S-1 financial statements run to 30 June 2026; customer metrics are as of 30 September 2026. Q3 revenue is not in the filing.
The audited numbers are better than the draft coverage suggested. Q2 2026 produced $1.0B of operating income at an 8% margin, and $2.2B of adjusted operating income at 19%. [9] H1 operating cash flow was $3,929M. [10]
Concentration is lower than reported. No single customer exceeded 10% of revenue for the nine months to September, and the top five were about 25%. [11] The draft-era report of two customers at one-fifth of H1 revenue does not appear in that form. [12]
The commitments are firmer in some places and looser in others. Microsoft is at least $44.9B, not $31B, and Google is at least $114.5B. [13][14][15] SpaceXAI becomes cancelable on 90 days' notice after early 2027. [16]
Anthropic is a Delaware public benefit corporation whose stated purpose is to responsibly develop and maintain advanced AI for the long-term benefit of humanity. [17] It sells Claude on a consumption basis, which was $13,064M of H1 2026 revenue, and on subscription, which was $3,222M. [18][19] It has applied to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. [20]
The investor question is whether revenue that went from $386M in FY2024 to $11.6B in Q2 2026 alone can stay ahead of a compute stack contracted out to 2045. [21][7][22] The filing shows $265.6B of fixed or minimum-guaranteed commitments at June, plus estimated Google TPU payments of $125-150B for 2028-2031 that sit outside the commitments table. [5][23] Against that, the company holds $131.7B of cash and short-term investments. [4]
The falsifiable near-term KPI is the Q4 exit rate. Investors expect $110B of annualized revenue by year-end, per The Wall Street Journal. [24] The filing defines run-rate revenue but leaves the 31 October figure blank. [25]
- Gross margin, derived from the S-1, went from 12.4% in H1 2025 to 85.4% in H1 2026. [26][27]
- Net revenue retention was about 900% at 30 September. [28]
- More than 1,900 customers spend at least $1M a year, and about 250 spend at least $10M. [29]
- GitLab's AI partnership manager calls Anthropic its largest AI vendor by spend. [30]
- The S-1 warns agentic workloads are more compute-intensive, which could make profitable pricing difficult. [31]
- OpenAI's 20-40% discounts lifted its share of Rippling's coding work from 16% to 52%. [32]
- A divided D.C. Circuit denied Anthropic's petition on the Department of War designation on 25 September. [33]
- The Founder LLC holds 50.1% of votes, and Anthropic will be a controlled company. [34][35]
The liquidity path is IPO-track and dated. Bloomberg reported marketing starts 9 November. [36] Fox Business's Liz Claman says a 19 or 20 November pricing remains on track, and the FT reports investors expect a $1.5-1.75T valuation. [37] The S-1 leaves share count and price range blank, and includes a selling-stockholder component. [38][39] The gate that would flip the read is a price range implying less than the $965B Series H post-money. [40]
- Price range, share count and run-rate revenue in the amended S-1. [25]
- Q3 in audited form; the September-quarter revenue line is blank in this filing. [41]
- Lock-ups: 180 days for most insiders and 365 days for founder-related holders, with an early release tied to the Q1 earnings release. [42]
- Department of War litigation; a rehearing petition was filed 13 October. [43]
- Public Benefit Fund milestones, which start from the IPO price. [44]
Anthropic monetizes model capability in three ways: metered tokens through its API and cloud marketplaces, seat-plus-usage enterprise subscriptions, and agentic products led by Claude Code. The former head of competitive intelligence at OpenAI describes Anthropic Enterprise as seat fees combined with metered usage. [45] Claude is available on the Claude Platform, Google Cloud's Vertex AI, Microsoft Foundry and Amazon Bedrock. [46]
The company was founded by siblings Dario and Daniela Amodei, and raised a $124M Series A in May 2021 led by Skype co-founder Jaan Tallinn. [47] It is headquartered in San Francisco. [48] Per the S-1, a single Class F share gives the Founder LLC 50.1% of total voting power, recalculated to hold that percentage as other voting shares change. [34] The company remains a Delaware public benefit corporation. [17]
Management's track record is two-sided. Execution on revenue is visible in the S-1: Q2 2026 revenue of $11,555M compares with $787M a year earlier. [7] Headcount rose about 170% in the year to September while TTM revenue rose about 1,200%. [49] The governance record is less settled. The S-1 says the Department of War designation followed Anthropic's public opposition to use of its technology for mass domestic surveillance or fully autonomous weapons. [50] Retention of builders is a live question after two core Claude Code engineers left on the same day in late October. [51] OpenAI CEO Sam Altman called those departures "not a good sign." [52]
- Frontier models
- Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1 sit in the current lineup offered through the cyber access tiers. [53] Opus 5.5 is the first model in the 5.5 family and targets long-running coding and knowledge work. [54]
- Claude Code
- Runs on developer machines, reads and modifies source code, and calls tools and MCP servers under a developer's credentials. [55]
- Claude Enterprise and Team
- Compliance and Analytics APIs expose session transcripts, tool calls and per-user usage to governance teams. [56]
- Claude Cowork
- A health system cut administrative workload by 20% using Cowork and Claude Code, per partner Intuitive.ai. [57]
- Security programs
- Project Glasswing partners found at least 129,000 verified software vulnerabilities between April and July 2026 using Mythos models. [58]
The S-1 splits revenue only into consumption-based ($13,064M in H1 2026) and subscription-based ($3,222M); Claude Code and other products are not broken out. [18][19][59]
Expert-call rows reflect individual practitioners' accounts, not company disclosures.
