Anthropic IPO Filing: Revenue and S-1 Discrepancies

Anthropic’s public S-1 (filed October 26, 2026) makes the audited half-year numbers solid and leaves the numbers investors have been trading—year-end run rate, third-quarter dollars, and the full compute stack—either blank or defined in a way that does not match how headlines use them. The filing is the first frontier-lab prospectus, ticker ANT, and the gaps are mostly mechanical, not evidence of a fabricated document. They are still large enough to misprice the equity if run rate, gross marketplace sales, or “$600–800 billion of commitments” are treated as interchangeable with GAAP revenue.

Run rate is a 28-day snapshot, and the S-1 leaves it blank

The number that has driven private valuations is not recognized revenue. Anthropic defines Run-Rate Revenue in the prospectus as trailing 28-day consumption fees multiplied by 13, plus monthly subscription fees multiplied by 12. That is a four-week annualization of metered usage, not contracted ARR, and the sentence that would give the figure as of October 31, 2026 is empty. The cover-style claim that Claude went from zero at commercial launch to a dollar amount for the quarter ended September 30, 2026 is also blank.

What is filled in is GAAP. First-half 2026 revenue was $16.286 billion, versus $1.163 billion a year earlier. The March quarter was $4.731 billion and the June quarter $11.555 billion—larger than all of 2025 ($4.585 billion). Annualizing the June quarter produces about $46 billion, which lines up with the roughly $47 billion run rate the company cited around its May Series H, not with the roughly $65 billion figure circulated for the end of July. A business that more than doubled from Q1 to Q2 will print a run rate far above trailing recognized revenue; the reverse is also true if a hot four weeks cools. The filing itself warns that revenue is mostly consumption-based and much of it is not under long-term contract, and that growth is not expected to continue at comparable rates.

• Q3 is stated only as a multiple: 86 times fourth-quarter 2024 revenue of $197 million, which is about $16.9 billion. The Financial Times and Bloomberg, citing an October investor meeting, reported $16.8 billion—roughly 45% sequential growth after 144% in Q2.
• Remaining performance obligations were only $10.8 billion at June 30. That is the contracted backlog. It is not the run rate.
• Spend is seasonal: customer spend fell about 20% in the two weeks ended January 2, 2026, versus the prior two weeks. A late-year 28-day window can overstate a full year.

For anyone underwriting a multiple on “$65–70 billion ARR,” the primary check is the amended S-1’s filled-in Run-Rate Revenue and the audited (or reviewed) September quarter, not a private slide from July.

Gross marketplace accounting inflates the top line relative to cash kept—and relative to OpenAI

About 48% of first-half 2026 revenue came through Amazon, Google, and Microsoft cloud marketplaces, up from about 11% in 2023, 32% in 2024, and 47% in 2025. Anthropic says it is the principal: it has the customer contract, sets price, and delivers the model, so it books the full amount and records platform fees in sales, marketing, and partnerships. The platforms bill the customer and remit cash net of those fees. Marketplace-originated balances were 58% of net receivables at June 30.

The public S-1 does not break out the fee rate for 2026. On the confidential draft, Reuters calculated that 2025 marketplace sales were about $2.16 billion and distribution fees about $351 million—roughly 16 cents per marketplace dollar. Applied to a 48% channel mix, that is several billion dollars of H1 “revenue” that never arrives as cash. OpenAI, by contrast, has told investors it records only its share of certain Microsoft-routed sales. The Financial Times reported OpenAI’s September annualized revenue near $50 billion on that net basis, about $20 billion below a roughly $70 billion figure that investors had built by grossing OpenAI up to look like Anthropic. Both treatments can be GAAP-compliant. They are not the same ruler. Comparing the two labs’ headline run rates without a fee bridge overstates Anthropic’s lead and, earlier this month, briefly overstated OpenAI’s.

• Amazon and Google are simultaneously investors, compute suppliers, resellers, and model competitors. A dispute that slows remittance hits cash even if the revenue has already been recognized.
• The S-1 does not quantify H1 platform fees, so a net-revenue restatement is not possible from the public document alone.

