Identify the strongest evidence and arguments that Anthropic's valuation is overstated or that the IPO range is too optimistic.

A $2 trillion IPO ask prices a future software franchise. The strongest bear case is that Anthropic is still a compute reseller with take-or-pay bills, flexible customers, and a price war—and that the numbers used to paper over that gap are not yet public, audited, or even consistent across the press.

The valuation rests on a run rate the prospectus has not yet proven can become cash

The reported target—about $2 trillion, more than double the $965 billion May 2026 Series H mark—is a bet on annualized revenue, not on 2025 results. Reuters’ review of the confidential draft prospectus shows 2025 revenue of nearly $4.6 billion, an operating loss of $8.06 billion (up from $2.98 billion in 2024), and a GAAP net loss of about $42 billion. Roughly $34 billion of that net loss was a non-cash remeasurement of financing that can convert into shares, not cash spent running the business. Compute and infrastructure alone was $7.33 billion—more than revenue, and about 58% of $12.65 billion in operating expenses. Cash and short-term investments at year-end 2025 were $20.28 billion. [1] [2] [3]

PitchBook’s reading of the same leak is the cleanest valuation critique: a $2 trillion price is about 435 times 2025 revenue and just under 31 times a late-July 2026 run rate above $65 billion. That multiple only works if gross margins rise, customers stay, and contracted capacity is used productively. PitchBook’s conclusion is explicit: the leak supports a valuation well above $1 trillion and does not justify $2 trillion until the public filing shows gross margin, the payment schedule on commitments, and a path to free cash flow. [4] [5] Private-market marks have not fully bought the ask either. Nasdaq Private Market put the company near $1.36 trillion as of October 1, 2026—about 41% above Series H, still roughly a third below $2 trillion. [6]

What this means: public investors are being asked to pay a multiple that assumes the July run rate is durable cash revenue. The last full year in the draft does not show that conversion.

Gross margin is the missing number, and the cost structure is the opposite of software

Anthropic does not have software economics yet. In 2025 it spent about $1.60 on compute and infrastructure for every dollar of revenue. [7] Epoch AI’s reconstruction from The Information reporting put 2025 inference compute near $2.7 billion on roughly $4.5 billion of revenue (implying about a 40% gross margin) and training compute near $4.1 billion—so serving customers and training models together dwarf everything else. [8] The Information reported that Anthropic cut its own 2025 gross-margin projection to 40% from an earlier 50% goal because inference costs ran about 23% above plan, even as cloud rental prices fell. OpenAI missed its margin forecast the same way. [9] [10]

The forward bill makes the margin question existential. The draft describes at least $518 billion of cloud, compute, and infrastructure obligations, about 80% non-cancelable or payable regardless of usage. Named pieces include at least $111.1 billion to Google (April 2026–July 2033), $110 billion to Amazon (May 2026–April 2036), $31.4 billion to Microsoft (non-cancelable except for uncured material breach), and about $161.2 billion of Broadcom-related equipment leases that neither side can cancel except on default. The filing language is blunt: if actual spend falls short, Anthropic must pay Google the difference, with similar terms on Amazon. [11] [12] PitchBook notes the draft still does not disclose gross margin—the figure that decides whether those contracts are a moat or a trap. [5]

Modeled tables circulating online (including quarterly gross margins climbing into the mid-50s and a Q2 2026 adjusted operating profit) should be treated as estimates, not prospectus facts. PitchBook’s only hard operating signal is a second-quarter 2026 adjusted operating profit excluding stock-based compensation, which it says does not establish cash generation or gross-margin improvement. [5] [13]

What this means: a competitor or a short does not need Anthropic to “fail.” It needs token prices or utilization to fall while minimum payments stay fixed. That mismatch is already written into the contracts.

Circular financing turns suppliers into underwriters of their own demand

Amazon, Google, and Nvidia are not arm’s-length vendors. They are investors, landlords, distributors, and—in Google and Amazon’s case—model competitors.

