on the same base. Google Cloud revenue reached $30.7 billion (+103%) with operating income of $11.1 billion (36.2% margin), while its cloud revenue backlog stood at $560.8 billion, including $16.8 billion from related party Crux AI174,175.
AWS revenue grew 39% year-over-year to $46.0 billion, delivering $17.8 billion in operating income (38.7% margin)173. AWS backlog rose to $578 billion from about $496 billion in Q2, which management described as “growing triple digits year-over-year”10,176. AWS disclosed no dollar AI run-rate but stated that AI revenue is expanding at triple digits, and Amazon announced a new AWS agreement with OpenAI173,176. Microsoft reported 48% growth in Azure and a 73% increase in commercial remaining performance obligation to $677 billion, stating that “Customer demand continues to exceed available supply”, while also providing no updated dollar AI run-rate179.
11.4 Signals Table (Section 10) After Q3
| Signal | At report date | After Q3 | Reading |
|---|---|---|---|
| Google API token growth | ~22 billion tokens per minute (Q2 2026) | 40% quarter-over-quarter growth to “nearly 32 billion tokens per minute” | Remains above the 35% “structure holds” line; subject to the caveat that Q2 reported “model APIs” while Q3 reported “first-party model APIs”180 |
| Hyperscaler leases not yet commenced | $329.1 billion at Microsoft (2026-06-30) | $364.6 billion at Microsoft (+10.8% in the quarter) | Below the “over 50% a quarter” crack line178 |
| Still pending | Pending Q3 filings | Amazon 10-Q (drivers of the $35.974 billion unrealized gain and Anthropic carrying values); CoreWeave Q3 results in November | Awaiting disclosure173 |
11.5 Overall
The report’s three core claims stand: (1) reported earnings depend on AI-related stake valuations while physical investment is increasingly financed with debt, as Q3 demonstrated the equity valuation line swinging negative alongside rising long-term debt; (2) contractual risk sits with model companies and supply timing sets prices, evidenced by expanding hyperscaler lease commitments and Microsoft’s capital outlays being dictated by delayed third-party deliveries; and (3) demand continues to rise, as evidenced by expanding token throughput, cloud revenues, and contract backlogs.
References
- Amazon, Second Quarter 2026 Financial Results (2026-07-30)
https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results - Amazon, Form 10-Q for the Quarterly Period Ended June 30, 2026
https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm - Alphabet, Q2 2026 Earnings Release (Form 8-K Ex.99.1) (2026-07-22)
https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm - Fortune, Alphabet Q2 earnings and Anthropic, SpaceX investment gains (2026-07-22)
https://fortune.com/2026/07/22/anthropic-spacex-investments-google-earnings-biggest-ever-profit-quarter/ - Microsoft, Form 10-K for the Fiscal Year Ended June 30, 2026 (2026-07)
https://www.sec.gov/Archives/edgar/data/0000789019/000119312526323660/msft-20260630.htm - Nvidia, Form 10-Q for the Quarterly Period Ended July 26, 2026
https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm - Anthropic, Expanding compute and investment partnership with Amazon (2026-04-20)
https://www.anthropic.com/news/anthropic-amazon-compute - Alphabet, Form 10-Q for the Quarterly Period Ended June 30, 2026 (2026-07-23)
https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm - Al Jazeera, OpenAI restructures into public benefit firm, Microsoft takes 27% stake (2025-10-28)
https://www.aljazeera.com/economy/2025/10/28/openai-restructures-into-public-benefit-firm-microsoft-takes-27-stake - Futurum Group, Amazon Q2 2026 Earnings Analysis (AWS RPO $496B) (2026-07)
https://futurumgroup.com/?p=93531 - CTech (Reuters), Anthropic draft IPO prospectus: revenue sources and cloud marketplace sales (2026-09-30)
https://www.calcalistech.com/ctechnews/article/r8sh14ux4 - Microsoft, FY2026 Q4 Earnings Call (2026-07)
https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 - Google, A message from Sundar Pichai on Alphabet Q2 2026 earnings (2026-07-22)
