Anthropic’s S-1.
The demand is real.
The funding test is ahead.

Executive call

Anthropic’s first public S-1 strengthens the case that enterprises are paying for frontier AI at extraordinary scale. It does not establish that the industry’s infrastructure buildout will earn its cost of capital. The important surprise is the combination: revenue and operating leverage have accelerated, while the company is moving toward much larger, less flexible infrastructure obligations and financing that can become more expensive or less available when growth slows.

H1 2026 revenue was $16.286bn, versus $1.163bn a year earlier. Q2 produced $0.974bn of GAAP operating income. H1’s $80.748bn net loss is not cash burn: two noncash instrument adjustments total $81.108bn. Conversely, $3.929bn of operating cash inflow is not proof of sustainable self-funding. Collections, prepayments, supplier payment timing, future leases and upfront construction spending all matter. S02: MD&A—Results of Operations and Liquidity; financial statements

Investment judgment: strengthen the demand thesis; re-underwrite the duration and credit quality of the supply response. NVIDIA, AMD and Broadcom have identifiable compute channels, but maximum contract values are not their revenue or guaranteed backlog. Hyperscalers gain distribution and utilization while assuming correlated customer, supplier and investment risks. Software faces both substitution and a larger implementation opportunity. Infrastructure lenders need cash collections and collateral recovery analysis, not a large AI headline.

Action posture: watchlist / re-underwrite / wait for proof. No Anthropic price target or IPO recommendation is defensible before price, share count, updated liquidity and a complete obligation schedule are available. Listed-company exposures below are research priorities, not portfolio allocations.

Event status and market baseline

Verified event. SEC accession 0001628280-26-067168, filed 26 October 2026. The index reports acceptance at 09:00:11 Eastern, equivalent to 13:00:11 UTC. This is a registration statement, not an effective/priced/completed IPO. Offering price, shares offered and net proceeds are blank. S01: SEC filing index

Information cutoff: 26 October 2026, 17:20 UTC for the source set; the report was completed later without incorporating later market events. Annual statements through 2025 are audited; the six months ended 30 June 2026 are unaudited. Selected operating metrics extend through September; subsequent events were evaluated through 26 October. Some October 31 run-rate fields are blank and refer to a future date. They are not observations. Cash below is a June balance, not an October estimate.

Observed market context, not an event study. The inv-AI quote feed at 17:15:27–35 UTC reported NVDA +2.55%, AMD +2.97%, AVGO +2.84%, MSFT +1.22%, AMZN +0.70%, GOOGL −0.37%, META −0.20%, MU +4.10%, ANET +3.81% and VRT +3.29%. Earlier independent finance-feed observations around 17:00 UTC showed NVDA +2.33%, AMD +2.63%, AVGO +2.57% and SPY +0.73%. Different timestamps are intentionally retained. The independent feed did not supply a verifiable underlying venue URL; it is a directional cross-check, not an exact-price verification. These are intraday feed changes; prior-close methodology was not independently reconstructed. They do not measure a filing-attributable return. Other financing announcements, positioning and market news coincide. M01: dated quote capture

The immediate repricing candidates are perceived AI demand durability, supplier order quality and funding costs. What is already priced is unresolved. The quote pattern supports co-movement in AI infrastructure shares; it supplies neither an expectations baseline nor evidence of undervaluation. No consensus revision, options event study or comparable enterprise-value dataset is assumed.

Source and freshness

The principal source is the actual full S-1, downloaded as approximately 14.17 MB of SEC HTML, parsed into full text and 171 tables, and retained with its hash. Analysis uses the prospectus, MD&A, financial statements, accounting policies, commitments and subsequent events together. References use section names and note titles because extracted page-break markers can differ from the prospectus table of contents. The retained response is 14,169,252 bytes; the SEC index lists 14,169,143 bytes. The retained HTML includes a dynamic server script, and its actual hash identifies this research copy. A local source ledger records URLs, periods, captures, limitations and hashes.

Primary issuer releases supply the ecosystem comparisons. Those releases mix calendar and fiscal quarters and are used as dated context, not a common-period peer ranking. Management forecasts, adoption claims and infrastructure schedules retain that status. No paid consensus dataset, independent enterprise ROI panel or token-level cost model was available. The important missing facts are the latest balance sheet, complete obligation maturity ladder, actual current financing draws, absolute cloud fee schedule and final IPO capitalization. Their absence limits valuation and precise runway, rather than invalidating the verified historical statements.

