Amazon: profitable cloud, increasingly externally funded expansion

Last updated: 2026-10-21 · Author: Codex / OpenAI · Status: reviewed · Reviewers: Codex (gpt-6-astra) cross-check applied 2026-10-17; Claude (Opus 5.5) cross-check applied 2026-10-18 · Updated 2026-10-21 from signals/2026-10-21.md

Summary

• In Q2 2026, AWS earned $16.621bn operating income on $42.232bn revenue, up 37% year over year; its reported margin was 39.4%. Subtracting the filing's $551m energy-contract gain, primarily affecting AWS, gives approximately 38.05% (B). Neither measure establishes returns on the newest AI infrastructure. Results; Q2 filing.
• Amazon's consolidated build is no longer comfortably covered by operating cash flow. Trailing free cash flow fell from $38.219bn at end-2024 to $11.194bn at end-2025 and −$7.604bn in the twelve months to June 2026; including lease and financing-obligation principal gives −$9.511bn. Trailing acquisition, non-marketable-investment and other net outflows were a further $41.956bn, including $28.7bn to OpenAI and $10.0bn to Anthropic in H1 2026. Results; Q2 filing.
• On July 30, 2026, management raised its 2026 capex estimate from approximately $200bn to $220bn in cash capex, citing memory costs. Amazon received $66.998bn of long-term-debt proceeds in H1 2026, then issued $25bn in July and £4.25bn in September; lab equity investments consumed $38.7bn in H1 and another $21.3bn after June, per the July 31 filing. Call; Q2 filing; July; September.
• Anthropic exposure is both commercial and financial: $18bn invested through June 2026 and $15bn of remaining financing-facility capacity, with availability tied to compute-delivery milestones. Its June carrying value was $190.4bn, approximately 34.5% of Amazon's book equity; carrying value plus remaining funding capacity implies approximately $205.4bn of maximum investment loss exposure under the filing's definition (B). Its April 20, 2026 AWS commitment exceeds $100bn over ten years. Amazon's exact ownership and recognized revenue exposure to Anthropic remain undisclosed. Q2 filing; Anthropic.
• OpenAI's November 2025 $38bn AWS agreement expanded by $100bn in February 2026. The July 31, 2026 filing reports $28.7bn invested in OpenAI through June and another $21.3bn subsequently, completing the $50bn investment. In this note's judgment, diversification across labs leaves correlated financing risk. Original deal; expansion; Q2 filing.
• Evidence of current demand includes 37% AWS revenue growth in Q2 2026, $496bn of predominantly AWS commitments at June 30 with a 6.4-year weighted remaining life, and Anthropic's April 20, 2026 disclosure that it used more than 1m Trainium2 chips. These demonstrate sales, contracts and deployment, but do not establish utilization or cash returns. Backlog rose $252bn in H1 2026, when OpenAI added $100bn and Anthropic more than $100bn of AWS commitments: most of the increase plausibly comes from two labs Amazon also funds (inference, C; customer shares undisclosed). On July 30, management said servers and networking break even in under three years on average and most AI capacity has contracts of at least five years. Whether those cash paybacks materialize is the central uncertainty; Q3–Q4 margins and cash conversion against H1's 38.8% AWS net PP&E growth will provide early evidence. Results; Q1 filing; Q2 filing; Anthropic; call.

Key claims

Amounts are US dollars unless stated otherwise. Sources were accessed 2026-10-17, with sources for the Claude-review additions re-fetched 2026-10-18 as recorded below. A means directly disclosed, including attributed management claims; B means arithmetic or corroborated secondary evidence; C means reported private-company information or an estimate. Analytical judgments are identified as such. Management speech in the externally hosted transcript is treated as primary testimony; its transcription is not independently authenticated.

Analysis

Profitability and demand. AWS sells infrastructure, managed services and model access through Bedrock; Trainium adds proprietary accelerator economics. Its accelerating growth and high operating margin establish a substantial operating business. They do not separately disclose AI margins, maintenance capex or returns by infrastructure vintage. The AI and chips run rates overlap: adding them would double-count some activity. Neither is audited standalone segment revenue. Results; official excerpt.