Concentration looks better than 2025, but the revenue is still optional and partly non-compliant

No single customer exceeded 10% of revenue in the nine months to September 30, and the top five were about 25%. Customers outside the top five were about 79% of third-quarter revenue, up from 64% in the first quarter. That is an improvement on the confidential-draft picture, in which two unnamed customers each supplied about 12% of 2025 revenue. It is not diversification in the SaaS sense. Eighty percent of H1 revenue was token consumption ($13.1 billion); subscriptions were 20% ($3.2 billion). Large accounts can cut usage without breaking a contract. Net revenue retention of about 900% as of September 30 measures what a qualified cohort spent in the latest three months versus the same three months a year earlier. At this growth rate that mostly says last year’s base was tiny and Claude Code expanded it, not that the book is locked. The metric excludes the fixed Pro and Max plans, whose fees do not expand the same way. Claude Enterprise users who also use Claude Code spend on the order of 20 times users who do not.

A separate quality issue sits inside the growth. Anthropic estimates that 5% to 7% of revenue for the nine months ended September 30 came from customers later found to be violating usage policy or commercial terms—largely distillation, in which other labs train on Claude outputs. That revenue is recognized as tokens are served, before accounts are cut. Detection was tightened in August and September. The company expects the share to fall and says the hit to future revenue and growth is uncertain in both timing and size. On a nine-month base near $33 billion, 5–7% is roughly $1.6–2.3 billion that was real cash and is not a durable customer. Pre-filing investor meetings used the same 5–7% range to explain part of the Q3 deceleration. The S-1 confirms the estimate; it does not name the accounts or split distillation from other violations.

Price cuts and reasoning workloads pull in opposite directions, and the 85% “gross margin” is not the compute bill

Unit prices are falling on purpose. The filing says Opus 5.5 cut cost per task by about 40% versus Opus 5, and that cheaper tasks drove more usage. Fable 5.1 cut typical workloads by about 25% and complex coding and agentic tasks by up to 45%. Models released in the prior six months accounted for more than 80% of third-quarter revenue, so Anthropic cannot sit on an old price list: the frontier release is the revenue. Open-weight models are named as pricing pressure on workloads that care more about cost than intelligence. Forrester’s read of the filing is that this is now a volume business: fixed compute has been bought, and customers have only $10.8 billion of remaining performance obligations against a much larger supplier stack, so buyers have leverage to demand lower task prices.

Reasoning and agents are the volume offset and the margin risk at once. Agentic loops reread context, call tools, and emit “thinking” tokens. That multiplies billable consumption—the mechanism behind Claude Code’s outsized spend—but it also multiplies inference compute per completed task. If price per task falls 25–45% per generation and reasoning tokens rise, serving margin depends on whether cache hits, cheaper chips, and higher utilization outrun the token inflation. The S-1 says they have so far, and that margins held through the recent price cuts. It also says the company expects to buy more short-term compute, at a higher unit cost, when demand exceeds long-term capacity, and that this will compress operating margin in the near term. New capacity is expensed when it comes online, often in large steps, before the revenue it supports. Operating margin is therefore expected to swing “at times meaningfully” even if underlying unit economics improve.

The 85% figure that circulated as a cloud-like gross margin is a functional allocation, not the full compute cost. Cost of revenue was $2.38 billion in the first half—substantially all customer-serving compute—on $16.286 billion of revenue, or about 85%, up from about 12% a year earlier. Total compute and infrastructure expense was $8.945 billion, or 55% of revenue (48% in the June quarter, down from 251% in Q1 2025). Roughly $6.6 billion of compute sits in research and development (training, research, internal Claude use) and in sales (free-tier serving), not in cost of revenue. H1 GAAP operating loss was $693 million; adjusted operating income was positive; the June quarter was the first GAAP operating profit at $974 million. An 85% gross margin does not mean 85 cents of every revenue dollar is available after the chips are paid for.

“The compute commitment” is several different numbers, and the company’s own total is blank

The prospectus states that, as of and after September 30, 2026, purchase commitments and leases—including short-term deals and leases not yet commenced—have an aggregate committed spend of “more than $____,” substantially all compute, with most of the remainder expected between 2031 and 2045. The dollar is not filled in. Everything else in the press is a reconstruction.