All amounts are US dollars. Fact denotes a filing or issuer disclosure; calculation denotes arithmetic from identified inputs; interpretation denotes inv-AI’s judgment; illustration denotes assumptions rather than company guidance. A source ledger accompanies the report; prospective uncertainty remains explicit.

What changed: operating leverage, with a larger capital burden

Source: S02: MD&A—Results and Cash Flows; statements and Note 8. H1 2025 SBC cash-flow addback is $0.279bn versus $0.286bn expense; the table above uses expense, and the cash bridge uses statement addbacks.

Calculation. Q2 2026 revenue of $11.555bn was 144.2% above Q1’s $4.731bn; operating margin moved from −35.2% to +8.4%. Revenue of $0.787bn in Q2 2025 makes the same-quarter growth 1,368% rounded. That is a different comparison from H1’s 1,300%. Multiplying Q2 by four gives $46.220bn, a scale reference only. It is not management’s Run-Rate Revenue, an FY2026 forecast or a contract backlog. S02: MD&A—Quarterly Results of Operations

The strongest bullish evidence is revenue outrunning a compute cost base committed in advance. The strongest caution is the same operating leverage in reverse: capacity becomes expense as delivered, whether current demand arrives or not. A very good quarter can precede a less favorable capacity ramp. Management explicitly warns that operating and adjusted margins may compress as committed compute arrives and higher-cost short-term compute is used. Note 17 also discloses a $2.0bn Q3 donation-matching charge, settled in October with approximately 28.0m Class N shares; the liability is remeasured through settlement. Q3 GAAP operating results absorb a noncash charge roughly twice Q2 operating income; adjusted operating income excludes it. This is also a dilution event, with no complete Q3 earnings statement supplied. New transparency is valuable; the filing alone cannot establish how far these facts differ from prior private-market expectations.

Revenue quality: paid use, young cohorts and policy leakage

Consumption revenue is recognized when intelligence is used; subscriptions are generally recognized ratably and have usage limits. Subscription terms are generally twelve months or less; many customers can reduce use or leave with little notice. Payment terms are generally under sixty days. This is monetization evidence, but its duration differs from a conventional multiyear contracted software annuity. S02: Note 1 revenue recognition; Risk Factors

There is meaningful broadening. Customers producing at least $100,000 of trailing-year revenue exceeded 6,000 at June 2026 versus over 600 a year earlier; they accounted for about $12.0bn, or 74%, of H1 revenue. By September the count exceeded 11,900, with more than 1,900 at $1m and roughly 250 at $10m. No customer exceeded 10% of nine-month revenue; the top five represented about 21% of Q3 revenue versus 36% of Q1; the nine-month top-five share was about 25%. Keep those September observations separate from H1: Customer C represented 11% of H1 revenue, while A and B were individually below 10%; June receivables included 13% for A and 10% for C. S02: MD&A—Enterprise Adoption and Results; Note 1 concentration

The approximately 900% September net revenue retention is striking, but narrowly defined. It compares current-quarter revenue from customers with at least $1 of revenue in the corresponding prior-year quarter against that base-quarter revenue, excluding Pro/Max subscribers. A tiny starting spend, fast deployment and a young cohort can produce extreme expansion. It is not 900% recurring ARR, a forecast or a standard SaaS peer comparison. Cohort aging, contraction, churn and customer-funded ROI matter more than extending that number forward. S02: Definitions—Net Revenue Retention; Summary—Strengths of Our Financial Model

A material qualification: management estimates 5–7% of revenue for the nine months ended September 2026 came from customers subsequently found in breach of usage policies or commercial terms; detection improved in August/September. Revenue represented service already delivered before termination. This creates a future demand and enforcement trade-off, not evidence of an H1 restatement or fictitious sales. Nine-month revenue dollars are not supplied in the completed statements, so an exact dollar deduction cannot be made. A 5–7% haircut to future revenue is a stress assumption, not a historical correction. S02: MD&A—Key Factors Affecting Our Performance—Revenue Quality

June remaining performance obligations were $10.8bn, with roughly 75% expected to become revenue over the following twelve months: $8.1bn by calculation. These are customer arrangements with noncancelable commitments. They are neither cash already received nor the same thing as Anthropic’s purchases from infrastructure vendors. Compare them with June deferred revenue of $1.650bn and with the broader, usage-driven revenue base; do not divide RPO into an annualized quarter and call the result revenue visibility. S02: Note 2

Interpretation. Sustainable demand requires customer productivity benefits that survive credits, model substitution and tighter policy enforcement. Strong spending proves willingness to buy at current terms. It does not reveal the customer’s realized return, the amount subsidized by venture funding, or demand at a higher net price. Prioritize cohort cash collections, new-versus-existing customer expansion, credits, churn and workload renewal evidence.