https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/ - Oracle, Q1 FY2027 Earnings Presentation (2026-09-10)
https://s23.q4cdn.com/440135859/files/content_files/Q1-FY27-Oracle-Earnings-Slides.pdf - CoreWeave, CoreWeave Reports Strong Second Quarter 2026 Results (2026-08-11)
https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx - KuCoin News (citing Goldman Sachs), $300B annual AI revenue needed to break even (2026-09-25)
https://www.kucoin.com/news/flash/goldman-sachs-ai-capital-spending-not-near-inflection-point-market-to-scrutinize-revenue-efficiency - AOL (Bloomberg), BofA: More than half of Oracle’s RPO is OpenAI (2026-06)
https://www.aol.com/finance/oracle-defused-key-risk-going-204912444.html - investinglive (Reuters), Anthropic draft IPO prospectus shows $42 billion 2025 net loss (2026-09-29)
https://investinglive.com/stock-market-update/anthropic-ipo-prospectus-shows-42-billion-2025-loss-and-518-billion-spending-plans/ - 24/7 Wall St. (WSJ), Anthropic has committed $450 billion to cloud computing, draft filing shows (2026-10-21)
https://247wallst.com/investing/2026/10/21/anthropic-has-committed-450-billion-to-cloud-computing-its-draft-ipo-filing-shows/ - Yahoo Finance, Leaked financials show OpenAI’s 2025 results (2026-06)
https://finance.yahoo.com/markets/stocks/articles/openai-2025-financials-leaked-38-121508294.html - Trades and Gains (Substack), Leaked OpenAI 2025 audited financials breakdown (2026-06)
https://tradesandgains.substack.com/p/openais-audited-numbers-leaked-2 - Techstrong.ai, OpenAI’s fast-growing revenues outpaced by massive scale of losses (2026-06)
https://techstrong.ai/articles/openais-fast-growing-revenues-outpaced-by-massive-scale-of-losses-ahead-of-expected-ipo/ - IMF, Global Financial Stability Report October 2026 (2026-10-13)
https://www.imf.org/en/Publications/GFSR/Issues/2026/10/13/global-financial-stability-report-october-2026 - BIS, Annual Economic Report 2026 Chapter 1 “Progress and peril” (2026-06-28)
https://www.bis.org/publications/aer-2026/progress-peril - Alphabet, Form 10-K for the Fiscal Year Ended December 31, 2025 (2026-02)
https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm - The Motley Fool, Alphabet will spend as much as $205 billion this year (2026-07-30)
https://fool.com/investing/2026/07/30/alphabet-will-spend-as-much-as-205-billion-this-ye - Meta, Form 10-Q for the Quarterly Period Ended September 30, 2026 (2026-10)
https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066773/meta-20260930.htm - Epoch AI, “Hyperscaler capex vs. cash flow” (2026)
https://epoch.ai/data-insights/hyperscaler-capex-vs-cash-flow - Computing, S&P downgrades Oracle to BBB- over AI expansion debt (2026-07)
https://www.computing.co.uk/news/2026/ai/oracle-under-pressure-after-ratings-downgrade-over-ai-expansion - FactSet Insight, Hyperscalers tap external financing as AI capex outruns cash flow (2026)
https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow - Aviva Investors, “Bond voyage”: Hyperscaler debt issuance trends (2026-07)
https://www.avivainvestors.com/en-be/views/aiq-investment-thinking/2026/07/bond-voyage/ - The Star (Bloomberg), JPMorgan sees tech bond sales surpassing half a trillion dollars in 2026 (2026-08-10)
https://www.thestar.com.my/business/business-news/2026/08/10/jpmorgan-sees-tech-bond-sales-surpassing-half-a-trillion-dollars-in-2026 - IFR, Amazon makes market history with largest corporate bond sale ever (2026-03)
https://www.ifre.com/bonds/2397565/amazon-makes-market-history-with-largest-corporate-bond-sale-ever - Bloomberg Government, Alphabet sells almost $32 billion in bonds (2026-02)
https://news.bgov.com/capital-markets/alphabet-sells-almost-32-billion-bonds-as-tech-races-to-fund-ai - IFR, Oracle returns with another jumbo bond to fund AI expansion (2026-02)
https://www.ifre.com/bonds/2379081/oracle-returns-with-another-jumbo-bond-to-fund-ai-expansion - Alphabet, Pricing Press Release on Capital Offering (2026-06-03)
https://s206.q4cdn.com/479360582/files/doc_news/2026/Jun/03/attachments/Alphabet-Pricing-Press-Release_2026-06-03.pdf - The Motley Fool, Nvidia will buy back $235 billion of its stock by 2028 (2026-10-08)