Named deals that analysts have added up come to roughly $637 billion: Broadcom equipment leases about $161.2 billion (largely non-cancellable), SpaceXAI up to $120.5 billion (cancelable on 90 days’ notice after early 2027), Google at least $114.5 billion, Amazon at least $110 billion, Microsoft at least $44.9 billion, Nscale up to about $44.6 billion, Lambda about $41.3 billion. Of that, on the order of $431 billion is minimum spend or lease rent owed largely regardless of usage; about $206 billion is “up to” or billed as capacity is accepted. A further $125–150 billion of TPU capacity for 2028–2031 is estimated and not finalized, and AMD hardware is expected to exceed $20 billion if fully taken. Other published totals measure different things: about $265.6 billion of fixed or minimum-guaranteed commitments at June 30; about $417 billion of future minimum obligations on agreements entered or amended since year-end 2025; Reuters’ pre-filing $518 billion; Wall Street Journal draft ranges around $450 billion of cloud deals and more than $600 billion overall; Fortune’s $657 billion with about $498 billion called non-cancellable; some secondary write-ups at $750–810 billion once estimates are stacked on top of caps.

The economic point does not depend on which total wins. Revenue is optional and monthly; a large core of Google, Amazon, and Microsoft spend is take-or-pay. Cash and short-term investments were $131.7 billion at June 30 ($112.4 billion cash plus $19.3 billion of investments), against obligations that, on any of the larger readings, run for most of a decade. H1 operating cash flow was positive at $3.9 billion. That does not fund a $400 billion-plus minimum stack. The $80.7 billion first-half net loss is mostly not cash either: about $72.4 billion is the mark-to-market on convertible notes, derivatives, and SAFEs as the private valuation rose. Treating that loss as burn, or treating every “up to” cap as a debt-like obligation, are both wrong.

What outlets got wrong, and what was never a company claim

Discrepancies are real; a fabricated S-1 is not what the record shows. The document is on EDGAR (accession 0001628280-26-067168, file 333-299633). Blanks for price range, share count, September revenue, October run rate, and the aggregate commitment total are ordinary in an initial S-1 filed weeks after quarter-end. An accounting professor told WIRED there are practical reasons the September close is not in the first public draft. Gary Marcus’s claim that blanking Q3 means “demand has dropped” and that the company is skirting the truth overstates the omission: the filing does give the 86-times multiple, and investor-meeting reporting of $16.8 billion is directionally the same number.

The unreliable layer is secondary arithmetic and pre-filing leaks that keep circulating as if they were the prospectus:

• Draft Microsoft commitment figures near $31 billion became at least $44.9 billion in the public filing. Draft “two customers are a fifth of revenue” does not match the nine-month concentration in the S-1.
• Chinese-language recaps have multiplied dollar figures by ten (H1 revenue rendered as $162.86 billion, the net loss as $807 billion). Those are unit errors, not new disclosures.
• OpenAI’s “$70 billion” was an investor gross-up using Anthropic’s method, not an OpenAI print. The company’s own figure to investors was nearer $50 billion net.
• Sacra’s “~$70 billion by October, growth almost stalled” and SemiAnalysis’s “~$74 billion annualized from the Q3 print” are outside estimates. The S-1 does not contain an October run rate.
• Social posts that treat the $80.7 billion loss as cash burn, or the 85% margin as profit after all compute, are misreading lines the filing separates.

Claims least reliable, and the evidence that would settle them

1. Any specific October or year-end run rate ($65 billion, $70 billion, $74 billion, $100–110 billion). Least reliable because the defined metric is blank and prior prints are point-in-time, gross, and uncontracted. Resolve with the amended S-1’s Run-Rate Revenue as of the stated measurement date, plus the monthly consumption bridge, and the reviewed Q3 and Q4 income statements.

2. A single compute-commitment total ($417 billion, $518 billion, $637 billion, $750–810 billion). These add different populations (minimums vs caps vs unsigned estimates vs post-balance-sheet deals). Resolve with the filled-in “aggregate committed spend of more than $____” and the contractual-obligations table, split into non-cancellable minimums, usage-based, and cancelable (especially SpaceXAI after the 2027 notice windows).

3. “85% gross margin” as the economic margin on serving. It excludes most of the compute bill by allocation. Resolve with a quarterly split of compute and infrastructure among cost of revenue, R&D, and sales, and with revenue per dollar of inference compute versus training compute—neither of which the first S-1 fully gives.

4. Apples-to-apples lead over OpenAI on run rate. Anthropic is gross on a channel that was 48% of H1 revenue; OpenAI has described itself as net on Microsoft-routed sales. Resolve with a quantified marketplace-fee line (H1 2026 platform fees, not the 2025 Reuters estimate) and, eventually, OpenAI’s own S-1 principal-versus-agent footnote.