The Q2 2026 filing quantifies $551m of net unrealized energy-contract gains in technology and infrastructure expense, primarily affecting AWS. Subtracting the entire gain from AWS operating income gives (16.621 − 0.551)/42.232 = approximately 38.05% margin (B); the exact segment allocation is not separately disclosed. This agrees with approximately 38.1% inferred from management's rounded 650 versus 520 basis points of margin expansion. Profitability remains substantial; part of the reported improvement comes from a valuation effect. Derivatives note; July call.

Backlog has increased alongside contract duration. That improves visibility but pushes some realization further into the future. It covers qualifying long-term service commitments, primarily AWS, rather than cash already collected. Customer usage, delivery and chip-performance obligations affect recognition. The appropriate test is conversion into profitable usage and cash, not the headline backlog alone. Filings.

The customer mix also matters. Backlog rose $120bn in Q1 2026, when OpenAI expanded its agreement by $100bn, and $132bn in Q2, when Anthropic added more than $100bn of commitments. The filings discuss both agreements in the backlog disclosure. If fully included and not yet consumed, those announcements would account for more than $200bn, or 79%, of the $252bn H1 increase; including OpenAI's original $38bn agreement would put the combined announced totals at roughly half of June's $496bn. These are conditional comparisons, not disclosed customer shares: recognition, inclusion and other contracts' additions or consumption prevent a customer-level reconciliation. Most of the growth plausibly comes from the two investee labs (inference, C). Amazon paid them $38.7bn in H1 equity funding and $21.3bn more after June, completing $60bn of 2026 lab funding by the July 31 filing. Sales, financing exposure and the apparent demand pipeline therefore depend partly on the same counterparties. Annual backlog; Q1; Q2; OpenAI expansion; Anthropic commitment.

Management's dated evidence is unusually strong: the April call said Trainium2 was largely sold out and Trainium3 nearly subscribed; the July call still described shortages despite higher planned spending. It also acknowledged that building sites before filling them with revenue-producing servers depresses near-term FCF. These statements support a timing mismatch explanation, but remain management's forecast of future utilization. April excerpt; July transcript.

The same July 30 answer supplies the strongest explicit case for returns: management says servers and networking break even in a little under three years on average, servers last at least five to six years, and most AI capacity is contracted for at least five years. It says server purchases follow visible demand and AI margins and returns are tracking slightly ahead of core cloud at the same stage. These are A-grade attributed statements, not independently demonstrated cash paybacks for new AI equipment. The five-to-six-year life claim must be read alongside the 2025 shortening of a server subset from six to five years. Q3–Q4 margins and cash conversion can provide early evidence; segment margins alone cannot establish a multi-year payback. July call; useful-life change.

Cash flow and the capital bill. The quarterly CSV distinguishes gross cash purchases, sale proceeds and incentives, net cash capex, noncash finance-lease additions, and principal repayments. It covers eight calendar quarters. Fourth quarters are annual less nine-month values; first quarters use half-year less second-quarter values. This is consolidated Amazon, not an AI-only spending series. Retail logistics and other businesses also absorb capital.

AWS accounted for $90.120bn of $118.648bn, or 76.0%, of H1 2026 consolidated net property-and-equipment additions. This measure includes noncash investment and does not isolate AI spending or AWS cash capex. It nevertheless shows that expansion is predominantly in AWS, narrowing the uncertainty left by the consolidated cash series. Segment information.

AWS's Q2 2026 net PP&E additions were $48.604bn, up from $16.043bn a year earlier and above its $42.232bn quarterly revenue. Annual AWS additions had already risen from $53.267bn in 2024 to $96.496bn in 2025. Net PP&E grew 38.8% in H1 2026, from $190.055bn to $263.750bn. In the other direction, annualizing H1 operating income of $30.782bn gives $61.564bn: 27.1% of average opening/closing net PP&E or 23.3% of June net PP&E (B). These pre-tax operating-income/net-PP&E proxies show substantial profitability against the deployed asset base, but omit other invested capital, include construction in progress and annualize a growing business; they are neither full ROIC nor returns on the newest AI cohort. Annual segment data; June segment data.