Concentration is moderate and falling. For the nine months to 30 September 2026, no single customer exceeded 10% of revenue and the top five made up about 25%. [11] In H1 2026 the largest disclosed customer, Customer C, was 11%. [69] In Q3, about 79% of revenue came from customers outside the top five. [70] The filing does not name the top five or say whether they are cloud providers. [71]
Expansion inside accounts is the strongest signal. The S-1 reports more than 11,900 customers with at least $100,000 of TTM revenue at 30 September 2026, up from more than 950 a year earlier. [6] More than 1,900 customers exceed $1M and about 250 exceed $10M. [29] Net revenue retention was about 900%. [28] JPMorgan's executive director reports overall AI usage growing more than 5% month over month, with Claude Code rollout still limited while controls are built. [72]
The ledger also shows where control of the workflow is slipping to intermediaries. At Microsoft, the move to Copilot CLI routes Claude underneath rather than abandoning it. [73] Adobe's FinOps manager says engineers switch between Anthropic and OpenAI on tooling, negotiated discounts and observed value. [66] Both accounts point to Claude winning the hardest workloads while losing direct ownership of the developer seat.
Gross margin and free cash flow are calculated from the reported lines. The S-1 runs to 30 June 2026.
Quarterly revenue per the S-1 was $376M, $787M, $1,382M and $2,040M through 2025, then $4,731M in Q1 2026 and $11,555M in Q2 2026. [7] Q2 2026 is the first profitable quarter: operating income of $1.0B, an 8% operating margin, and adjusted operating income of $2.2B at a 19% adjusted margin. [9]
Q3 is not in the filing; the S-1 leaves the September-quarter revenue figure blank. [41] The $16.8B Q3 figure comes from the October investor day and remains unaudited. [8] It is up 45% QoQ from Q2, after roughly 147% in Q2. [84]
The $80.7B H1 2026 net loss is overwhelmingly a $72.4B fair-value charge. [82][81] Anthropic elected fair-value accounting for its convertible notes, so changes in their estimated value run through the income statement. [85] Convertible notes were carried at $108.6B on the balance sheet at 30 June. [86]
Strip the charge and the net loss is about $8.4B. That is below the ~$8.7B the press estimated from the draft. [8] The operating loss narrowed to $693M in H1 2026 from $2,135M in H1 2025 on 14x the revenue. [2]
The adjusted metric now has a definition. Adjusted operating income excludes stock-based compensation, non-cash charitable contributions and the Bartz settlement. [87] On that basis H1 2026 adjusted operating income was $1.618B, a 10% margin. [88]
Cash generation turned positive in the S-1 period. Operating cash flow was $3,929M in H1 2026 against capex of $482M. [10][3] That makes H1 free cash flow about $3.4B, which contradicts the "deeply negative" read given before the filing was public. [89]
The balance sheet holds $112.4B of cash and $19.3B of short-term investments, a combined $131.7B. [4] Debt is $6.2B of Amazon notes principal. [90] Separately, a Broadcom note-purchase facility of up to $42B exists solely to pay TPU lease fees, at 9% interest. [91]
Marketplace revenue is booked gross. Anthropic says it is principal on cloud-marketplace sales and recognizes them on a gross basis. [92] Those sales were about 48% of H1 2026 revenue, up from 32% in 2024. [93] They were also 58% of net receivables at 30 June. [94] The marketplace fees that CSPs deduct are not separately quantified, so a net-revenue figure is unavailable.