5. 900% net revenue retention as proof of locked, high-quality expansion. The cohort math on a tiny year-ago base, the exclusion of fixed consumer plans, and the 5–7% policy-violator slice all inflate the quality read. Resolve with NRR ex-distillation and ex-top-five, the dollar revenue subsequently reversed or foregone after the August–September enforcement, and the share of consumption under minimum-commit contracts versus pure pay-as-you-go.

6. Q3 as either a collapse or a clean $16.8 billion print. The multiple and the investor-meeting figure agree within rounding; neither is a reviewed statement, and adjusted profit for that quarter was described to investors without a published reconciliation. Resolve with the amended financial statements for the quarter ended September 30, including the GAAP-to-adjusted bridge and whether marketplace fees and violator revenue are called out.

7. Claims that the S-1 was fabricated, secretly rewrote Q3, or hid a demand crash by leaving blanks. The filing is public, the Q3 multiple is in the business section, and blank forward items are standard. Resolve by comparing the October 26 S-1 to the next amendment on EDGAR, not by treating an empty field as a negative disclosure.

---

Recent Findings Supplement (October 2026)

Anthropic filed its S-1 on or around October 26, 2026, disclosing H1 2026 (ended June 30) results that clarify several headline distortions while introducing new specifics on accounting treatment, commitments, and risks.[1][2]

The filing shows $16.286 billion in revenue for the period (up ~1,300% YoY from $1.163 billion), with Q2 alone at ~$11.555 billion; full-year 2025 revenue was $4.585 billion. A net loss of $80.7 billion included a $72.4 billion non-cash charge tied to fair-value adjustments on convertible notes, derivatives, and SAFEs as the company’s valuation rose. Operating results improved markedly, with Q2 delivering the first reported operating profit of $974 million (H1 operating loss narrowed to $693 million).[3][4]

Marketplace sales (primarily AWS Bedrock, Google Vertex AI, and Azure) accounted for 48% of H1 2026 revenue (up from 47% in 2025, 32% in 2024, and 11% in 2023), with 58% of receivables also flowing through these channels.[1][5]

Anthropic acts as the principal in these arrangements, maintaining direct customer contracts, control over the service, and pricing authority. It therefore recognizes the full customer payment as gross revenue; cloud partners bill, collect, and remit net of their fees (treated as a sales/marketing expense). Channel partner fees drove most of the rise in sales, marketing, and partnerships expense to $3.168 billion in H1 2026 (including a $2.7 billion increase).[6][1]

• This gross recognition contrasts with reports that OpenAI books marketplace sales net of partner cuts, making direct topline comparisons misleading.[7]
• Consumption-based (token-metered) revenue comprised ~80% of H1 2026 totals ($13.1 billion), with subscriptions at ~20%.[2]
• Net revenue retention reached ~900% as of September 30, 2026 (enterprise customers spending ~9x year-over-year on average).[3]

This structure inflates reported revenue relative to net-of-fees peers and highlights dependency on a small number of cloud partners that simultaneously supply critical (and rationed) compute capacity, creating potential conflicts.[8]

For competitors or new entrants, the gross-book treatment rewards scale through marketplaces but exposes margins to rising channel fees (already pressuring sales expenses to 19% of revenue). Success requires either bypassing these channels or negotiating better terms while matching Anthropic’s consumption-based model.

Customer concentration has moderated but remains a noted risk, with top-five customers accounting for ~25% of revenue in the first nine months of 2026 (no single customer >10%); many large accounts lack long-term contracts.[1][2]

The customer base expanded rapidly: >11,900 customers with ≥$100k trailing-twelve-month (TTM) revenue as of September 30, 2026 (vs. >950 a year earlier); >1,900 with ≥$1 million; and ~250 with ≥$10 million. U.S. billing addresses accounted for ~72% of H1 revenue.[3][2]

• Revenue from customers outside the top five rose to ~79% in Q3 2026 (from 64% in Q1).[3]
• The filing flags that top customers can reduce or discontinue spending at any time.

Broadening adoption reduces single-customer risk compared with earlier periods, but the high NRR and consumption model mean revenue remains sensitive to usage patterns among a still-concentrated set of large enterprise and partner-driven accounts.[2]

Entrants must demonstrate sticky, diversified usage to compete on valuation multiples; heavy reliance on a few hyperscalers or large logos invites the same concentration scrutiny.