Gross versus net: normalize the economics before comparing peers

Anthropic is the principal for marketplace sales: it directly contracts with customers, sets pricing and is primarily responsible for the service. CSPs do not control the service before transfer. It therefore recognizes revenue gross and pays cloud channel fees as expenses. Marketplaces represented 48% of H1 2026 revenue, versus 44% in H1 2025, and 58% of June receivables. Clouds act as collection agents. This is distribution concentration; it is not evidence that the clouds themselves are the ultimate buyers of all those services. S02: Note 1—Revenue Recognition and Concentration

The functional expense schedule permits a clean bridge:

Source: S02: Note 8, Supplemental Operating Expense Information. Gross profit/margins are calculations using the functional costs, not separately supplied headline KPIs. H1 2025 derived gross margin on the same definition was 12.4%.

Substantially all cost of revenue is customer-serving compute and infrastructure, with personnel and overhead. R&D includes training, research, internal usage and personnel. Channel costs are in sales and marketing. Therefore 85.4% is not an all-in frontier-model economic margin. Nor is revenue less all compute, 45.1%, a GAAP gross margin: that subtraction mixes research and serving costs while omitting other functional costs.

MD&A discloses a $2.7bn increase in channel fees. The entire H1 2025 nature-classified sales/marketing/partnerships expense was only $0.120bn, so its channel-fee component could not exceed that total. Allowing approximately $0.05bn of rounding in the disclosed increase implies a rough H1 2026 fee bound of $2.65–2.87bn, rather than an unconstrained unknown. This is a bound from rounded disclosures, not an exact fee total. A counterfactual presentation exercise uses values inside that bound:

Illustration: subtract K from both gross revenue and sales/marketing, leaving all other costs fixed. Gross margin becomes (revenue − K − cost of revenue)/(revenue − K). This is not a proposed GAAP restatement; K values remain assumptions within the disclosure-implied range. Operating dollars do not change, while revenue multiples and margins do. A peer comparison must also reconcile serving arrangements, free usage, shared R&D, SBC and leases. Never subtract all 48% marketplace sales, treat the disclosed $2.7bn increase as the exact fee total, or add lab and cloud gross revenue as independent end demand. S02: Results of Operations—Sales, Marketing, and Partnerships; Note 8

Inference and training: improvement is evident; token economics remain unknown

H1 revenue per dollar of total compute expense improved from $0.498 to $1.821. This is a useful company-level utilization and scale indicator, with a fixed-capacity and product-mix component. It cannot identify inference cost per token: the numerator includes gross distribution revenue and the denominator includes research, serving and internal use. No sufficient token volume, model-level price realization, cache mix, training/inference allocation or vintage-level return is disclosed. S02: MD&A—Key Factors and Results; Note 8

Management attributes demand expansion to better intelligence per dollar and reports approximately 40% lower customer task costs for Opus 5.5 versus Opus 5. That is a company comparison combining capability, price and task efficiency; it is not a measured industry price elasticity. Longer reasoning, autonomous tasks, caching and model mix can move tokens per paid outcome in opposite directions. More tokens are not automatically more useful work, and lower cost per task does not automatically reduce total compute spending.

The relevant operating identity is paid workloads × realized price per workload − serving costs − channel fees − continuing model-development costs. A stronger model can increase willingness to pay and lower customer labor costs even as token prices fall. Competition can transfer those gains to customers. Training remains economically recurring as long as remaining at the frontier is necessary for retention; removing R&D from the investment case would treat renewal of the core product as optional.

The sensitivity chart indexes starting revenue and serving costs to 100 and 15. It changes realized price and serving cost per unit, holding paid units fixed. A 30% price fall with only a 10% serving-cost decline compresses pre-channel contribution from 85 to 56.5 per initial revenue 100. If both fall 30%, the contribution rate remains 85% but contribution dollars fall 30%. Higher units can offset this, but required utilization, capacity and ongoing R&D must be funded. These are arithmetic illustrations, not estimates of Anthropic’s price or serving cost.

Cash flow: reconcile the loss, then stress its quality

Source: S02: Consolidated Statements of Cash Flows. Pre-working-capital cash earnings and the final measure are inv-AI calculations, not company non-GAAP measures. They include investment income and exclude noncash acquisitions/new lease assets.