- Google: at least ~$114.5B from April 2026 to July 2033, regardless of usage, non-cancelable. [14] A further ~$125-150B is estimated for TPU capacity in 2028-2031 and is not in the commitments table. [23]
- AWS: at least $110B from May 2026 to April 2036, regardless of usage. [95]
- Microsoft: at least $44.9B from November 2026 to May 2033; undelivered tranches can be cancelled. [13]
- Broadcom: five-year equipment leases totalling ~$161.2B of rent, excluding interest. [96]
- SpaceXAI: up to $120.5B into 2029 at ~$3.1B a month once delivered; cancelable on 90 days' notice after early 2027. [16]
- Nscale and Lambda: up to ~$44.6B and ~$41.3B of NVIDIA capacity, paid only as it is delivered and accepted. [97][98]
- AMD: no fixed amount; above $20B if the full capacity is procured. [99]
The filing gives three aggregates that measure different things, and none is the $450B figure reported from the draft. [100] Fixed or minimum-guaranteed commitments were about $265.6B at 30 June 2026, excluding compute deals with no minimum. [5] New or amended agreements signed after December 2025 carry about $417B of future minimum payments. [101] Leases not yet commenced account for about $321.4B of undiscounted obligations. [102]
The scale-up happened in 2026. At 31 December 2025, non-cancelable purchase commitments totalled $54.6B. [103] The S-1 says substantially all commitments relate to compute and infrastructure, with most of the post-2030 tail falling in 2031-2045. [22]
Against revenue, the $265.6B contracted floor is about 5.7 years of Q2 2026 revenue annualized ($11,555M x4). [7] On the unaudited Q3 run-rate of $67.2B it is about 4.0 years. [104] Flexibility is greater than the draft coverage implied: SpaceXAI becomes cancelable on notice from 2027, and Nscale and Lambda pay only on delivery. [16][97]
The filing frames compute as the binding constraint. Anthropic says it has operated in a compute-constrained environment, relies on a limited number of suppliers, and may need more expensive short-term compute during shortages. [105] It names a conflict at the core of the supply chain: Amazon, Google and Microsoft supply substantial compute while also offering competing models. [106]
Owned capacity is arriving through partners. Theseus Infrastructure, formed with Macquarie Asset Management and GIC, will develop sites with Anthropic as anchor tenant. [107] Fluidstack provides data-center development and operations under a colocation agreement, with no dollar figure disclosed. [108] AMD's first gigawatt of MI450-series GPUs starts deploying in H1 2027. [67]
Productivity inside the company is the offset to compute intensity. Anthropic had about 5,300 employees at 30 September 2026, and TTM revenue per employee rose from $1.4M to over $6.6M. [49] In September 2026 Claude wrote over 90% of contributions to Anthropic's own codebase. [109]
Distribution is split almost evenly between direct and cloud marketplaces. CSP marketplace sales rose from about 11% of revenue in 2023 to 47% in 2025 and 48% in H1 2026. [71] GitLab's partnership manager says usage across the first-party API, Vertex and Bedrock can count toward both Anthropic commitments and cloud burndowns, which has repeatedly helped Anthropic win business. [110] Sales and marketing spend rose to $3,808M in H1 2026 from $239M a year earlier. [79]
The services layer is being built to move pilots into production. Anthropic and Hg plan a European consultancy launching by year-end, following a similar $1.5B US initiative. [111] Claude Frontier Academy is a $100M programme to train 10,000 engineers. [46] Block's commercial executive still describes Anthropic as less mature than OpenAI or Google in standardized packaging and procurement. [112]
Each cell is one practitioner's account from a dated expert call; spend shares are proxies, not disclosures.