• Amazon has put about $18 billion in (convertible notes plus preferred), with more capacity available, and carries the stake near $190 billion. Anthropic owes AWS about $110 billion through 2036 on take-or-pay terms. An April 2026 expansion took Amazon’s potential commitment toward $33 billion in exchange for more than $100 billion of AWS spend over a decade, including Trainium. [14] [15] [16]
• Google is owed at least $111.1 billion on similar shortfall terms. Separate reporting describes Google at the center of a large TPU financing network (private-credit SPVs that buy hardware and lease it to Anthropic), so Google is chip supplier, investor, and credit backstop at once. [11] [17]
• Nvidia committed up to $10 billion of equity in a late-2025 arrangement with Microsoft, and later structures go further: a reported $35 billion Lambda cloud deal in which an Nvidia-backed neocloud serves Anthropic inside a data center whose lease Nvidia holds. Trade press put Nvidia’s contracted value with Anthropic above $180 billion. Nvidia’s CFO has rejected the “circular financing” label; the Financial Times’s cleaner description is old vendor financing—writing checks so customers can buy more product than they could otherwise afford. [18] [19] [20]

The accounting loop feeds the valuation loop. Amazon marked Anthropic notes from $42.2 billion on March 31 to $97.9 billion on June 30—a $55.7 billion one-quarter increase driven by private-round prices, not by cash distributions. [5] Marketplace distribution closes the circle in the income statement: 47% of 2025 sales, about $2.16 billion, ran through Amazon and Google, up from 11% in 2023 and 32% in 2024, with roughly $351 million paid back as distribution fees booked in operating expense rather than as a contra-revenue. [21]

What this means: demand, capex, and private marks are partly the same capital recycled. If public investors haircut the equity value, supplier marks, credit appetite, and “committed” capacity can reprice together.

Customer concentration and run-rate durability are the revenue-side mirror of those fixed bills

Two unnamed customers each produced 12% of 2025 revenue. The company warns that many of its largest customers are not on long-term contracts and can cut or stop spending. Consumption-based usage was about $3.8 billion of 2025 revenue versus $789 million of subscriptions, and management expects consumption to remain the substantial majority. Cloud partners collected 60% of the $909 million in customer bills outstanding at year-end 2025. [21] [21]

That is the bear mechanism in one sentence: revenue can fall inside a quarter; minimum infrastructure payments run into the 2030s. [5] Earlier reporting tied a large slice of the coding boom to Cursor and GitHub Copilot. Those are sophisticated buyers who can multi-home the day a cheaper model is “good enough.” Even the company’s own growth narrative concedes the base is lumpy: high-value API accounts, not a broad installed base of sticky seats, move the top line. [22] [23]

Run-rate headlines have also outrun recognized revenue before. Company and press figures put year-end 2025 annualized revenue near $9 billion and, in at least one CNBC-cited account, “actual” 2025 revenue near $10 billion. The draft prospectus puts recognized 2025 revenue near $4.6 billion. Q1 2026 revenue of $4.73 billion and preliminary Q2 revenue above $11.5 billion show the business did accelerate—but they also show why annualizing a peak month is not the same as a year of GAAP sales. [24] [25] [5]

What this means: the IPO case treats $65 billion, or $100–120 billion by December, as the valuation base. The filing’s own risk factors say the customers behind that curve are not locked in.

Price competition is already attacking the margin expansion the multiple requires

September 2026 made commoditization operational, not theoretical. OpenAI priced GPT-6.1 Sol at one-fifth the token cost of its own most capable model. Anthropic launched Claude Opus 5.5 at token prices 20% below Opus 5 and claimed about 40% lower effective cost than Opus 5 because the model uses fewer tokens. Gartner’s Anushree Verma described general-purpose models as increasingly interchangeable, with vendors grabbing share on price. [26] [27]

Chinese open-weight labs set the floor. Juniper Research, citing OpenRouter, said leading closed providers’ share of work on that platform fell from about 70% to 30% in a year, and that Chinese models typically run 60–90% cheaper. DeepSeek’s V4.1 Flash launch in September hit listed Chinese AI stocks and was framed as fresh pressure on labs including Anthropic. [28] [29]

The non-obvious implication: Anthropic’s scarcity thesis—“limited principally by the availability of compute”—justifies $518 billion of take-or-pay. A price war says the scarce input is being turned into a cheaper output faster than those contracts depreciate. If intelligence per dollar keeps falling, utilization can stay high and revenue per committed megawatt can still disappoint. [11] [26]

What this means: competing labs do not need to beat Claude on benchmarks. They need to be close enough that procurement switches on price, which is exactly when fixed compute bills hurt most.