Two mistakes point in opposite directions. Removing $81.108bn of noncash instrument adjustments from the net loss leaves +$0.360bn, not an $80bn operating cash hole. But positive OCF is helped by a $1.168bn net operating working-capital inflow. Payables and prepayments alone contributed $4.773bn while receivables absorbed $4.982bn. CSP payables rose from $0.947bn in December to $2.810bn in June; prepaid current expenses fell from $2.537bn to $0.258bn. Those balance movements are contextual and not substituted for cash-flow-statement movements. S02: Note 8; MD&A—Cash Flows

Normalization must be a range of treatments, not a magic burn rate.

These are selective mechanical sensitivities, not forecasts or complete normalized free cash flow. Zero working capital is generous for a rapidly growing usage business. Removing all favorable movements while preserving receivables investment is deliberately adverse. Deferred receipts can be a legitimate operating financing source. Interest income of $1.111bn boosts earnings and depends on the financing-funded investment balance; it is not model-serving profitability. Subtracting it from pre-working-capital cash earnings leaves approximately $1.650bn, a limited operating sensitivity rather than a full normalized cash measure. Earlier prepaid hosting consumption also carries a timing history: management attributes much of FY2024 cash use to upfront hosting purchases and part of FY2025 improvement to consuming capacity paid in prior periods. Releasing prepayments is not repeatable free compute. The proper forward model needs billings, cash collections, payment terms, prepayments, cash capex and lease starts by period.

H1 investing outflow of $17.181bn includes net short-term security purchases of $16.362bn. That portfolio reallocation is not infrastructure burn. Other investing outflow was $0.337bn and should be examined separately rather than automatically treated as recurring capex. A further economic cash qualification sits outside that measure: the company deposited $2.5bn into infrastructure-related escrow in April, within restricted cash rather than an OCF/investing outflow. Deducting that known deposit from $3.443bn leaves $0.943bn of cash generation after those selected uses; this is an analytical deployable-liquidity measure, not GAAP cash flow. Restricted cash is already excluded from the liquidity pool below, so this adjustment must not be deducted again there. The $1.2bn Bartz settlement liability still current at June is another future cash claim, despite no H1 settlement expense. S02: Notes 8, 13 and 16

Financing inflow of $110.774bn, including $110.432bn net preferred issuance, explains the balance-sheet expansion far more than OCF does. Cash, equivalents and restricted cash rose $97.517bn to $114.918bn. The cash-flow identity reconciles: 3.929 − 17.181 + 110.774 − 0.005 FX = 97.517. S02: cash-flow statement

Liquidity and runway: distinguish cash from conditional capacity

At June 30, cash and equivalents were $112.389bn, short-term investments $19.262bn, and restricted cash $2.529bn. The first two sum to $131.651bn; restricted cash is separate and excluded. The company may have to restrict up to an additional $10bn under certain arrangements, though no such requirement was in effect at June. There is no defensible October cash roll-forward from the filing’s blank latest fields. S02: balance sheet, Notes 3 and 8

Management says available resources should support requirements for at least twelve months from the prospectus, while warning that capacity investment can require significant net cash use even if operating performance improves. The move from CSP service purchases toward owned/leased equipment, construction and upfront deposits makes a backward-looking OCF extrapolation particularly unsafe. S02: MD&A—Liquidity and Capital Resources

Illustrative liquidity endurance from the June 30 starting point, excluding all future financing and IPO proceeds:

Formula: 12 × deployable starting funds / annual net cash use. The buffered pool is $116.651bn. The $5bn is an analytical operating cushion inspired by the later covenant; it is not a reconstruction of legally defined covenant liquidity. Actual covenant definitions and eligible facilities matter. This constant-use exercise is not runway measured from October, a cash-flow forecast, or a demonstration that funds can meet lumpy payments. It excludes acquisitions, changes in collateral, seasonality and liquidity haircuts. Contract duration can exceed any of these horizons.

Compute obligations: the cancellation clause changes the trade

Source: S02: MD&A—Contractual Obligations; Notes 7, 13 and 17. These are overlapping contract descriptions and balance-date disclosures. Do not sum this table into a backlog, current liability or unique lifetime cash requirement.