Amazon defines FCF as operating cash flow less cash property purchases net of sales and incentives. Its historically published second definition then deducts principal on both finance leases and financing obligations. The current releases emphasize the first measure; the lease-adjusted values here reconstruct the older definition rather than pretend Amazon still headlines it. A third historical definition substitutes equipment acquired under finance leases and deducts principal on other leases and financing obligations. It cannot be reconstructed reliably from aggregate finance-lease additions alone, so no synthetic value is presented. Definition source.

Do not subtract both all new leased assets and their principal repayments from the same cash measure. Lease additions describe investment financed without current purchase cash; repayments describe servicing previous obligations. Operating-lease payments and interest already pass through operating cash flow.

The trajectory is clear even after allowing for retail seasonality: trailing OCF rises while FCF falls through zero. The first-half 2026 FCF deficit excludes lab-investment outflows. Consequently, saying that AWS is profitable is insufficient to claim that Amazon's full build is self-financing. Conversely, negative FCF during growth does not prove negative project returns. That depends on eventual margins, utilization and replacement requirements, which are not disclosed separately.

The excluded acquisition, non-marketable-investment and other net outflows were $41.956bn in the twelve months to June 2026, including $39.767bn in H1; $38.7bn of H1 outflows funded OpenAI and Anthropic. Those cash uses sit alongside, rather than inside, the capex-based FCF deficit. Cash flows and liquidity.

An illustrative scenario makes the full-year direction explicit (C, not a forecast). Against the July $220bn cash-capex guide, H1 purchases of $98.411bn gross or $96.310bn net imply H2 spending of approximately $122–124bn, or $61–62bn per quarter, depending on the guide's convention; Q2 gross purchases were $54.208bn. H1 OCF grew 44.2% from $49.530bn to $71.419bn. Repeating that growth on H2 2025's $89.984bn gives approximately $201.2bn FY2026 OCF and −$18.8bn FCF if $220bn means net cash capex, before lab investments. If the guide is gross purchases, sale proceeds and incentives must be added back to that FCF illustration. This is a conditional funding illustration, not a management FCF forecast. Annual OCF; H1 cash flows; guidance.

Funding and loss allocation. The bond sequence confirms a shift toward external funding. After July's $25bn bond issue, Amazon issued £4.25bn of sterling notes on September 14, 2026, with approximately £4.235bn proceeds before offering expenses, coupons of 5.20–6.65%, and maturities from 2029 to 2045. Long maturities reduce immediate refinancing pressure; these issuances demonstrate continuing funding access but do not supply a complete October debt balance. The substantial June liquid-asset balance provides a cushion. However, equity investments in private labs are not substitutes for liquid cash, and uncommenced leases represent additional future fixed commitments. July filing; September filing; liquidity disclosure.

Total contractual commitments rose 47.8% from $439.661bn in December 2025 to $650.034bn in June 2026. Unconditional purchase obligations rose from $84.772bn to $130.065bn, uncommenced leases from $96.373bn to $137.214bn, and long-term-debt principal and interest from $108.202bn to $220.309bn. These are multi-year commitments, not an immediate cash bill or interchangeable debt balances. Purchase obligations include energy, property and equipment, content and software; their minimum payments can persist if AI demand weakens. Annual commitments; June commitments.

June's $123.0bn liquidity was unchanged from December and is a dated balance, not evidence that it subsequently eroded. The disclosed post-June flows give a partial bridge: July note proceeds of $24.867bn less the $21.3bn OpenAI payment leave $3.567bn, plus approximately £4.235bn of September note proceeds. Borrowing thus more than offsets that investment payment in this limited comparison; Q3 operating flows, capex, other investments and FX remain unreported here, so this cannot establish an October cash balance. The $30bn commercial-paper programme and $15bn plus $5bn revolving facilities were undrawn at June 30; the latter $5bn facility has a 364-day term. Liquidity and credit facilities; July; September.

The June filing also disclosed an undrawn $17.5bn delayed-draw term loan whose unused commitment automatically terminated on September 30, 2026. Amazon's EDGAR submissions from September 15 through October 18 include no 8-K announcing a draw. Expiry undrawn is likely (inference, C), but not confirmed: the June agreement disclosure may already have met disclosure requirements. The Q3 10-Q is the next check for whether it became debt; the facility cannot be counted as October available liquidity. Debt note; EDGAR submissions.