Anthropic holds the high-capability end of enterprise coding, and the price competition is aimed at that end. The former head of competitive intelligence at OpenAI cites enterprise spending data showing Anthropic overtaking OpenAI among paying businesses, followed by signs of OpenAI recovering ground. [126]
OpenAI is buying share back with price. It is offering 20-40% off list for enterprise models including Codex, and Replit and DataCamp received 30-40% offers and switched primary models. [32]
Anthropic has cut prices too. A review of June-October pricing describes aggressive cuts by both labs, driven by low switching costs and open-weight competition. [127]
No defensible TAM figure appears in the S-1 coverage or company releases reviewed, so a top-down market size is unavailable. The bottom-up evidence comes from spend shares. In Block's commercial executive's peer proxy, OpenAI holds over 50% of model spend, Anthropic is close behind, and Anthropic may capture the largest share of marginal new spend. [122] Anthropic's Q3 annualized revenue of $67.2B is the observable floor for its own serviceable demand. [104]
Pricing power is eroding at the list level and holding at the workload level. The managed-services director at a North American advisory firm estimates negotiated discounts are down about 10% when customers can commit volume or compare platforms. [128] The same expert estimates comparable completed workflows cost about 15% less through optimization rather than vendor cuts. [128]
Model efficiency is the offset. In Kiro's internal benchmarking, Opus 5.5 completes tasks with about 40% fewer tool calls and half the tokens of Opus 5, and its credit multiplier fell to 2.0x from 2.2x. [129] The former OpenAI competitive-intelligence head argues a more token-intensive Claude task can still be cheaper if it finishes correctly with less human intervention. [45]
The orchestration layer is moving to customers and intermediaries. A Microsoft principal engineer says the shift to enterprise wrappers is chiefly economic: routing, context management, token accounting and budget controls let simple requests go to smaller models while complex work stays on Claude or GPT-class models. [130] Block's executive notes sophisticated buyers increasingly use routing and abstraction layers. [131]
Open-weight models are closing the gap from below. A Cohere technical-staff member estimates the best open-weight models are roughly one generation behind leading closed models. [132] Moonshot AI's Kimi K3.1 Pro rivals leading US frontier models on some benchmarks. [133] Adobe is testing Qwen for basic coding tasks and agents because of its low cost. [115]
The counterweight is that on-prem models are neither cheap nor easy to maintain; they require skilled engineers and continuing infrastructure support. [134] The near-term exposure is therefore concentrated in high-volume, lower-complexity tokens rather than in the frontier coding work that carries Anthropic's premium.
Public shareholders will buy economics with limited control. A single Class F share gives the Founder LLC 50.1% of total voting power, recalculated to preserve that share as other voting stock changes. [34] The Class F share sunsets 450 days after the Founder LLC has had fewer than three qualifying members for 90 consecutive days. [135] Anthropic will be a controlled company under Nasdaq rules. [35]
A second layer sits with non-shareholders. The Long-Term Benefit Trust holds the Class T share and can generally elect and remove a bare majority of the board. [136] It is governed by five voting trustees. [137] Harvard astrophysicist Christopher Stubbs was named to the trust on 21 October. [138] Employees other than the co-founders hold Class E shares, which carry limited voting rights and no economic rights. [139] Class N shares are non-voting and convert to Class A on a qualifying transfer. [140]
The third layer is the Anthropic Public Benefit Fund. Anthropic has already contributed 846,840,960 Class N shares to it, equal to 5% of fully diluted shares at 30 June assuming full conversion of the Amazon notes. [141] After the IPO the board may issue up to a further 2% of outstanding shares each time the share price rises 50% above the prior milestone, starting from the IPO price. [44] A milestone requires closes above the threshold on 90 of the trailing 120 trading days, and the filing states no separate stockholder approval for those issuances. [142] For a public buyer, this escalator is dilution that kicks in at +50%, +125% and so on from the offer price.
The risk factors centre on four exposures. Anthropic says it has operated in a compute-constrained environment and may need more expensive short-term compute. [105] It warns that competition could pressure prices and that agentic workloads make profitable pricing harder. [31] It flags that Amazon, Google and Microsoft supply its compute while offering competing models. [106] It also warns that models could pose catastrophic or existential risks if safeguards prove insufficient. [143]
On litigation, Anthropic entered a $1.5B settlement in the Bartz copyright case in 2025 and warns adverse fair-use rulings could force injunctions or retraining. [144] On the Department of War designation, a district court ruled in Anthropic's favour in August 2026, but a divided D.C. Circuit denied its petition on 25 September. [33]
Use of proceeds is general corporate purposes, including compute and infrastructure capex, hiring, possible acquisitions and RSU tax withholding. [145] Anthropic receives nothing from shares sold by selling stockholders. [39] Twenty-two underwriters are named, led by Goldman Sachs and Morgan Stanley. [146]
Severity is an editorial ranking of exposure size and proximity, not a probability.