SpaceXAI’s full-capacity fee is approximately $3.1bn a month, with full delivery expected in December 2026 and different ramp fees. Initial noncancelable periods end in January 2027 and March 2027 under the respective agreements; afterward either party can terminate on 90 days’ written notice. The $120.5bn maximum is neither guaranteed vendor backlog nor a fully irrevocable take-or-pay obligation through 2029. The right cuts both ways: Anthropic gains a capacity escape route but also supplier-continuity risk. The exact minimum exposure requires individual ramp schedules, notice mechanics and fees not reconstructed here. Multiplying $3.1bn by twelve gives $37.2bn of annual full-ramp fee pace, conditional on capacity and continuation; it is not an incremental addition to other disclosed totals. S02: Contractual Obligations—Infrastructure Service Arrangements

The notes show the scale more clearly than a misleading grand total. At June, $265.6bn of fixed/minimum purchase commitments ran one to twenty years; $321.4bn of estimated undiscounted leases had not commenced, some subject to conditions. Commenced lease liabilities were much smaller: $1.079bn operating and $0.443bn finance. Subsequent new/amended arrangements brought approximately $245bn of future minimum obligations plus $21bn of short-term minimum noncancelable payments. Those later figures include amendments, and named agreements overlap the categories. The integrated latest obligation/maturity table contains blanks; a precise roll-forward cannot be established by addition. S02: Notes 7, 13, 17

Interpretation. The risk is a mismatch between cancelable, usage-sensitive receipts and long-lived, partly fixed payments. Different contracts put delivery, power availability, financing and idle-capacity risks on different parties. Defaults and cross-defaults can defeat the reassuring impression of legal entities separated from the parent. Current recognized lease balances do not describe the future economic exposure.

Financing: availability deteriorates when it is most valuable

Source: S02: Liquidity—Revolving Credit and Convertible Notes; Note 17. The June predecessor revolver had no borrowings and $78m letters of credit; that does not establish the September facility’s undrawn amount on October 26. NPA capacity totals $77bn at June, not $77bn cash on hand. No financing-capacity subtotal is added to the runway table.

Google’s option for up to $10bn in an IPO-concurrent private placement reduces its $20bn NPA capacity if exercised; it is not another $10bn layered on top. Amazon’s ownership cap is 12%; Google’s is 14.99%, on their contractual capitalization definitions. Rollover notes can defer conversion, with contingent cash obligations and resets; ownership caps do not eliminate claims. S02: Liquidity—Convertible Notes and Note Purchase Agreements

The $42bn Broadcom facility finances only quarterly rent under the 2026 Lease Arrangement, whose rent is approximately $125.2bn excluding interest; it is not funding for every obligation in the broader $161.2bn master-lease description. The Broadcom facility makes the downside convex. Notes generally accrue 9% payment-in-kind interest and convert at 90% of trailing 30-day VWAP. Notes issued when certain revenue conditions are unmet accrue 13% and convert at 85%; sustained failure can terminate access. Conditions also reference minimum liquidity and lease performance. In a slowdown, cash need can rise while access shrinks and the cost of surviving rises. A change-of-control/liquidation cash option at 150% of then-outstanding principal and accrued interest adds a separate contingent claim. S02: Liquidity—Convertible Notes—Broadcom

Illustration, one $10bn draw: after three years with annually compounded PIK, principal is $12.950bn at 9% or $14.429bn at 13%. Dividing by the conversion discounts produces $14.39bn / $16.98bn of then-market equity value represented at conversion, if the VWAP equals that illustrative market price. This is an exposure illustration, not the actual cash cost, precise contractual accrual schedule or a forecast of conversion. Actual issue dates, compounding, VWAP, stock price, redemption/default outcomes and financing availability determine the result. At unchanged current equity value, lower prices require more shares for the same dollar claim.

Accounting loss, stock compensation and IPO dilution

The $72.388bn expense principally reflects higher fair values of outstanding notes as Anthropic’s valuation rose. The $8.720bn separate cash-flow adjustment covers other financial-instrument recognition/remeasurement, including note-purchase economics. Note 8 disaggregates other expense into $9.405bn NPA expense, $0.039bn holdback expense, $0.724bn warrant gain and $0.037bn other expense: net $8.757bn. The $8.720bn adjustment is not identical to the entire other-expense line. These marks are noncash in the period but describe valuable claims and financing concessions; ignoring their economic dilution is also wrong. S02: Notes 4, 8 and 9

June balance-sheet convertible notes of $108.566bn are fair value, not $108.566bn cash principal debt. The outstanding Amazon principal was $6.203bn; disclosed fixed split-adjusted conversion prices imply about 1.840bn Class N shares before ownership-cap mechanics. A valuation must handle the convertible claim consistently in the cash/debt/share bridge, not subtract its fair value as cash debt and also count all conversion shares. S02: Liquidity—Convertible Notes—Outstanding Notes; Note 9