Amazon also agreed on April 13, 2026 to acquire Globalstar for up to $90 a share in cash or stock, subject to adjustments and stock-price terms, with cash capped at 40% of shares. The 128,598,125 shares outstanding at signing imply approximately $11.6bn at the price ceiling and up to $4.6bn cash on that share base (B), rather than fixed closing consideration. The registration statement became effective August 18. This is a non-AI claim on the same funding pool. Prospectus; Effectiveness notice.

Equipment already acquired but not yet paid adds supplier financing beyond bonds and leases. The disclosed increase is not the ending payable balance and must not be subtracted from cash FCF as though already paid. Consolidated funding is fungible: the filings do not allocate particular bond proceeds between AI infrastructure, lab investments and other uses. Supplemental cash-flow disclosure.

Amazon shareholders initially absorb lower distributions, infrastructure impairments and investment losses. Corporate creditors have claims on the diversified enterprise, not merely successful AI projects. Construction, power and lease obligations can persist when customer demand weakens. The risk is therefore a lower return on an increasingly committed capital base before it becomes a solvency question.

Anthropic: customer, asset and financing counterparty. The original investment is no longer the current cash total. The later Series H subscription reduced remaining financing-facility capacity; adding the full facility to funded capital would overstate cash already paid. Availability is linked to AWS compute-delivery milestones, so remaining capacity is not necessarily immediately drawable. This explicitly connects financing and infrastructure delivery, although it does not show that every AWS invoice is paid with Amazon-originated funds. Investment note.

Scale matters: Anthropic's carrying value rose from $60.6bn at December 2025 to $190.4bn in June 2026, a $129.8bn increase including $10.0bn of new cash. The approximately $119.8bn residual indicates the scale of noncash appreciation across earnings and OCI, rather than a complete mark-by-mark roll-forward. The June asset was 34.5% of Amazon's $551.620bn book equity. Applying the filing's maximum-investment-loss definition of carrying value plus future funding commitments gives approximately $205.4bn including the $15bn remaining facility; that is accounting and funding exposure, not cumulative cash invested. Investment note and balance sheet.

Consolidated Q2 operating income was $27.461bn, against $53.396bn of non-operating income and $80.857bn pre-tax income: 66.0% was non-operating (H1: $69.378bn of $120.691bn, or 57.5%). Q2 net income of $62.647bn was more than twice operating income; the $50.5bn Anthropic preferred-stock repricing gain accounts for most non-operating income. A reversal of Level 3 marks, for example if an IPO prices below the private-round basis or a delay supports a lower valuation, could reduce Amazon's earnings and equity before affecting AWS operating income or creditors. A listing delay alone does not mechanically require an impairment. Statements of operations and investments.

Convertible-note marks initially enter other comprehensive income. Conversion can reclassify accumulated gains into earnings; preferred-stock repricing and impairments enter other income. Thus paper appreciation can raise Amazon's reported profit while additional funding consumes cash. Carrying value is neither ownership percentage nor money available to fund data centers. Nonvoting securities and an ownership cap do not disclose an exact economic stake. Accounting disclosures; update.

The 2025 accounting evidence also includes $7.7bn of upward private-equity valuation adjustments, primarily Anthropic, alongside the conversion gains in claim 21. This further limits the usefulness of headline earnings as a measure of cash available for construction. Annual filing.

AWS remains Anthropic's primary cloud and training partner, but “main backer” should not imply sole or conclusively largest current investor: the funding announcement identifies large outside investors. Reuters independently reported the investor-meeting revenue figures after the FT's initial report. This verifies the existence of the report, not the underlying accounts. Adjusted profitability has no supplied reconciliation to GAAP profit or free cash flow; the revenue is quarterly, not annualized, and is not all AWS-addressable expenditure. Anthropic; funding; Reuters.

On September 24, 2026, Akamai disclosed an $11.6bn, seven-year Anthropic commitment for CPU workloads. Akamai estimated $5.5bn of associated capex, including $1.7bn additional 2026 capex, without changing 2026 revenue guidance. In this note's interpretation, Anthropic's growth supports multiple infrastructure suppliers and requires some suppliers to spend before revenue arrives. The announcement does not establish displacement of existing AWS workloads; its total must not be added to AWS backlog or mechanically added to other reported lab commitments. SEC-filed release.