https://www.fool.com/investing/2026/10/08/nvidia-will-buy-back-235-billion-of-its-stock-by/ - Meta, Form 10-K for the Fiscal Year Ended December 31, 2025 (2026-01)
https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm - Fortune, Moody’s: Hyperscalers carry $662 billion in off-balance-sheet data center commitments (2026-02-25)
https://fortune.com/2026/02/25/hyperscaler-risk-off-balance-sheet-662-billion-data-center-commitments-meta-amazon-microsoft-oracle-alphabet/ - CFO Dive, Microsoft changes building useful lives while maintaining AI capex plans (2026-07)
https://www.cfodive.com/news/microsoft-holds-line-ai-spending-plans/826648/ - Dealroom, Anthropic chip lease SPV rated junk at BB+ by KBRA (2026)
https://dealroom.co/news/163182-anthropic-rated-junk-at-bb-as-it-chases-2t-ipo/ - The Synthesis, “The Chip Loan” — Anthropic Big Sky and North Star chip financing (2026-10-19)
https://thesynthesisai.substack.com/p/the-chip-loan - The Star (Reuters, citing FT), Anthropic offers Blackstone double-digit yield in chip financing deal (2026-10-17)
https://www.thestar.com.my/tech/tech-news/2026/10/17/anthropic-offers-blackstone-double-digit-yield-in-chip-financing-deal-ft-reports - IFR, The Meta-Blue Owl deal broken down: off-balance-sheet gymnastics (2025-10)
https://www.ifre.com/opinion/2363199/the-metablue-owl-deal-broken-down-off-balance-sheet-gymnastics-24-years-after-enron - Bisnow, Meta pushes its largest data center project off balance sheet with $27B joint venture (2025-10)
https://www.bisnow.com/news/national/data-center-capital-markets/meta-pushes-its-largest-data-center-project-off-its-books-with-27b-joint-venture-131490 - CoreWeave, Form 10-K for the Fiscal Year Ended December 31, 2025 (2026-02)
https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm - LetsDataScience News, CoreWeave shares and bonds slide as debt fears grip AI buildout (2026-10)
https://letsdatascience.com/news/coreweave-shares-and-bonds-slide-as-debt-fears-grip-ai-build-b082a38e - Dealroom, Volta lines up $5B debt deal as banks curb AI loan risk (2026)
https://dealroom.co/news/161809-volta-lines-up-5b-debt-deal-as-banks-curb-ai-loan-risk/ - ROIC.ai (citing FT), SpaceX seeks $40 billion debt financing for Nvidia AI chips (2026-10-07)
https://www.roic.ai/news/spacex-seeks-40-billion-debt-financing-for-nvidia-ai-chips-shares-slip-10-07-2026 - CoreWeave, Form 8-K: Nvidia unsold capacity purchase agreement of $6.3B (2025-09-09)
https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm - AI Weekly, Oracle, Nvidia, Alphabet, SpaceX CDS hit records on AI debt (2026-10)
https://aiweekly.co/alerts/oracle-nvidia-alphabet-spacex-cds-hit-records-on-ai-debt - The Edge Malaysia, Meta Hyperion bond prices drop (2026)
https://theedgemalaysia.com/node/819893 - Bloomberg Government, Firmus misfire casts shadow on data center IPOs in the pipeline (2026-10)
https://news.bgov.com/capital-markets/firmus-misfire-casts-shadow-on-data-center-ipos-in-the-pipeline - Computing, Microsoft exits Nscale deal, placing IPO at risk (2026-10)
https://computing.co.uk/news/2026/ai/microsoft-exits-nscale-deal-ipo-at-risk - MarketBeat, Micron Technology FY2026 Q4 earnings call highlights (2026-09-30)
https://www.marketbeat.com/instant-alerts/transcript-micron-technology-q4-earnings-call-highlights-2026-09-30/ - CBRE, North American data center demand continues to outpace supply despite record construction (2026-09-08)
https://www.cbre.com/press-releases/north-american-data-center-demand-continues-to-outpace-supply-despite-record-construction - Hut 8, Hut 8 signs 15-year 245 MW AI data center lease at River Bend campus ($7.0B TCV) (2025-12-17)
https://canada.hut8.com/resources/press-releases/hut-8-signs-15-year-245-mw-ai-data-center-lease-at-river-bend-campus-with-total-contract-value-of-usd7-0-billion - Aker ASA, Q3 2025 Presentation (Aker-Nscale to Microsoft GB300 deal) (2025-11-04)
https://mb.cision.com/Public/18835/4261024/994c2d210bb66616.pdf - Dealroom, Anthropic’s draft S-1 reveals $16.3B revenue, $80.7B loss, and up to $750B commitments (2026-10)