SBC is a recurring economic cost even when it does not consume immediate cash. H1 expense was $2.185bn. At June, unrecognized RSU expense was $26.6bn, including about $2.9bn for service already wholly/partly satisfied and $23.7bn future cost; unrecognized option compensation was $2.8bn. Listing effectiveness can satisfy RSU performance conditions and trigger a large catch-up. Later grants of 91m RSUs, with $5.8bn grant-date value, and October executive modifications mean the June catch-up is not an exact IPO charge. The Q3 $2.0bn donation-matching charge and October settlement in roughly 28.0m shares add a separate expense/dilution event; the final settlement-period mark is not yet quantified. Adjusted operating margin excludes specified donation charges, so it should sit beside, not replace, GAAP and dilution analysis. S02: Critical Accounting Estimates—Stock-Based Compensation; Notes 10 and 17

The September ten-for-one split applies to common shares and awards; preferred issue prices retain their own units. Mixing a preferred financing price with split-adjusted common shares can create a tenfold error. The 2026 equity plan reserves 2.950bn shares and ESPP 160m, with annual increases; reserves are not all issued today. Blank offering terms prevent a precise post-IPO denominator and proceeds calculation. We publish neither an invented IPO market capitalization nor a precise enterprise value. S02: Note 17, capitalization and dilution

Governance: public economics without public control

The LTBT generally elects/removes a majority of authorized directors through non-economic Class T stock. The founders’ Class F share carries 50.1% of the general voting power, including election of non-Class-T directors; public Class A generally cannot determine outcomes. These rights coexist; they should not be added as if voting on an identical constituency. The PBC’s duty balances shareholders, affected stakeholders and its stated public-benefit purpose. This can support long-term trust and responsible deployment, but investors cannot assume conventional shareholder primacy or takeover discipline. S02: Our Governance Structure and Its Goals; Risk Factors—Governance

The initial 846.841m Class N public-benefit shares, representing 5% of June fully diluted shares assuming Amazon conversion, sit in a wholly owned fund initially. The accounting/ownership treatment differs before external transfer; treating them as both treasury-like and fully outside-held would double count. Future donations or sales can change minority economics. The board also has a discretionary framework for up to 2% additional shares at successive 50% price milestones, with 90-of-120-day and 10-day-average tests. It is not an automatic annual 2% issuance. S02: MD&A—Our Commitment to Public Benefit; governance and shares

Governance warrants an explicit assumption in any eventual valuation: how much cash/equity can be directed to public benefit, which decisions public holders can influence, and the return required for those rights. That is a real economic question rather than a numerical discount selected without evidence.

From Anthropic to the market: a transmission map

The causal spine is customer outcomes → paid usage and collections → model-lab contribution → compute utilization and infrastructure payments → supplier cash conversion → reinvestment or credit loss. Equity expectations can reprice immediately; recognized semiconductor revenue follows delivery/accounting, while infrastructure cash flows can lag years. A model-quality gain may grow both demand and competition. A cancellation can relieve Anthropic while weakening a supplier’s asset utilization. The same event has different signs across the chain.

Issuer support: E03: AMD partnership, E04: NVIDIA Q2 FY27, E06: Broadcom Q3 FY26, E07: Microsoft FY26 Q4, E08: Amazon Q2, E09: Alphabet Q2, E11: Micron FY26 Q4. Networking, electrical and software rows identify mechanisms for diligence; issuer-specific orders and fair values have not been established.

The hardware read-through is broader than GPU unit volume: foundry capacity, advanced packaging, HBM, server DRAM, SSDs, optics and power are shared bottlenecks. Substitution can redistribute rent from a GPU platform to custom silicon while preserving some memory/power demand. It does not ensure all suppliers win: lower precision, better caching, utilization and shorter model inference can reduce resource intensity per task.

Geographic risk concentrates in semiconductor manufacturing, packaging and memory supply even if model customers diversify. Export controls can limit addressable markets and reorder procurement. Predominantly dollar revenue reduces one near-term FX translation issue for Anthropic, but global customers’ budgets and supply chains still transmit currency and geopolitical shocks. There is no basis for a precise country-level earnings allocation from this filing.

Compare primary disclosures on their actual definitions

Amounts in $bn. Sources E04, E05, E06, E07 call, E08, E09, E10.

This table deliberately does not rank margins, sum sales into an AI market size, or annualize different fiscal quarters into a supposedly comparable panel. It shows why an infrastructure supplier’s cash-generative business and a model lab’s funding requirement can both be real. Financing-related investment gains should be kept out of operating demand comparisons, just as Anthropic’s instrument losses should be kept out of cash-burn estimates.