Compute and infrastructure commitments have ballooned, with non-cancelable or minimum obligations reaching hundreds of billions—far outpacing current revenue and creating long-term margin and liquidity pressure.[9]

At June 30, 2026, the filing disclosed $265.6 billion in future purchase commitments and $321.4 billion in leases not yet commenced. Post-June 30 additions totaled ~$245 billion (some conditional); earlier aggregates referenced ~$417 billion in minimum obligations from late 2025 onward, with further estimates pushing totals toward $542–567 billion or higher when including Broadcom (~$161 billion), Google (~$114.5 billion+), AWS (~$110 billion+), Microsoft (~$44.9 billion+), and others. Many contracts require payment “regardless of actual usage.”[10][11]

Compute/infrastructure expense fell sharply as a percentage of revenue (from 201% in H1 2025 to 55% in H1 2026, or ~48% in Q2), enabling the Q2 operating profit despite absolute spend rising to ~$8.9–9 billion for the half.[2][9]

• The S-1 leaves some aggregate post-September 30 figures blank or estimated, and sources vary in how they sum floors vs. caps or post-period amendments.[11]
• Partners have rationed capacity below requested levels in some cases.[8]

These fixed or minimum commitments lock in massive future cash outflows (potentially >$800 billion across disclosed and estimated deals) while revenue depends on filling that capacity with usage.[12]

New entrants face a high barrier: either secure comparable long-term deals (risking overcommitment) or rely on spot/short-term capacity at higher unit costs, compressing margins until scale is achieved.

Recent model releases delivered substantial cost-per-task reductions (up to 45% for certain workloads with September 2026 updates) while the company claims margins were preserved, amid broader price competition from Chinese rivals and OpenAI discounts.[13]

Open-weight and lower-cost models (e.g., from Moonshot, DeepSeek) are cited as sources of pricing pressure; Anthropic notes it may need to adjust pricing or shift to task-based models. Token prices have declined ~50% year-over-year in some datasets due to efficiency gains, cache hits, and competitive cuts.[14]

• Gross margins (calculated) reached ~85% in H1 2026 vs. much lower prior periods.[15]
• Channel fees and rising absolute compute spend still create margin headwinds even as percentages improve.

Efficiency gains allow price cuts without immediate margin collapse, but sustained competition risks eroding the premium pricing that supports high gross margins and the ability to service compute commitments.[16]

Competitors can gain share by undercutting on cost-per-task or offering deeper discounts, forcing Anthropic (and peers) to accelerate efficiency or accept lower margins to retain volume.

Headline net losses are dominated by non-cash valuation adjustments rather than operations, and some pre-filing draft figures or extrapolations created minor discrepancies across outlets; no evidence of fabricated S-1 claims emerged.[4]

The $80.7 billion H1 net loss was driven overwhelmingly by the $72.4 billion fair-value charge. Earlier leaked or draft reports referenced different 2025 loss figures (~$42 billion net vs. ~$8 billion operating) or customer concentration shares. The public S-1 includes blanks for certain Q3 or aggregate figures and leaves some commitment reconciliations to footnotes or estimates.[17][18]

• Revenue figures (~$16.3 billion H1) have been consistent across WSJ, Reuters, Bloomberg, and direct analyses.[19]
• Run-rate extrapolations (e.g., from Q2 or NRR) appear in commentary but are caveated by the consumption-based, usage-dependent nature of revenue.

The least reliable claims (based on available post-April 2026 reporting) are:

• Precise aggregate compute commitment totals (e.g., “$500B+” or “~$800B”) — vary by inclusion of estimates, post-period adds, and conditional vs. firm obligations; primary evidence needed is the full detailed schedule and any amended S-1 reconciliations.
• Direct revenue or margin comparisons to OpenAI (or peers) without adjusting for gross-vs-net marketplace treatment — distorts scale; primary evidence is side-by-side audited revenue recognition policies.
• Run-rate or annualized extrapolations implying sustained hyper-growth without usage caveats — sensitive to customer concentration and competition; primary evidence is full-year 2026 audited results and Q3/Q4 usage trends.
• Pre-filing draft specifics on customer concentration or losses — superseded by the public S-1; primary evidence is the final filed document and any SEC comments/amendments.

These elements underscore that while operational metrics show rapid scaling and improving efficiency, structural dependencies on cloud partners, massive fixed commitments, and competitive pricing dynamics introduce material risks not fully captured in headline revenue or growth rates. Updated filings or earnings will be key to resolving remaining gaps.