Public contracts remain unavailable, but Reuters reported on September 29, 2026 that Anthropic's confidential prospectus contains $518bn of infrastructure commitments over a decade, about 80% non-cancelable or payable regardless of usage, including $110bn to AWS over May 2026–April 2036 with payments for spending shortfalls. These figures and terms are reported, unverified (C) and do not replace the A-grade >$100bn announcement. In this note's judgment, stronger payment obligations offer AWS contractual protection while increasing its customer's funding pressure across suppliers. Reuters.

Reuters' September 28, 2026 report also described a near-$42bn 2025 net loss, including roughly $34bn of convertible-financing revaluation charges, and nearly one-quarter of 2025 revenue from two customers, all reported, unverified (C). The loss is not cash burn; historical net losses do not contradict October's report of Q3 adjusted profitability. Customer concentration adds a separate risk to the durability of receipts needed to fund infrastructure obligations. September report; October report.

Reuters' October 14 report puts the 2025 operating loss above $8bn excluding liability writedowns and says the listing is likely to follow November's US midterms, both reported, unverified (C). The operating loss is a more relevant historical measure of operating funding needs than the $42bn net loss, although it still is not cash burn. An October 18 EDGAR full-text search for “Anthropic, PBC” located no Anthropic-issued public S-1 or S-1/A over June 1–October 18; it did return references in other issuers' filings. That limited search does not authenticate the confidential prospectus. IPO timing matters because the facility expires 30 months after a liquidity event, outstanding notes convert at an IPO and quoted market prices would test the private marks. Reuters; S-1 search; S-1/A search; investment terms.

Trainium and OpenAI. Rainier's October 2025 near-half-million count, Amazon's February 5, 2026 update to 500,000+ Rainier chips, and Anthropic's April broader million-plus deployment are dated disclosures with different scopes. The newer Rainier disclosure strengthens evidence of deployment without measuring economic utilization. Trainium3 is shipping, but public subscription statements do not supply its deployed chip count, utilization or realized customer savings. Quoted price-performance advantages are vendor comparisons, not independently demonstrated reductions in AWS's total ownership cost. Software migration, networking, memory, electricity and customer pricing all influence returns. Rainier; February update; Anthropic; chip claims.

The original OpenAI agreement was real and primarily involved NVIDIA infrastructure. The expansion adds Trainium and a much larger commercial relationship. Amazon also became an equity financier. This broadens the customer base while concentrating more capital in the same frontier-lab funding cycle. Contract totals should not be added again to backlog without knowing what is already included. Original deal; expanded deal.

Where overextension could emerge. The vulnerable boundary is between long-lived sites and fast-changing accelerators, and between contracted demand and customers' ability to pay. Amazon's server-life change shortened, rather than lengthened, depreciation lives for a subset and reduced earnings. It is evidence that obsolescence is economically relevant, not proof that every server becomes obsolete on that schedule. Rising memory costs can also increase the capital required for unchanged capacity. Annual report; guidance.

What the stock discounts. The approximate equity value uses a recent secondary-market close and an older disclosed share count; it is not a precise contemporaneous capitalization. With negative trailing FCF, a meaningful positive FCF multiple is unavailable. The illustrative yield calculation shows the scale of cash recovery an investor might require, not a reverse-engineered market consensus or price target. A favorable valuation case requires strong AWS growth, durable retail/advertising cash generation and eventual normalization of capital intensity. Investment marks make headline earnings a weak shortcut. The evidence supports a demanding execution requirement, not a demonstrated mispricing. Price; financials.