Legal and government exposure is no longer a footnote, and some of it is already adjudicated

Copyright is a settled cash cost plus an open tail. In Bartz v. Anthropic, the company agreed to a $1.5 billion class settlement—described by plaintiffs’ counsel and the court record as the largest known U.S. copyright recovery—covering roughly 482,000 to 500,000 works at about $3,000 per work. Judge William Alsup had held that training on lawfully acquired books can be fair use, but that downloading and storing pirated copies was not. Final approval came on July 20, 2026. The release covers past acquisition and copying through August 25, 2025. It does not release output claims or future conduct. Funding is in installments through September 2027. [30] [31] [32] $1.5 billion is small next to a $2 trillion ask, but the open output docket is the live risk, and authors and publishers are still fighting over who gets paid. [33]

Government risk is fresher and closer to the IPO. After Anthropic refused to allow military use for autonomous weapons and mass domestic surveillance, the Pentagon designated it a supply-chain risk—the first time that label was applied to a U.S. company. On September 25, 2026, the D.C. Circuit upheld the designation 2-1. Anthropic has said the fight cost billions in lost business and damaged its reputation ahead of the IPO. A separate California ruling limited a broader ban; the Pentagon bar itself stands. [34] [35] [36]

In June 2026 the Commerce Department, using export-control authority, ordered Anthropic to cut foreign-national access to Claude Fable 5 and Mythos 5. The company disabled the models broadly to comply. Restrictions were lifted on June 30 after Anthropic coordinated mitigations with the government. The prospectus warns that government attitudes can hurt commercial customers and partners, not just public-sector revenue, which the company says is under 1% of sales. It also discloses that advanced AI could pose “catastrophic or existential risks to humanity,” and Reuters reported an FTC industry probe that includes Anthropic. [37] [38] [39]

What this means: the copyright settlement is largely a known liability. The regulatory pattern—export “is informed” letters, a supply-chain designation upheld on appeal, and a prospectus that flags contagion into commercial accounts—is a recurring tax on the enterprise growth story the IPO is selling.

Bubble talk, delayed listings, and precedents that rhyme—even without an Anthropic down round

Anthropic itself has not taken a private down round. The mark went from the May $965 billion round to about $1.36 trillion on Nasdaq Private Market. The stress shows up around it. The IPO slipped from an October target to after the November midterms, with marketing discussed for as early as the week of November 9. OpenAI postponed its own listing, with Sam Altman arguing it would be ill-advised to go public now. [40] [6]

Skeptics are arguing from capital structure, not vibes. Ed Zitron has called private-credit financing of the AI buildout a “brewing crisis,” pointing to CoreWeave’s $35.6 billion of debt as of June 30 and Goldman’s count of lower-rated AI borrowing. His June claim that Anthropic could not sustain a $47 billion run rate was later contradicted by the company’s own higher figures—which is a useful reminder that bears have been early, and wrong, on the top line. The $518 billion commitment stack is what remains after that miss. [41] Wealth managers quoted by Reuters have said they cannot underwrite multi-trillion equity values on companies that lose billions and require massive capex. [42] Commentary around the filing has revived Scott McNealy’s post-dot-com warning on paying rich sales multiples for businesses that never earned their cost of capital, and telecom-style vendor financing as the historical rhyme for Nvidia’s balance-sheet support. [43] [20]

There is no clean public precedent of an Anthropic-scale lab repricing in a down round yet. The nearer precedents are adjacent: neocloud leverage, Oracle credit spreads cited in bust scenarios, and the gap between derivative pre-IPO prints above $2 trillion and the last institutional round at $965 billion. Decentralized perpetual futures have implied valuations above $2 trillion; those markets are not price discovery for an S-1. [44] [45]

Leaked and press-reported figures have already been wrong, incomplete, or non-comparable

Treat every number that is not in a public EDGAR filing as provisional. As of late September, searches of SEC filings had not produced a public Anthropic S-1; Reuters reviewed a confidential draft, and Anthropic declined to comment. [1] [46]

Documented reliability problems:

• Gross versus net revenue. In an April 2026 internal memo reported by CNBC and The Verge, OpenAI’s chief revenue officer Denise Dresser told staff Anthropic’s then-$30 billion run rate was inflated by about $8 billion because Anthropic grosses up cloud-marketplace revenue while OpenAI reports Microsoft-channel revenue net. Both treatments can be GAAP-compliant. Anthropic’s defense, given to Reuters, is that it is the principal and sets price and delivery. OpenAI’s claim is adversarial, but Reuters later confirmed the gross-up is real: full marketplace billings hit revenue, partner cuts hit operating expense. [47] [21]
• Run rate versus recognized revenue. End-2025 “run rate” figures near $9 billion, and at least one press account of $10 billion in 2025 “actual” revenue, do not match the prospectus figure of nearly $4.6 billion. Annualizing a strong month is a company-preferred metric, not a GAAP year. [25] [1]
• The $42 billion loss is easy to misuse. About $34 billion is fair-value accounting on convertibles, which rises when the company is marked higher. The operating loss above $8 billion is the economic number. Bull notes that ignore it, and bear notes that lead with $42 billion of “cash burn,” are both wrong. [1]
• Forward revenue is investor talk. The $100–120 billion year-end 2026 range came from Anthropic backers talking to the Financial Times, not from audited guidance. One investor’s “800% growth justifies 30x revenue, so $3 trillion” is a circular multiple, not a forecast. [48]
• Margin path has already slipped once. Internal 2025 gross-margin hopes were cut before the year closed. Any third-party quarterly margin table that is more precise than the draft prospectus should be labeled a model. [9] [5]
• 2026 quarterly revenue in the leak is partly preliminary. PitchBook flags second-quarter revenue above $11.5 billion as preliminary. [5]