https://dealroom.co/news/164610-anthropics-draft-s-1-reveals-16-3b-revenue-80-7b-loss-and-up-to-750b-com - Augment Market, OpenAI reportedly carries zero debt and $665B in commitments (2026)
https://augment.market/pulse/openai-reportedly-0-debt-665b-in-commitments - TechCrunch, Sam Altman says OpenAI has $20B ARR and ~$1.4 trillion in data center commitments (2025-11-06)
https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/ - OODA Loop (citing The Information), OpenAI’s planned cloud spending hits $750 billion (2026-07)
https://oodaloop.com/briefs/technology/openais-planned-cloud-spending-hits-750-billion-as-computing-efforts-ramp-up/ - Nscale, Form S-1 Registration Statement (2026-09-18)
https://www.sec.gov/Archives/edgar/data/0002110365/000119312526395475/ck0002110365-20260918.htm - Sherwood News, Microsoft passes on $12 billion worth of CoreWeave data center capacity (2025-03)
https://sherwood.news/tech/microsoft-passes-on-usd12-billion-worth-of-coreweave-data-center-capacity - Dealroom, Oracle files force majeure on 2.45 GW New Mexico Stargate data center (2026-09)
https://dealroom.co/news/156472-oracle-files-force-majeure-on-2-45gw-new-mexico-stargate-data-center/ - Riot Platforms, Riot Platforms reports second quarter 2026 financial results (2026-08-10)
https://www.riotplatforms.com/riot-platforms-reports-second-quarter-2026-financial-results-and-strategic-highlights/ - TIKR, CoreWeave stock fell 25% on bond distress: Is it a buy now? (2026-10)
https://www.tikr.com/blog/coreweave-stock-fell-25-last-week-on-bond-distress-is-it-a-buy-now - Reuters (TradingView, citing The Information), Oracle falls on report of razor-thin GPU margins (2025-10-07)
https://kr.tradingview.com/news/reuters.com%2C2025%3Anewsml_L3N3VO0YP%3A0-oracle-falls-on-report-of-razor-thin-gross-profit-margins-in-cloud-business - Cointelegraph, Google largest shareholder in TeraWulf AI data center with Fluidstack (2025-08)
https://cointelegraph.com/news/google-largest-shareholder-terawulf-ai-data-center-fluidstack - Decrypt, Google dives deeper into bitcoin mining by backstopping Cipher-Fluidstack deal (2025-09)
https://decrypt.co/341339/google-dives-deeper-bitcoin-mining-backstopping-ai-hosting - SpaceX, Form S-1/A Registration Statement (2026-06-03)
https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm - Yahoo Finance, Anthropic’s $2 trillion IPO pitch shows gross margins and compute efficiency
https://finance.yahoo.com/technology/ai/articles/anthropic-2-trillion-ipo-pitch-045117254.html - Yahoo Finance (GuruFocus, citing The Information), Anthropic raises compute budget ceiling for 2026 (2026-01-29)
https://finance.yahoo.com/news/anthropic-quietly-raises-ceiling-121501064.html - AI Weekly (FT), Anthropic tells investors Q3 revenue hit $16.8B, adjusted operating margin >11% (2026-10-15)
https://aiweekly.co/alerts/anthropic-tells-investors-q3-revenue-hit-168b-up-45 - The Star (Reuters, citing FT), OpenAI forecasts cash burn near $280 billion by 2030 (2026-09-19)
https://www.thestar.com.my/tech/tech-news/2026/09/19/openai-forecasts-cash-burn-near-280-billion-by-2030-ft-reports - Anthropic, Higher limits with SpaceX Colossus compute (2026-05-06)
https://www.anthropic.com/news/higher-limits-spacex - Memphis Flyer, Google and SpaceX reach $30B lease deal for Colossus compute capacity (2026-06-08)
https://www.memphisflyer.com/google-spacex-reach-30b-rent-deal-for-colossus-compute-space/ - TechCrunch, Anthropic continues compute-gobbling streak in $45B deal with Nscale (2026-08-26)
https://techcrunch.com/2026/08/26/anthropic-continues-compute-gobbling-streak-in-45-billion-deal-with-nscale/ - TheEnergyMag, Anthropic, Lambda, Nvidia, and Hut 8 in Texas Beacon Point $35B deal (2026-09-01)
https://theenergymag.com/news/market-news/anthropic-lambda-nvidia-hut-ai-texas-beacon-point - Analytics Insight, Anthropic linked to $10B Norway deal for 121 MW compute (2026-08-05)
https://www.analyticsinsight.net/news/anthropic-linked-to-10b-norway-deal-for-121mw-nvidia-compute - The Decoder (citing WSJ), OpenAI signs $300 billion cloud contract with Oracle (2025-09)