Consensus belief. The market had priced Anthropic for a $100B+ year-end run-rate, with some investors projecting $120B. [84] What the evidence shows. The S-1 shows an 85.4% derived gross margin, positive operating cash flow and a first profitable quarter in Q2 2026. [27][10][9] The contested point. Whether that margin survives $265.6B of fixed commitments, plus up-to contracts and $125-150B of expected Google TPU payments, as they reach full delivery in 2027-2031. [5][23] The signal that changes the read. The Q4 exit rate against the $110B investors expect, and gross margin as SpaceXAI, Nscale and Lambda capacity comes online. [24]
Anthropic S-1 (26 Oct 2026). H1 2026 operating cash flow was $3,929M against capex of $482M. [10][3] Q2 operating income was $1.0B at an 8% margin, with adjusted operating income of $2.2B at 19%. [9] This is the audited version of the bull case. Cost of revenue was only $2,380M on $16,286M of H1 revenue. [27]
Anthropic, via the Financial Times (15 Oct 2026). "Cash-flow positive, generating $1.70 in revenue for every dollar spent on computing power during the third quarter." [24] That extends the H1 cash generation into Q3 on management's definition.
Memeburn (22 Oct 2026). Its ratios show compute spend falling from about $2 per $1 of revenue to about 55 cents. [157] This independent recalculation matches the direction of the company figure. The Arora Report (15 Oct 2026). Calls 45% QoQ growth extraordinary at this scale. [84] On the bull reading, the $131.7B cash pile covers half of the fixed commitments on its own. [4][5]
Aswath Damodaran, NYU Stern (23 Oct 2026). He called the adjusted Q3 profit "often a fictional number" and said free cash flow remains deeply negative. [89] He argues $1.5T hinges on 30% margins and that the IPO will be priced on market dynamics rather than intrinsic value. [158] The S-1's positive H1 free cash flow cuts against his cash point. His margin point stands: H1 adjusted operating margin was 10%. [88]
Anthropic S-1 risk factors (26 Oct 2026). The company itself warns that agentic workloads are more compute-intensive, which could make profitable pricing difficult. [31] It says shortages may force it into more expensive short-term compute. [105] R&D, which carries training compute, was $9,741M in H1 2026, 60% of revenue. [78]
Leaked-S-1 reporting via RTTNews (21 Oct 2026). Anthropic is struggling to secure enough capacity from AWS, Google and Microsoft as they come to see it as a competitor. [159] Jason Lemkin, 20VC (22 Oct 2026). "Nobody 10xes forever." [160] On the bear reading, growth halving to 45% QoQ while commitments run to 2045 is how a mismatch begins. [84][22]
What it is. The question is whether contracted compute is a capacity asset or a fixed liability. The S-1 shows $265.6B of fixed or minimum-guaranteed commitments at 30 June 2026. [5] Contracts signed since December 2025 add about $417B of future minimum payments. [101]
Role. Compute is both the product input and the largest cost. It sits in cost of revenue for inference and in R&D for training. H1 2026 R&D was $9,741M against cost of revenue of $2,380M. [78][27]
Exposure. Named counterparty maximums total about $657B: Broadcom $161.2B, SpaceXAI $120.5B, Google $114.5B, AWS $110B, Microsoft $44.9B, Nscale $44.6B, Lambda $41.3B and AMD above $20B. [96][16][14][95][13][97][98][99] Google TPU capacity for 2028-2031 adds an estimated $125-150B. [23]
Strength and risk. The strength is flexibility in part of the stack: Nscale and Lambda bill only on delivery, and SpaceXAI is cancelable on notice from 2027. [97][16] The risk is the take-or-pay core: Google, AWS and Microsoft are owed regardless of usage. [14][95][13]
The numbers. At $1.70 of revenue per compute dollar, covering the $265.6B floor requires about $452B of cumulative revenue. [24][5] Over seven years that is about $65B a year, slightly below the $67.2B Q3 run-rate. [104] Covering the ~$657B of counterparty maximums requires about $1.1T of revenue, or about $160B a year over seven years.