What this means: the bull case is built on the least audited number (run rate) and the bear case is sometimes built on the most misleading one (GAAP net loss). The figures that survive both filters—operating loss, compute spend above 2025 revenue, 80% non-cancelable commitments, two customers at 12% each, 47% channel sales—are the ones in Reuters’ reading of the draft.

Bull counterpoints that a serious bear has to answer

The growth is not fake. Revenue rose about twelvefold in 2025. First-quarter 2026 revenue exceeded all of 2025. Preliminary second-quarter revenue topped $11.5 billion. By late July the run rate was above $65 billion. If December annualized revenue really lands at $100–110 billion, $2 trillion is roughly 18–20 times sales, which Morningstar’s Michael Field called less extreme than SpaceX’s debut multiple. [1] [49] [42] FT’s Lex column goes further: if 2028 sales hit the levels bulls imply, $2 trillion could be a mid-single-digit to low-teens multiple of that year’s revenue—and the bull case can be stretched to absurdity ($5–10 trillion) by inflating TAM. That is an argument about narrative elasticity, not about 2025 cash flow. [50]

Other fair counters:

• Inference cost per dollar of revenue has been falling in company projections and in third-party models. A sustained gross margin in the 40–50% range, which some pre-filing analysis expected, would confirm half the private-round underwriting. The prospectus can confirm that half. It cannot confirm 2030 revenue. [51]
• Enterprise mix is a real difference versus a consumer-heavy peer, if it holds. Channel partners are also distribution, not only concentration.
• Training on legally acquired books was held to be fair use; the $1.5 billion deal is finite and partly paid.
• Export controls on Fable and Mythos were lifted within weeks. Direct government revenue is under 1%.
• If compute really is the binding constraint, take-or-pay contracts are capacity insurance, and fixed costs leverage margins up as usage grows inside contracted power. PitchBook states that upside explicitly. [5]
• Ed Zitron’s mid-2026 call that the $47 billion run rate would not hold has already been overtaken by higher company figures. Top-line bears have a losing recent record. [41]

The disagreement is not whether Claude is selling. It is whether a business that in 2025 spent more on compute than it earned, that has locked in hundreds of billions of dollars it must pay even if usage disappoints, and that sells a product whose price is being cut by rivals and by itself, should be capitalized at twice the price sophisticated investors paid four months earlier. PitchBook’s line is the fairest summary of the evidence now in hand: fast growth is established; $2 trillion is not. [5]

---

Recent Findings Supplement (October 2026)

Anthropic’s September 2026 IPO prospectus (reviewed by Reuters) reveals 2025 revenue of ~$4.59 billion (12x growth from $386 million in 2024) against an $8.06 billion operating loss and $7.33 billion in compute/infrastructure spend, with $518 billion in future cloud/compute commitments (roughly 80% non-cancelable or take-or-pay). This structure shows how explosive top-line growth coexists with capital intensity that far exceeds current cash ($20.28 billion at year-end 2025) and revenue, while improving gross margins in some analyses (e.g., quarterly models showing progression toward 50%+) still leave the company dependent on continued hyperscaler financing and usage-based revenue that may not cover fixed obligations.[1][2]

• Compute spend tripled year-over-year and represented ~58% of 2025 operating expenses; Q2 2026 showed early adjusted operating profit signals in some reports, with gross margins cited around 52% or higher before partner shares/training costs in select analyses.[3]
• Projections in the filing and related reporting include 2028 revenue of $190–200 billion, but the $518 billion commitment horizon (e.g., $111.1B to Google through 2033, $110B to Amazon through 2036) dwarfs near-term cash flow.[4]
• For competitors or new entrants: Securing equivalent long-term compute at scale requires either deep Big Tech ties or alternative hardware paths; pure-play model developers face margin pressure unless they achieve materially better utilization or vertical integration.