https://the-decoder.com/openai-signs-300-billion-cloud-contract-with-oracle-according-to-wsj/ - IREN, IREN secures $9.7B AI cloud contract with Microsoft (2025-11-03)
https://iren.gcs-web.com/news-releases/news-release-details/iren-secures-97bn-ai-cloud-contract-microsoft - The Decoder, Anthropic commits $200 billion to Google Cloud over five years (2026)
https://the-decoder.com/anthropic-commits-200-billion-to-google-cloud-over-five-years/ - Yahoo Finance (PitchBook), xAI’s neocloud pivot eats into compute market (2026-10)
https://finance.yahoo.com/technology/ai/articles/xais-neocloud-pivot-eats-82-222456008.html - Silicon Data, H100 Rental Price Index
https://www.silicondata.com/products/silicon-index/h100 - 9to5Google, Google to invest up to $40 billion more in Anthropic (2026-04-24)
https://9to5google.com/2026/04/24/google-anthropic-investment-40-billion/ - Utility Dive, Meta-Constellation PPA could be first of many deals for existing reactors (2025-06)
https://www.utilitydive.com/news/meta-constellation-ppa-could-be-first-of-many-deals-for-existing-reactors/750567/ - US Energy Information Administration (EIA), Electric Power Monthly: Average Retail Price of Electricity to Ultimate Customers (2026-10-23)
https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_3 - Blockspace, Morgan Stanley: Powered shell deal wave (2026-10)
https://blockspace.media/insight/morgan-stanley-powered-shell-deal-wave-october/ - Modo Energy, US data center on-site gas and speed to power (2026)
https://modoenergy.com/research/us-data-center-onsite-gas-speed-to-power - Commercial Search, JLL Midyear Data Center Outlook: Demand hits record levels (2026-08-13)
https://www.commercialsearch.com/news/data-center-demand-hits-record-levels-this-year-jll/ - Janus Henderson, “Data center power: why the key risk is under-delivery, not overbuild” (2026-03-05)
https://www.janushenderson.com/en-gb/institutional/article/data-center-power-why-the-key-risk-is-under-delivery-not-overbuild/ - Utility Dive (Sightline Climate press release), Sightline analysis finds 30.5% of 2026 data center capacity under construction (2026-03-05)
https://www.utilitydive.com/press-release/20260305-sightline-climate-launches-powerstack-newsletter-with-analysis-finding-30-5-1 - NewKerala (Reuters), Morgan Stanley: AI race may depend more on access to power infrastructure (2026-10-17)
https://www.newkerala.com/news/a/ai-race-may-depend-more-access-power-infrastructure-978.htm - GE Vernova, Second Quarter 2026 Financial Results (2026-07-22)
https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000147/gevpressrelease2q26.htm - Utility Dive, GE Vernova gas turbine backlog hits 100 GW as prices rise (2026-04-23)
https://www.utilitydive.com/news/ge-vernova-gas-turbine-backlog-hits-100-gw-as-prices-rise/ - POWER Magazine, Transformers in 2026: Shortage, scramble, or self-inflicted crisis? (2026-01-02)
https://powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/ - Lawrence Berkeley National Laboratory, “Queued Up: 2026 Edition”
https://emp.lbl.gov/publications/queued-2026-edition-characteristics - Mysteel (China NBS), China’s 2025 power output rises 2.2% YoY (2026-01-19)
https://www.mysteel.net/news/5110599-nbs-chinas-25-power-output-rises-22-yoy-thermal-posts-first-fall-in-decade - US Energy Information Administration (EIA), Electric Power Monthly: Net Generation by Energy Source (2026-10-23)
https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_1_1 - GMW.cn (CCTV, China NEA), Power generation capacity reaches 4,103 GW by late August 2026 (2026-09-24)
https://m.gmw.cn/2026-09/24/content_1304567652.htm - Mysteel (China NEA), China’s thermal power capacity rises, utilization falls in 2025 (2026-01-29)
https://www.mysteel.net/news/5111881-nea-chinas-thermal-power-capacity-rises-utilization-falls-in-2025 - Caixin, In-depth: Chinese tech giants struggle to power AI data center boom (2025-05-16)