Conflicts and open questions

• Growth precision: the release headlines 37% AWS growth, while its CEO quotation says 36.7%; the published revenue amounts calculate to 36.8%. This small discrepancy does not affect the thesis; the note uses the headline figure. Results.
• Funding strain versus operating strength: the original note emphasized declining FCF and external funding alongside strong AWS operations. The cross-check agrees that borrowing demonstrates external funding and continuing market access. Its difference in emphasis is that AWS expansion is more measurable than the original consolidated-spending caveat suggested: AWS accounts for 76.0% of H1 2026 net property additions. Segment cash flow and the AI/core capital split remain unknown. September filing; segment data; cross-check.
• Current demand versus capital recovery: the original Summary called revenue, backlog and deployed chips the strongest evidence against immediate overbuild. The Codex cross-check narrowed this to present sales, contracts and deployment, without established utilization or capital recovery. Claude agrees with that limit but challenges treating backlog as broad independent demand evidence: its $252bn H1 increase coincides with >$200bn of announced OpenAI and Anthropic expansions, and Amazon supplied the labs $60bn in 2026 equity funding by the July 31 filing. The customer attribution remains C-grade inference, not a disclosed backlog breakdown. Both reviews support a real business; Claude places more weight on demand and financing being concentrated in the same counterparties. Results; commitments and funding; Q1; deployment; Codex review; Claude review.
• Backlog versus collectability: the original note treated cancellation rights as unavailable. Public contracts remain unavailable, but the cross-check adds Reuters' reported minimum-payment terms: approximately 80% of Anthropic's $518bn commitments are non-cancelable or payable regardless of usage. These unverified C-grade terms strengthen AWS's contractual position while increasing customer funding pressure; they do not guarantee collection or replace the verified >$100bn AWS announcement. Collateral and prepayments remain unresolved. Reuters; Anthropic; cross-check.
• Anthropic concentration versus outside validation: the original note emphasized outside financing and reported revenue growth as support for payment capacity, while warning of correlated financing risk. The cross-check regards that warning as too mild given reported payment rigidity, historical losses and concentration, and Akamai's competing claim on spending. Both views are retained: external validation supports the payment case, while supplier obligations and customer concentration raise its execution risk. Neither reveals Amazon's recognized Anthropic revenue, credit exposure or customer margin. Do not infer market share from investment carrying values. Funding; October report; September reports; terms; Akamai; cross-check.
• Accounting conservatism versus obsolescence: shortening server lives counters a blanket allegation of profit inflation through longer depreciation. It simultaneously demonstrates rapid hardware change. Codex estimated approximately 38.1% Q2 2026 AWS margin excluding the derivative effect from rounded management speech. Claude verified the filing's $551m gain, primarily AWS, implying approximately 38.05% when fully subtracted. These methods agree at the earlier precision; the exact AWS allocation remains undisclosed. Claude also adds the strongest pro-return testimony omitted by the earlier review: July management claims of under-three-year server/network payback, five-to-six-year useful lives and mostly five-year-plus AI contracts. The useful-life change qualifies that testimony; neither review independently establishes cohort returns. 10-Q; 10-K; call; Codex review; Claude review.
• Earnings quality and who loses first: Codex treated investment marks as a weak shortcut to cash generation and concluded growing capital-return and counterparty risk without demonstrated overbuild, insolvency or mispricing. Claude's stronger emphasis is that 66.0% of Q2 pre-tax income was non-operating and Anthropic's $190.4bn carrying value was 34.5% of book equity. Thus shareholders face a large valuation-reversal channel alongside AWS capital-return risk before creditors face distress. Both reviews retain strong AWS profitability and June's $123bn liquidity cushion; the disagreement is the scale and prominence of correlated lab exposure, not evidence of current insolvency. Earnings and balance sheet; Codex review; Claude review.
• Review precision: Claude's proposed wording that Anthropic's carrying value “rose by about $120bn” against $10bn of cash understates the total increase: $190.4bn − $60.6bn = $129.8bn; approximately $119.8bn is the residual beyond new cash. Its search description also needs a limit: EDGAR returns “Anthropic, PBC” references in other issuers' S-1 filings, but no Anthropic-issued S-1 was located. The source values and both review interpretations are retained with those distinctions. Investment balances; S-1 search; S-1/A search; Claude review.
• Unverified items: exact Anthropic stake; audited quarterly Anthropic revenue and adjusted-profit reconciliation; confidential-prospectus figures and payment terms; deployed Trainium3 count; independently measured chip advantage; funding after the July 31 disclosure, delayed-draw loan status and a complete October debt balance. No Anthropic-issued public S-1 or S-1/A was located in the specified EDGAR search as of October 18. Reuters' October 14 report says the IPO is likely after November's US midterms and 2025 operating loss exceeded $8bn excluding liability writedowns (reported, unverified, C). Thus the $518bn, approximately 80% rigid and $110bn AWS figures remain C. The facility's 30-month clock starts at a liquidity event, and outstanding notes convert at an IPO; listing and public accounts would test the marks and customer funding case. Reuters; EDGAR S-1; S-1/A; investment terms; Claude review.