Anthropic’s relationships with Amazon, Google, and Nvidia illustrate circular financing where investors double as suppliers, revenue collectors, competitors, and (in Nvidia’s case) potential IPO anchors, creating interdependent cash flows that may not reflect arm’s-length demand.[5]

• Amazon and Google (major investors and primary cloud partners) accounted for 47% of 2025 revenue routed through their marketplaces (~$2.16 billion), with Anthropic paying ~$351 million in distribution fees (~16% take); these same firms collect bills and compete directly in AI.[6]
• Nvidia has discussed up to $10 billion as an IPO anchor (in addition to prior commitments) while Anthropic buys Nvidia-powered capacity through partners; similar loops exist with Microsoft and others.[7]
• Implications: New entrants or rivals must navigate or replicate these ecosystems; any disruption (e.g., partner prioritization of their own models or regulatory scrutiny of vertical integration) could cascade to both funding and distribution.

Two unnamed customers each contributed 12% of 2025 revenue (24% combined), with many large customers lacking long-term contracts and able to reduce spending without corresponding cost relief; 47% of revenue flows through Amazon/Google channels whose share has risen sharply.[8]

• Run-rate figures accelerated from ~$9 billion annualized at end-2025 to $47 billion by May 2026 and >$65 billion by end-July 2026, with investor expectations of $100–120 billion by year-end 2026.[9]
• OpenAI has reportedly questioned gross vs. net revenue recognition on some partner deals, highlighting potential inflation in reported run-rates.[10]
• For market participants: High concentration and flexible customer terms create downside asymmetry—revenue can drop faster than costs (especially fixed compute commitments)—favoring diversified or enterprise-locked models over pure usage-based plays.

Reported IPO targets of >$2 trillion (more than double the $965 billion May 2026 private valuation) imply ~30x the July 2026 run-rate or higher multiples assuming continued 800%+ growth, yet the prospectus underscores losses, commitments, and risks that have prompted AI-bubble comparisons and down-round precedents elsewhere in tech.[11]

• Bull case (investor views in FT reporting): Sustained hyper-growth and margin expansion (some quarterly models project positive adjusted operating margins by late 2026) could justify premiums akin to or exceeding SpaceX’s $1.77 trillion IPO.[12]
• Bear case (PitchBook/Morningstar analysis of leaked figures): At $2 trillion, valuation assumes stable customers, productive use of contracted capacity, and rising margins that have not yet produced sustained free cash flow; compute costs per revenue dollar have declined but remain material.[13]
• Implications: Public-market scrutiny may force more conservative pricing or disclosure; entrants betting on similar multiples face execution risk if growth decelerates.

New legal and regulatory developments include final approval of a $1.5 billion copyright settlement (largest known U.S. copyright recovery) in July 2026 over unauthorized book downloads for training, plus a September 2026 appeals court ruling upholding a Pentagon “supply-chain risk” designation restricting DOD use of Anthropic products.[14][15]

• Some authors opted out and continue separate cases; the settlement requires destruction of pirated files and provides ~$3,000 per work.
• The DOD designation (typically for adversary-linked firms) survived one appeal despite a partial earlier court block, with Anthropic considering further review.
• Implications: Copyright exposure remains live for non-settled claims and peers; government relations risks (export controls or procurement bans) could limit addressable markets for frontier labs.

September 2026 model launches by Anthropic (Claude Opus 5.5) and OpenAI featured aggressive price/performance cuts (20–80% token price reductions plus efficiency gains yielding ~40%+ effective savings), intensifying competition with each other and cheaper open-weight models (including Chinese offerings), supporting commoditization concerns.[16]

• Anthropic has resisted deep enterprise discounts in favor of strict metered billing, contrasting with more flexible competitor approaches.
• Past run-rate or valuation reports (e.g., earlier private rounds or leaked figures) have shown variability in recognition and growth assumptions, underscoring the need for prospectus-level verification over press estimates.
• For competitors: Price competition compresses margins and rewards efficiency/segmentation; durable advantages may shift toward distribution, data, or vertical applications rather than raw model capability.

These post-April 2026 disclosures (primarily the late-September prospectus and related reporting) provide the most concrete, company-sourced data yet, tempering optimism around valuation multiples with explicit warnings on costs, concentration, and obligations. Bullish growth narratives persist but rest on execution assumptions that public markets will test directly.