https://www.caixinglobal.com/2025-05-16/in-depth-chinese-tech-giants-struggle-to-power-ai-data-center-boom-102320408.html - SemiAnalysis, “The Chinese AI infrastructure boom” (2026-09-25)
https://newsletter.semianalysis.com/p/the-chinese-ai-infrastructure-boom - Caixin Global, Computing shortage forces Chinese AI firms to ration services (2026-04-21)
https://www.caixinglobal.com/2026-04-21/computing-shortage-forces-chinese-ai-firms-to-ration-services-102436452.html - Epoch AI, “Huawei’s roadmap to 2031” (2026-09-24)
https://epoch.ai/publications/huaweis-roadmap-to-2031 - AI Weekly (FT summary), Chinese tech firms faced AI chip supply shortfall in 2026 (2026-10-13)
https://aiweekly.co/alerts/ft-chinese-tech-firms-bought-1m-nvidia-rtx-pro-6000d-and-5000-chips-in-2026-for - DeepSeek, API Pricing (accessed 2026-10-26)
https://api-docs.deepseek.com/quick_start/pricing - Alibaba Cloud, Model Studio Pricing
https://www.alibabacloud.com/help/en/model-studio/model-pricing - Z.ai, API Pricing (accessed 2026-10-26)
https://docs.z.ai/guides/overview/pricing - eesel.ai, Kimi K3 Pricing (2026-07-17)
https://www.eesel.ai/blog/kimi-k3-pricing - China Daily, China’s daily token output surpasses 140 trillion (2026-03-24)
https://www.chinadaily.com.cn/a/202603/24/WS69c21887a310d6866eb3f91f.html - TechNode, China’s daily AI token usage tops 500 trillion (2026-08-28)
https://technode.com/2026/08/28/chinas-daily-ai-token-usage-tops-500-trillion-as-compute-demand-grows/ - Micron, Micron Technology reports results for the fourth quarter of FY2026 (2026-09-30)
https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm - The Next Platform, Micron’s fiscal Q4 2026 financial analysis (2026-10-02)
https://www.nextplatform.com/a/5300701 - TrendForce (DRAMeXchange), HBM profitability falls below DDR5 server memory (2026)
https://dramexchange.com/WeeklyResearch/Post/2/12718.html - TechNews (Taiwan), HBM wafer share vs bit output (2026)
https://technews.tw/?p=1565421 - TrendForce, 2027 HBM ASP projected to rise 121% (2026-09-29)
https://www.trendforce.com/presscenter/news/20260929-13255.html - TrendForce, Memory share in hyperscaler capex to expand (2026-08-25)
https://www.trendforce.com/presscenter/news/20260825-13198.html - Epoch AI, “AI chip component cost shares” (2026-05-21)
https://epoch.ai/data-insights/ai-chip-component-cost-shares - Nvidia, Nvidia reports financial results for second quarter fiscal 2027 (2026-08-26)
https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm - Webull (citing earnings call), Nvidia Q2 FY2027 earnings and margin outlook (2026-08-26)
https://www.webull.com/blog/304-Nvidia-Q2-FY2027-Earnings-Beats-Revenue-EPS-Estimates-Guides-Q3-to-108B - TrendForce, Nvidia Vera Rubin SOCAMM memory cut by half (2026-06-10)
https://www.trendforce.com/presscenter/news/20260610-13090.html - Yahoo Tech (citing The Information), Nvidia reportedly testing lower memory on Rubin Ultra (2026-08)
https://tech.yahoo.com/computing/articles/nvidia-reportedly-testing-lower-memory-164700228.html - The Star (Bloomberg, citing Morgan Stanley), Nvidia, Broadcom shielded as AI power crunch hits supply chain (2026-10-05)
https://www.thestar.com.my/tech/tech-news/2026/10/05/nvidia-broadcom-shielded-as-ai-power-crunch-hits-chip-supply-chain-says-morgan-stanley - Alphabet, Q3 2025 Earnings Call Transcript (2025-10-29)
https://s206.q4cdn.com/479360582/files/doc_events/2025/Oct/29/2025_Q3_Earnings_Transcript.pdf - Alphabet, Q4 2025 Earnings Call Transcript (2026-02-04)
https://s206.q4cdn.com/479360582/files/doc_events/2026/Feb/04/2025_Q4_Earnings_Transcript.pdf - Alphabet, Q1 2026 Earnings Call Transcript (2026-04-29)
https://s206.q4cdn.com/479360582/files/doc_events/2026/Apr/29/Alphabet-2026_Q1_Earnings_Transcript.pdf - Alphabet, Q2 2026 Earnings Call Transcript (2026-07-22)
https://s206.q4cdn.com/479360582/files/doc_events/2026/Jul/22/2026_Q2_Earnings_Transcript.pdf - Google, Google I/O 2025 Keynote (2025-05-20)
https://blog.google/technology/ai/io-2025-keynote/ - Google, Google I/O 2026: Sundar Pichai Keynote (2026-05-19)
https://blog.google/innovation-and-ai/sundar-pichai-io-2026/ - Pulse 2.0, Alphabet reports $109.9 billion in Q1 2026 revenue as Google Cloud accelerates (2026-04-29)