The contested question. Whether the ratio holds once the newer capacity arrives. The floor is already covered at today's run-rate. The maximums need revenue to more than double again. The freshest evidence, positive H1 cash flow, leans constructive on the floor and unresolved on the maximums. [10]
Both sides of this debate are drawn from enterprise buyers on dated expert calls. Each voice is one practitioner, so the panel shows direction rather than market share.
AI partnership manager, GitLab (29 Jul 2026). "The flexibility that we've had with Anthropic, I would say, is probably the one reason why they have continuously won our business." [161] Anthropic is GitLab's largest AI vendor by spend, ahead of OpenAI and Google, within an eight-figure budget. [30] Cheaper, more capable models led GitLab to use more, not spend less. [116]
AI transformation executive, Hanger (25 Oct 2026). "Swapping would in fact be a project on its own." [162] Claude replaced ChatGPT Enterprise and delivered a ~40% cut in approval-cycle time and 50%+ more cases closed per day by senior clinicians. [63]
Former head of competitive intelligence, OpenAI (14 Oct 2026). "Workflow-level switching costs are much higher." [163] Replacing an agent wired into repositories, testing and deployment is far harder than swapping models. [123] Director of AI product security, HP (30 Jul 2026). "Claude is very hyperspecific and really well-loved by our engineering teams." [164]
OpenAI pricing campaign, per IT-trade reporting (26 Oct 2026). OpenAI is offering 20-40% off list for enterprise models including Codex. [32] At Rippling, OpenAI's share of coding work rose from 16% in May to 52%, and Replit and DataCamp switched primary models. [32]
Commercial executive, Block (29 Jul 2026). Claude Code uses about four times as many tokens per task as Codex on the same work, making real-world cost high and hard to predict. [114] Internal estimates rose three to five times when usage or plans changed. [165]
Principal engineer, Microsoft (8 Sep 2026). "The biggest issue with the customer that they are complaining is the token economics." [166] Microsoft cancelled most internal Claude Code licences for Copilot CLI. [65] Executive director, JPMorgan Chase (8 Oct 2026). "For 90% of the tasks, you really don't need Opus." [167] FinOps manager, Adobe (2 Sep 2026). Adobe is moving to model-agnostic tools partly because Claude Code is less flexible for other models. [66]
Every voice on both sides agrees Claude leads on the hardest coding work. Even Adobe's FinOps manager says "the Anthropic models still hold a little bit of an edge" for coding. [168] The split is over who captures the value. GitLab and Hanger describe seat-level lock-in through workflow integration. [110][169] Microsoft, JPMorgan and Adobe describe the routing layer moving in-house, which keeps Claude for complex tasks while pushing the 90% of simpler tokens to cheaper models. [130][170][171] If that pattern spreads, revenue per enterprise account grows more slowly than usage.
- How fast does gross margin compress as the SpaceXAI, Nscale, Lambda and Broadcom capacity reaches service? H1 cost of revenue was $2,380M. [27]
- Who are the top five customers behind about 25% of revenue, and are any of them cloud providers? The filing does not say. [71]
- How large are the marketplace fees netted out of the 48% of revenue booked gross through CSPs? They are not quantified. [92][93]
- What does run-rate revenue at 31 October show? The figure is blank in this filing. [25]
- How many shares sit in the fully diluted count once convertibles carried at $108.6B convert? [86]
- Will the Department of War designation survive rehearing, and what share of revenue is government-related? [33]
Funding-round record, $169.4B total funding across 20 recorded transactions. Per-round lead investors are not separated in the record. [172]
The private-mark ladder steps up without a down round: $61.5B in March 2025, $183B in September 2025, $380B in February 2026 and $965B in May 2026. [172] The two April 2026 tranches at $350-355B sit below the February Series G mark and read as strategic tranches priced off earlier terms, not as a down round.
Press accounts describe the May round as $65B at a $965B valuation. [40] The funding record splits it into a $50B close plus $15B of April tranches. [172]
Secondary pricing ran ahead of the primary mark. In August, Rainmaker Securities CEO Glen Anderson said "the few sellers on our books are around $1.5 trillion." [149] Eric Schmidt holds about 1% from a 2021 Series A investment, per Bloomberg. [173]
Provider EV/Sales basis; the sales period differs across companies, so read the column as a range rather than like-for-like. None is a pure-play frontier-model developer; no listed pure-play exists.