Financing the ecosystem: interdependence is not proof of fake revenue

Meta’s October 26 announcement launched approximately $35bn of offerings: $22.5bn of common stock associated with forward sales and $12.5bn of mandatory-convertible depositary shares, before options. It expects no initial proceeds from forward purchasers’ common-stock sales; cash depends on later settlement and terms. The separate potential $60bn ATM was not contracted, with timing/size/terms undetermined and establishment not expected until Q1 2027. NVIDIA/AMD/Broadcom interest in up to $5bn aggregate common purchases was nonbinding. Do not call $95bn cash raised today, or treat supplier interest as completed financing. E01: Meta launch

SpaceX’s October 24 binding commitment is to enter an Apollo-managed GPU master finance lease, with closing conditions and expected definitive entry before year-end. Schedules are expected to begin H1 2027 as GPUs arrive; roughly $40bn of aggregate principal payments span approximately five-to-five-and-a-quarter-year tranches. This is not $40bn cash already received. The primary 8-K leaves the fixed rate to pricing; a reported 8.75% on the most expensive tranche is not a verified blended rate and is not used in this report’s calculations. E02: SpaceX 8-K signed Oct 26

Interpretation. Supplier equity, credit support and customer commitments can be rational ways to finance a valuable technological transition. They also correlate balance sheets: a weaker model customer can reduce utilization, impair a supplier investment, increase collateral needs and weaken a lender’s residual recovery at the same time. Funding an asset is not proof that independent users will pay enough to support it. Nor does the presence of supplier financing prove fraudulent or circular booked revenue.

The diagnostic is cash outside the financing loop: identify ultimate customer collections, discounts and credits; distinguish investors from buyers; reconcile receivables and related-party balances; test whether purchase obligations are arm’s-length and funded; compare capacity utilization with fee minima; and establish how much risk has been transferred versus retained. Marketplace gross presentation can create overlapping corporate sales without invented end-user activity. Independent financing reduces one link, not the physical utilization risk.

For banks, insurers, BDCs, infrastructure funds and equipment lessors, credit analysis needs a dated maturity ladder, recourse and cross-default map, replacement-cost/residual-value stress, and concentration across apparently separate SPVs. Rapid GPU obsolescence can shorten economic asset lives relative to debt duration. A lender can own the equipment and still suffer if power, compatible systems, resale demand or legal enforcement disappoint. No numerical system-wide loss estimate is justified from these issuer disclosures alone.

Scenarios and sensitivities: useful ranges without invented IPO terms

The following are illustrative FY2027 operating cases, designed to expose fixed-capacity risk rather than forecast a precise result. They are not management guidance, consensus, probabilities, fair values or an exhaustive distribution. All inputs are assumptions. Revenue is anchored only to Q2 scale; aggregate compute is deliberately held constant across cases. No FY2026 full-year revenue or future contract cash schedule is invented.

The $65bn compute level is not a contractual minimum or a demonstrably conservative forecast. The completed contract maturity, delivery and accounting schedule is missing; actual compute expense could be higher. It includes assumed finance-lease ROU amortization as part of operating compute expense, but we do not assert an unsupported amortization amount. Other expenses include personnel, distribution, non-compute research and overhead, with more variable spending at higher revenue. Other costs include channel fees at roughly the H1 2026 rate of approximately 16–18% of revenue; a lower rate requires evidence of mix or fee-term changes. SBC assumptions in the mid-teens or above reflect the large unrecognized award pool and remain estimates, not a replay of award vesting. The $2bn annual donation charge is a separate illustrative provision, not a prediction that the Q3 2026 event repeats unchanged. New instrument marks and an IPO RSU catch-up are not projected; this is an operating, not net-income, exercise.

Bear mechanism: weaker renewal, policy enforcement and price competition arrive while fixed compute expense persists; financing tests tighten and collateral can consume cash. Base mechanism: paid usage grows strongly but frontier research and delivered capacity absorb much of the benefit; capital markets remain important. Bull mechanism: mature enterprise deployment and high utilization spread the same assumed capacity cost over more revenue. A larger capacity purchase or higher serving intensity would reduce the bull margin.

Cash cases remain separate from these earnings cases. A credible cash bridge requires ROU amortization, finance-lease principal and interest, cash equipment purchases, construction deposits, working capital and investment income by period. Large principal payments with a small D&A addback would be internally inconsistent once finance leases commence at scale. The filing does not support precise 2027 schedules, so none is supplied. The separate $20/$40/$60/$80bn annual cash-use table tests liquidity endurance without pretending to derive those uses from these operating cases. Net interest cannot be assumed constant when cash declines and financed equipment grows. GAAP improvement, OCF improvement and deployable-liquidity improvement can differ.