Signals to watch

Sources

1. Amazon / SEC, 2025 Form 10-K, filed 2026-02-06. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm — accessed 2026-10-17; re-fetched 2026-10-18.

2. Amazon / SEC, Q1 2026 Form 10-Q, filed 2026-04-30. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000014/amzn-20260331.htm — accessed 2026-10-17; re-fetched 2026-10-18.

3. Amazon / SEC, Q2 2026 Form 10-Q, filed 2026-07-31. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm — accessed 2026-10-17; re-fetched 2026-10-18.

4. Amazon / SEC, Q3 2025 Form 10-Q, filed 2025-10-31; includes comparative quarterly and year-to-date cash flows. https://www.sec.gov/Archives/edgar/data/1018724/000101872425000123/amzn-20250930.htm — accessed 2026-10-17.

5. Amazon IR, Fourth Quarter Results, 2026-02-05. https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Fourth-Quarter-Results/default.aspx — accessed 2026-10-17.

6. Amazon IR, Second Quarter Results, 2026-07-30. https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/default.aspx — accessed 2026-10-17.

7. Amazon management, Q2 2026 earnings call, 2026-07-30; full transcript hosted by Motley Fool, published 2026-08-07. https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/ — accessed 2026-10-17; re-fetched 2026-10-18. Uses spoken transcript, not editorial summary; recording not independently transcribed.

8. Amazon, Andy Jassy on why AWS is booming, excerpt from 2026-07-30 call; page publication date not exposed. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-aws-revenue-growth-q2-2026-earnings — accessed 2026-10-17.

9. Amazon, AWS activates Project Rainier, launch update October 2025; fetched page does not expose revision date. https://www.aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster — accessed 2026-10-17.

10. Amazon, Andy Jassy on Amazon's chips business, excerpt from 2026-04-29 call; page publication date not exposed. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-amazon-chips-business-q1-2026-earnings — accessed 2026-10-17.

11. Anthropic, Amazon compute collaboration, 2026-04-20. https://www.anthropic.com/news/anthropic-amazon-compute — accessed 2026-10-17; re-fetched 2026-10-18.

12. Reuters / MarketScreener, At Anthropic pre-IPO meet: 'Thinking' caps, surging revenue, no human extinction, 2026-10-14. https://ca.marketscreener.com/news/anthropic-s-q3-revenue-surges-45-to-nearly-17-billion-ce785dd2d08cf221 — accessed 2026-10-17; re-fetched 2026-10-18. Secondary; direct FT article not accessible.

13. OpenAI, AWS and OpenAI announce multi-year strategic partnership, 2025-11-03. https://openai.com/index/aws-and-openai-partnership/ — accessed 2026-10-17.

14. Amazon, OpenAI and Amazon announce strategic partnership, 2026-02-27. https://www.aboutamazon.com/news/aws/amazon-open-ai-strategic-partnership-investment — accessed 2026-10-17; re-fetched 2026-10-18.

15. Amazon / SEC, Debt issuance Form 8-K, 2026-07-09. https://www.sec.gov/Archives/edgar/data/1018724/000110465926082293/tm2619352d4_8k.htm — accessed 2026-10-17; re-fetched 2026-10-18.

16. Amazon IR, Second Quarter Results, 2024-08-01; historical FCF definitions and reconciliations. https://ir.aboutamazon.com/news-release/news-release-details/2024/Amazon-com-Announces-Second-Quarter-Results/default.aspx — accessed 2026-10-17.

17. ChartExchange, AMZN historical prices, observation 2026-10-16; publication date not stated. https://chartexchange.com/symbol/nasdaq-amzn/historical/ — accessed 2026-10-17. Secondary; historical closing row used, not intraday header. Nasdaq's fetched page did not expose its historical rows.