At $1.5T on Q3 annualized revenue of $67.2B, Anthropic would trade at about 22x run-rate revenue. [154][104] That is above NVIDIA at 16.6x [176] and below Palantir at 56.5x [174].
The denominator is gross. Anthropic books cloud-marketplace sales on a gross basis as principal. [92] Those sales were about 48% of H1 2026 revenue. [93] CSPs remit amounts net of their fees, and the S-1 does not quantify those fees. [94] A net-revenue multiple would therefore be higher than 22x by an undisclosed amount, so read 22x as a floor on a net basis.
Multiples and required revenue are derived from the cited anchors, Q3 annualized revenue [104], the $110B year-end expectation [24] and NVIDIA's 16.6x [176]. Per-share values divide by ~16.94B fully diluted shares, implied by the 846,840,960 Benefit Fund shares equalling 5% of fully diluted shares at 30 June assuming Amazon-note conversion. [141] The offering will add primary shares that are not yet disclosed. Equity value approximates EV: cash and short-term investments of $131.7B sit against $6.2B of Amazon notes principal. [4][90]
The anchors are ordered $1.5T, $1.75T, $2.4T, and each sets a revenue requirement rather than a fair value. At NVIDIA's 16.6x [176], the $1.5T bear anchor requires about $90B of run-rate revenue, which is below the $110B investors expect by year-end. [24] The $2.4T bull anchor requires about $145B, more than double the Q3 run-rate of $67.2B. [104]
Damodaran puts the requirement in margin terms: a $1.5T valuation hinges on 30% margins. [158] The S-1 shows an H1 2026 adjusted operating margin of 10%, rising to 19% in Q2. [88][9] Q2 alone covers about two-thirds of that 30% requirement, so the bear anchor is underwriting operating leverage that is already showing up.
Prediction markets frame the distribution. Polymarket traders put the chance of a first-day close above $2T at about 42%, with $1.75-2T the most popular bracket at 28%. [186]
The evidence points one way, to an IPO in November. Signal, implication and source for each path:
- IPO track (best supported). Public S-1 for a Nasdaq Global Select listing under ANT, with 22 underwriters led by Goldman Sachs and Morgan Stanley. [20][146] Marketing from 9 November. [36]
- Stay private (displaced). Secondary sellers were scarce even at ~$1.5T in August, but the public filing ends that path for now. [149]
- M&A (not applicable). Anthropic is a buyer, not a target; the S-1 lists possible acquisitions or investments among uses of proceeds. [145]
- Distress (no signal). The record shows no down round in the primary ladder and a $42B debt financing on 1 October, which adds fixed claims ahead of equity. [172] The S-1 separately describes a $42B Broadcom note-purchase commitment, usable only for TPU lease fees, at 9% interest. [91]
The gate that would flip the read is a range below the $965B Series H mark, or a deferral past the 19-20 November window. [37] OpenAI has ruled out its own 2026 debut, which leaves Anthropic as the only large AI listing this year. [187]
- FY2026E revenue
- $57.8B
- +1,161% YoY
- vs $4.6B FY2025 actual
- FY2027E revenue
- $242.5B
- +319% YoY
- Q4 2027E $88.4B
- FY2027E adj. EBIT margin
- 28.1%
- Peak year of the forecast
- FY2028E adj. EBIT margin
- 17.5%
- -10.6pp YoY
- Compute spend doubles to $161.7B
- FY2030E revenue
- $800.4B
- Adj. EBIT margin 20.5%
- FY2030E cash
- $416.3B
- From $112.4B at 30 Jun 2026
All figures on this tab are Canalyst model output unless a citation is shown. The model is a first-look build on the S-1, with no consensus loaded yet.
Canalyst model. FY2024-FY2025 are actuals; FY2026 onward are forecasts.
The model forecasts compute spend as a fixed ratio of revenue. It holds revenue at 2.08x compute spend for H2 2026, matching the actual Q2 2026 ratio. It steps that to 3.0x through 2027, then drops it to 2.2x for FY2028-FY2030.
That step-down is the entire FY2028 margin dip. Compute moves from 33% of revenue in FY2027 to 45% in FY2028, adjusted EBIT margin falls from 28.1% to 17.5%, and net income falls from $47.5B to $31.0B while revenue still grows 47%.