The main sensitivity is joint revenue and total operating cost, not an isolated revenue CAGR:

Cells are operating income. This grid varies costs independently; the full cases above use costs of $100/$107/$115bn, respectively. At base revenue of $90bn, an additional 5% revenue leakage removes $4.5bn of operating income if costs are fixed; a 7% leakage removes $6.3bn. Applying those percentages is a forward sensitivity inspired by the nine-month policy disclosure, not a correction to H1. At a hypothetical incremental contribution rate of 50%, a 6% revenue loss instead costs $2.7bn before fixed-cost action. Revenue dollars alone do not settle profit.

For eventual valuation, ask for a range of sustainable after-R&D cash returns, reinvestment, dilution, discount rates and terminal competitive outcomes. Every $1bn of durable annual distributable cash has a different value from a temporarily unspent $1bn financing balance. With price and shares unset, a precise per-share target would hide the central uncertainty rather than resolve it.

Catalysts, falsifiers and the research queue

The shareholder release summary is subject to holder-specific bases, exceptions, underwriter waivers, pro-rata release triggers for other holders and next-trading-day rules. The filing leaves share counts blank; no calendar IPO date is assumed. For some holder groups, post-FY26 release increments occur at 13/27/41/55/69 days; extended holders have further Q1/Q2-linked stages. Eligibility is not a sale, and the table is not a full legal lock-up schedule. S02: Underwriting—Lock-up and Market Standoff Agreements; Shares Eligible for Future Sale

Dates are filing/issuer expectations or explicit contract windows, not confirmed earnings-release dates. The filing discloses June worldwide Commerce restrictions that temporarily disabled Fable 5 and Mythos 5 for all customers until lifted, an October FTC civil investigative demand, and a September putative competition-law class action. These are concrete product-availability, compliance and litigation risks; outcomes are uncertain. It also discloses conflicting court outcomes relating to government restrictions; this report does not supply legal advice or assume final resolution. S02: Risk Factors, Notes 13, 16 and 17, E02, E03.

Falsification thresholds are research rules, not statistical forecasts. Revisit the operating-leverage thesis if two consecutive reported quarters show compute/infrastructure growing faster than revenue while cash collections lag receivables and capacity expands. Revisit the funding-risk thesis favorably if updated disclosures show positive OCF after a full capacity ramp, modest working-capital support and affordable independent financing without stronger supplier recourse. Revisit the broad supplier-demand thesis if material accepted capacity is canceled, delivered utilization misses economic break-even or counterparties need repeated concessions.

Next issuer work: build delivery/financing exposure maps for NVDA/AMD/AVGO; reconcile hyperscaler cash capex with operating and finance leases; quantify memory/network content only from documented configurations; separate software AI upsell from seat displacement. Portfolio action requires dated valuations, existing holdings, active weights and mandate, none of which this report assumes.

Red team: three ways this conclusion can be wrong

Demand may compound faster than fixed commitments. The strongest countercase is already visible: revenue scaled much faster than compute in H1 and Q2 became operating-profitable. If frontier quality creates durable, high-value autonomous work, today’s commitments may secure scarce capacity cheaply. A financing-risk focus could mistake an unusually productive investment cycle for weak economics. Milestone financing can partly match capacity delivery: Amazon, Google and AMD unlock capital when contractual delivery, revenue or deployment tests are met, and management says Broadcom growth tests were significantly below recent actual growth. That is a substantive funding advantage, although conditions, timing, ownership caps and permitted uses prevent treating those amounts as unrestricted cash. The answer is to follow cash collected per delivered capacity cohort, not dismiss growth because it is rapid.

The supply response may be more flexible than headlines imply. Delivery conditions, resale/assignment, termination rights and architectural substitution may reduce cash rigidity. Our constant-use runway stresses and illustrative case costs could overstate the near-term burden. Conversely, delay provisions can start rent before a facility is useful and cross-defaults can make risk greater than a simple schedule suggests. Contract-by-contract underwriting must replace a single aggregate number.

Competition may create much more value for users than for model shareholders. Extraordinary adoption can coexist with falling industry rents. Open models, hyperscaler bundling and competing frontier labs can force faster reinvestment or lower prices. Strong infrastructure earnings may persist for a period even as long-term model-lab cash economics weaken; hardware excess eventually reverses that relationship. Demand growth, industry revenue and shareholder value are related but distinct outcomes.

The report’s conclusion survives those objections only in its disciplined form: Anthropic establishes substantial paid AI demand and improving historical operating economics. It leaves open whether the next infrastructure and financing cycle can convert that demand into durable returns. Wait for the completed terms and cash schedule before underwriting the IPO; evaluate listed suppliers on funded delivery, utilization, cash conversion and price already paid for growth.