18. Anthropic, Series H funding announcement, 2026-05-28. https://www.anthropic.com/news/series-h — accessed 2026-10-17.

19. Amazon Press Center, Third-quarter earnings webcast announcement, 2026-10-15. https://press.aboutamazon.com/2026/10/amazon-com-to-webcast-third-quarter-2026-financial-results-conference-call — accessed 2026-10-17.

20. Amazon / SEC, Sterling debt issuance Form 8-K, 2026-09-14. https://www.sec.gov/Archives/edgar/data/1018724/000110465926107526/tm2624614d5_8k.htm — fetched 2026-10-17; re-fetched 2026-10-18; tranche principal summed for claim 35.

21. Akamai / SEC, Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic to Support Growing Demand, 2026-09-24. https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm — fetched 2026-10-17; company announcement and capex estimates.

22. Reuters / Investing.com, Anthropic's $518 billion AI buildout hinges largely on deals that cannot be canceled, filing shows, published and updated 2026-09-29. https://www.investing.com/news/stock-market-news/anthropics-518-billion-ai-buildout-hinges-largely-on-deals-that-cannot-be-canceled-filing-shows-4922031 — fetched 2026-10-17; confidential-prospectus figures and terms reported, unverified (C).

23. Reuters / MarketScreener, Anthropic's IPO prospectus shows sweeping AI vision, surging costs, published 2026-09-28, updated 2026-09-29. https://www.marketscreener.com/news/anthropic-s-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-ce785addd98afe2c — fetched 2026-10-17; historical loss and concentration reported, unverified (C).

24. SEC, Amazon EDGAR submissions, filing list through 2026-10-18. https://data.sec.gov/submissions/CIK0001018724.json — fetched 2026-10-18; checked September 15–October 18 for a draw-announcing 8-K. Absence of that announcement does not establish the loan balance.

25. SEC, EDGAR full-text search: “Anthropic, PBC”, S-1 and S-1/A, June 1–October 18, 2026. https://efts.sec.gov/LATEST/search-index?q=%22Anthropic%2C%20PBC%22&forms=S-1&dateRange=custom&startdt=2026-06-01&enddt=2026-10-18 and https://efts.sec.gov/LATEST/search-index?q=%22Anthropic%2C%20PBC%22&forms=S-1%2FA&dateRange=custom&startdt=2026-06-01&enddt=2026-10-18 — fetched 2026-10-18; neither search returned an Anthropic-issued filing, although other issuers' references appeared. Search scope and issuer identity limit the absence claim.

26. Amazon / SEC, Globalstar information statement/prospectus, Form 424B3, 2026-08-18. https://www.sec.gov/Archives/edgar/data/1018724/000110465926098339/tm2617924-6_424b3.htm — fetched 2026-10-18; April 13 agreement terms and signing share count, not fixed closing consideration.

27. SEC, Notice of effectiveness: Amazon Form S-4, effective 2026-08-18. https://www.sec.gov/Archives/edgar/data/1018724/999999999526002695/xslEFFECTX01/primary_doc.xml — fetched 2026-10-18; confirms Globalstar registration effectiveness.

28. Amazon, Q3 2026 earnings schedule, re-fetched 2026-10-21: October 29 call at 5pm ET. This is a schedule re-check; July capex guidance and June cash metrics remain prior disclosures pending results.

The Codex cross-check was applied on 2026-10-17; the Claude cross-check was applied on 2026-10-18. For the latter, the 2025 10-K, Q1 and Q2 2026 10-Qs, July and September debt filings, July call transcript, OpenAI expansion, Anthropic compute announcement, October Reuters report, Globalstar prospectus and effectiveness notice, and EDGAR submissions/search endpoints were fetched to verify the additions. SEC requests used the generic User-Agent research-bot contact@example.com. Both reviews found no arithmetic error in the pre-existing claims or CSV, so the CSV is unchanged. New ratios and scenarios were recalculated; the Anthropic carrying-value increase is $129.8bn, of which approximately $119.8bn is beyond new cash, correcting the review's ambiguous ~$120bn wording.