https://pulse2.com/alphabet-reports-109-9-billion-in-q1-2026-revenue-as-google-cloud-growth-accelerates-to-63/ - Ramp, Ramp AI Index September 2026 Edition (2026-09-09)
https://ramp.com/data/ai-index-sept-2026 - Fortune (citing Ramp), AI compute tokens are getting cheaper (2026-09-09)
https://fortune.com/2026/09/09/ai-compute-tokens-cheaper-boom/ - Techstrong.ai (citing Gartner), Agentic inference costs to rise fivefold even as tokens get cheaper (2026)
https://techstrong.ai/articles/agentic-inference-costs-to-rise-fivefold-even-as-tokens-get-cheaper-gartner-says/ - arXiv 2601.10088, Empirical Analysis of Token Consumption Dynamics on OpenRouter (2026-01)
https://arxiv.org/html/2601.10088v1 - Epoch AI, “The plunging price of thought” (2026-09-22)
https://epoch.ai/publications/the-plunging-price-of-thought - OpenAI, API Pricing Documentation
https://developers.openai.com/api/docs/pricing - Anthropic, Claude API Pricing (accessed 2026-10-25)
https://platform.claude.com/docs/en/about-claude/pricing - Google, Gemini API Pricing (accessed 2026-10-25)
https://ai.google.dev/gemini-api/docs/pricing - Ramp, Ramp AI Index October 2026 Edition (2026-10)
https://ramp.com/data/ai-index-oct-2026 - McKinsey, “The State of AI in 2026: On the road to ROI” (2026)
https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi.pdf - BCG, AI is starting to pay off: Half of companies generate value (2026-09-30)
https://www.bcg.com/press/30september2026-ai-starting-to-pay-off-companies-generate-value - Gartner, Gartner survey finds only 13% of CIOs have realized significant value from AI (2026-10-19)
https://gartner.com/en/newsroom/press-releases/2026-10-19-gartner-survey-finds-only-13-percent-of-cios-say-their-enterprise-has-realized-significant-or-transformational-value-from-ai - Gartner, Gartner says 44% of CEOs struggle to demonstrate AI financial returns (2026-10-20)
https://gartner.com/en/newsroom/press-releases/2026-10-20-gartner-says-44-percent-of-ceos-struggle-to-demonstrate-ai-financial-returns-as-technology-becomes-central-to-growth - Federal Reserve FEDS Notes, Monitoring AI adoption in the US economy (2026-04-03)
https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html - Fortune, MIT report: 95% of generative AI pilots at companies are failing (2025-08-18)
https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo - TechCrunch, ChatGPT reaches 900M weekly active users (2026-02-27)
https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users - Shopifreaks (citing The Information), OpenAI forecasts $8 ChatGPT Go tier will drive subscriber growth (2026)
https://www.shopifreaks.com/openai-forecasts-8-chatgpt-go-tier-will-drive-consumer-subscribers-to-122m-this-year-as-plus-subscribers-fall-80/ - DesignRush News, ChatGPT introduces ads to free users as OpenAI seeks profitability (2026-02)
https://news.designrush.com/chatgpt-introduces-ads-free-users-openai-profitability - AI Is Real (Substack), “AI Economics: A Synthesis, Q2 2026” (2026)
https://aiisreal.substack.com/p/ai-economics-a-synthesis-q2-2026 - Microsoft, FY2026 Q3 Earnings Call (2026-04)
https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q3/press-release-webcast - ROIC.ai (citing Goldman Sachs), Goldman Sachs sees AI capex at $1.2 trillion by 2027 (2026-09-25)
https://www.roic.ai/news/goldman-sachs-sees-ai-capex-at-12-trillion-by-2027-fueling-debt-financed-buildout-09-25-2026 - Gartner, Gartner says over $3.5 trillion of AI infrastructure spending exposed to impairment by 2030 (2026-10-14)
https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030 - Bain & Company, $2 trillion in new revenue needed to fund AI’s scaling trend: 6th Annual Tech Report (2025-09)
https://www.bain.com/about/media-center/press-releases/20252/$2-trillion-in-new-revenue-needed-to-fund-ais-scaling-trend---bain—companys-6th-annual-global-technology-report/ - David Cahn (Sequoia), “AI’s $1.5T Question” (2026-07)
https://dcahn.substack.com/p/ais-15t-question - OfficeChai (citing SemiAnalysis), Inference share steady as pre-training shifts to post-training (2026-10-18)