The 2028 timing lines up with the S-1's capacity calendar. Google's additional TPU capacity is expected to be delivered between 2028 and 2031. [23] Undelivered leases are expected to commence between 2026 and 2028. [102]
Canalyst model output. The model groups costs as compute and infrastructure, S&M and G&A (personnel included), which differs from the S-1's cost of revenue / R&D / S&M / G&A presentation. The model's 'Total revenue' line is year-to-date for reported quarters; quarterly figures here use its in-period revenue line.
The model's historical quarters tie to the filing. Q1 2026 is $4,731M and Q2 2026 is $11,555M, matching the S-1. [7] Q2 EBIT of $974M is consistent with the S-1's ~$1.0B Q2 operating income, and the model's 19.3% Q2 adjusted margin matches the 19% filed. [9]
The model's Q3 2026E of $17,031M sits about 1% above the $16.8B Anthropic gave investors on an unaudited basis. [8] The model's Q3 adjusted EBIT margin of 12.8% compares with the adjusted operating margin above 11% reported from the investor day. [104]
The model's Q4 2026E revenue of $24,502M annualizes to about $98B. That is close to the $100B year-end run-rate investors had targeted, but below the $110B annualized revenue investors expect by year-end, per The Wall Street Journal. [24]
It needs Q4 to grow about 44% QoQ, close to the 45% Q3 pace. [84] Growth then decelerates through 2027, with Q4 2027E at $88,350M, about 20% above Q3 2027E.
The model therefore assumes no further step-down in sequential growth before 2027, which is the point Lemkin and other skeptics dispute. [160]
Cumulative modelled compute spend over FY2026E-FY2030E is about $878B. That exceeds the $265.6B of fixed or minimum commitments at 30 June 2026. [5] It also exceeds the ~$657B of named counterparty maximums plus the $125-150B of expected Google TPU payments. [23]
On the model's path the take-or-pay core is not a constraint, because forecast usage outruns contracted minimums by FY2027. The risk inverts: if revenue undershoots the model, the fixed Google, AWS and Microsoft minimums become the binding cost. [14][95][13]
On the model's numbers, every valuation anchor looks inexpensive on forward revenue. Even the $2.4T bull anchor is under 10x FY2027E revenue, against NVIDIA at 16.6x [176] on its own provider basis. The Valuation tab asked what run-rate revenue each anchor requires at NVIDIA's multiple; the bull anchor needed about $145B, and the model has FY2027E at $242.5B.
The earnings multiples tell a different story because of the FY2028 margin step-down. At $1.5T the stock would be 31.6x FY2027E net income but 48.3x FY2028E, as modelled compute intensity rises and earnings fall 35%. A buyer at the bear anchor is therefore underwriting three things: revenue roughly quadrupling in FY2027, compute efficiency improving to 3.0x, and the efficiency loss in FY2028 being temporary.
Damodaran's 30% margin test for $1.5T is met only in the model's FY2027 peak, where the adjusted EBIT margin reaches 28.1% before stock-based compensation. [158] On a reported EBIT basis the model never exceeds 19.2% in any period shown. That gap between adjusted and reported margin is about $24B of modelled stock-based compensation in FY2027, and it is the single largest reconciling item between the bull and bear readings of this model.
- Scale is confirmed and audited. H1 2026 revenue was $16,286M against $1,163M a year earlier. [1] Q2 alone was $11,555M. [7]
- The business turned profitable in Q2. Operating income was $1.0B at an 8% margin, and H1 operating cash flow was $3,929M. [9][10]
- The headline loss is the cap table. $72.4B of the $80.7B H1 net loss is a fair-value charge on convertibles. [81][82]
- The cost base is pre-committed but funded. Fixed or minimum commitments are $265.6B, against $131.7B of cash and short-term investments. [5][4] Expected Google TPU payments of $125-150B for 2028-2031 sit outside that figure. [23]
- Concentration is lower than the draft suggested. No customer exceeded 10% of nine-month revenue, and the top five were about 25%. [11]
- Half the revenue runs through rivals' marketplaces. CSP sales were 48% of H1 revenue, booked gross. [93][92]
- Control stays inside. The Founder LLC holds 50.1% of votes, the Long-Term Benefit Trust elects a board majority, and the Benefit Fund can take 2% more stock for every 50% gain from the IPO price. [34][136][44]
- The mark has already reset. Investor talk moved from ~$2T to ~$1.5T, about 22x Q3 run-rate revenue on a gross basis. [154][104]