https://officechai.com/ai/while-share-of-inference-has-held-steady-over-the-last-2-years-pre-training-compute-has-moved-to-post-training-semi-analysis/ - Quasa, OpenAI leaked financials: Bullish revenue growth amid soaring compute costs (2026)
https://quasa.io/media/openai-s-leaked-financials-bullish-revenue-growth-amid-soaring-ai-compute-costs - Epoch AI, “Can AI companies become profitable?” (2026)
https://epoch.ai/gradient-updates/can-ai-companies-become-profitable - StorageNewsletter, SK hynix Fiscal 4Q23 and FY2023 financial results (2024-01-29)
https://www.storagenewsletter.com/2024/01/29/sk-hynix-fiscal-4q23-financial-results/ - CoreWeave, Form 8-K: DDTL 3.0 Credit Agreement Amendment (2026-01-02)
https://www.sec.gov/Archives/edgar/data/1769628/000176962826000003/crwv-20251231.htm - CoreWeave, Form 8-K: Offering of $4.2 Billion 2.875% Convertible Senior Notes (2026-09-22)
https://www.sec.gov/Archives/edgar/data/1769628/000176962826000432/crwv-20260917.htm - CoreWeave, Form 10-Q for the Quarterly Period Ended June 30, 2026 (2026-08-12)
https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm - 24/7 Wall St, Musk predicts guppy-level death for neoclouds as CoreWeave stock falls 38% in October (2026-10-26)
https://247wallst.com/investing/2026/10/26/musk-predicts-guppy-level-death-for-neocloud-companies-as-coreweave-stock-falls-38-in-october - Federal Reserve, Open Market Operations: Federal Funds Target Range History
https://www.federalreserve.gov/monetarypolicy/openmarket.htm - FRED (Federal Reserve Bank of St. Louis), Secured Overnight Financing Rate (SOFR)
https://fred.stlouisfed.org/series/SOFR - Nebius Group, Nebius Group announces closing of private offering of convertible senior notes ($4.3B) (2026-03-20)
https://nebius.com/newsroom/nebius-group-announces-closing-of-private-offering-of-convertible-senior-notes-with-aggregate-gross-proceeds-of-approximately-4-3-billion - Nebius Group, Form 6-K: Closing of $5.75 Billion Convertible Senior Notes Offering (2026-08-24)
https://www.sec.gov/Archives/edgar/data/1513845/000110465926100347/tm2623863d1_ex99-1.htm - Sergey (@SergeyCYW), Post on X: Debt and leverage comparison across 8 neoclouds (2026-10-19)
https://x.com/SergeyCYW/status/2112169649936945490 - moomoo, Morgan Stanley sizes up $1.58 trillion in 2027 AI capex destinations (2026-10-16)
https://www.moomoo.com/community/feed/morgan-stanley-sizes-up-1-58-trillion-in-2027-ai-117450233544710 - Beinsure, Morgan Stanley forecasts $1.58T AI boom as Nvidia exposure hits 91.7% (2026-10-20)
https://beinsure.com/news/morgan-stanley-forecasts-1-58-tn-ai-boom-as-nvidia-exposure-hits/ - P Equity Research (@pequityresearch), Post on X: Morgan Stanley 2027E capex to supplier revenue mapping chart (2026-10-17)
https://x.com/pequityresearch/status/2111503254375031148 - Amazon, Q3 2026 Earnings Release (8-K Ex. 99.1) (2026-10-29)
https://www.sec.gov/Archives/edgar/data/1018724/000101872426000038/amzn-20260930xex991.htm - Alphabet, Q3 2026 Earnings Release (8-K Ex. 99.1) (2026-10-28)
https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000080/googexhibit991q32026.htm - Alphabet, Form 10-Q (quarter ended 2026-09-30) (2026-10-29)
https://www.sec.gov/Archives/edgar/data/1652044/000165204426000084/goog-20260930.htm - Investing.com, Amazon Q3 2026 Earnings Call Transcript (2026-10-29)
https://www.investing.com/news/transcripts/earnings-call-transcript-amazon-q3-2026-misses-eps-as-shares-fall-after-hours-93CH-4985489 - Investing.com, Alphabet Q3 2026 Earnings Call Transcript (2026-10-28)
https://www.investing.com/news/transcripts/earnings-call-transcript-alphabet-q3-2026-profit-miss-sparks-10-afterhours-drop-93CH-4981037 - Microsoft, Form 10-Q (quarter ended 2026-09-30) (2026-10-28)
https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm - Microsoft, FY2027 Q1 Earnings Release and Conference Call (2026-10-28)
https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1 - Google, Sundar Pichai’s Q3 2026 Earnings Message (2026-10-28)
https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/