Is Anthropic worth $965 billion?

1. Revenue and Profitability

The revenue numbers are where the $965 billion valuation either justifies itself or does not.The revenue numbers are where the $965 billion valuation either justifies itself or does not.

The trajectory. Anthropic's annualised revenue run rate over the past 18 months [2, 6, 7, 48]:

Early 2025: approximately $1 billion annualised
August 2025: approximately $5 billion annualised
End of 2025: approximately $9 billion annualised
April 2026: approximately $30 billion annualised
May 2026: $47 billion annualised, disclosed in Anthropic's 28 May 2026 Series H announcement [57]

That trajectory is a 47x increase in annualised revenue in approximately 18 months. Q1 2026 closed at approximately $4.8 billion in revenue, and Anthropic disclosed in its 28 May 2026 Series H announcement that run-rate revenue had crossed $47 billion earlier in the month [57]. No company on the public record has grown this fast at this scale.That trajectory is a 47x increase in annualised revenue in approximately 18 months. Q1 2026 closed at approximately $4.8 billion in revenue, and Anthropic disclosed in its 28 May 2026 Series H announcement that run-rate revenue had crossed $47 billion earlier in the month [57]. No company on the public record has grown this fast at this scale.

These headline figures are not undisputed. The methodology behind both the ARR number and the Q2 operating margin claim has been challenged on substantive grounds by OpenAI, by PitchBook's research desk, and by independent investigative reporting from Ed Zitron. Section 10 of this article walks through the criticism in detail. The reader is encouraged to read what follows with those questions in mind.These headline figures are not undisputed. The methodology behind both the ARR number and the Q2 operating margin claim has been challenged on substantive grounds by OpenAI, by PitchBook's research desk, and by independent investigative reporting from Ed Zitron. Section 10 of this article walks through the criticism in detail. The reader is encouraged to read what follows with those questions in mind.

Annualised run-rate evolution

Anthropic OpenAI

Source: Reported annualised run-rate disclosures compiled from Anthropic press releases (Series F, Series G, Series H, Claude Code, Google & Broadcom partnership), OpenAI announcements, Bloomberg, CNBC, The Information, Financial Times, New York Times, Wall Street Journal. Hover any data point for the original source.

Peer comparison on revenue. OpenAI's most recently disclosed annualised revenue is approximately $24 billion, on a $2 billion per month run rate stated in its 31 March 2026 funding-round announcement [8]. That number was $20 billion at end-2025 per the company's earlier scaling-with-intelligence post, so OpenAI added roughly $4 billion of annualised revenue in three months. OpenAI is growing rapidly but at a slower absolute and relative trajectory than Anthropic. xAI's revenue is not reliably disclosed but estimated in the single-digit billions. Mistral's revenue is reported in the hundreds of millions, and Perplexity reached approximately $500 million annualised in April 2026 [47]. Google DeepMind's contribution is folded into Google Cloud and Search revenue and not separately disclosed; Meta AI's similarly into Meta's segments. The state-aligned Chinese frontier labs (DeepSeek, Qwen) do not disclose revenue for Western markets.

The revenue mix. Anthropic does not publicly disclose a clean segmentation, but the reporting permits a working approximation [6, 7, 9]. Enterprise API, sold directly and through AWS Bedrock, is the largest segment. This revenue is high-margin, sticky once integrated, and tied to consumption patterns that scale with the customer's underlying use case. The enterprise base is concentrated and growing fast: Anthropic disclosed at its May 2026 developer event that more than 1,000 customers now spend over $1 million annually on its products, doubled from approximately 500 in February 2026 [36]. Consumer (Claude.ai and Claude Pro) is meaningful but smaller and exposed to product preference cycles. Reseller revenue through AWS Bedrock is a hybrid: Anthropic recognises revenue on the underlying API consumption while AWS captures the customer relationship and a margin layer. Strategic deal revenue from the Amazon and Google commitments flows through compute and partnership structures and is not separately broken out.

The first profitable quarter. For most of Anthropic's history, the profitability conversation was theoretical. The company operated at a loss, the loss was funded by capital raises, the question was when the unit economics would inflect. That conversation changed in May 2026 when the Wall Street Journal reported that Anthropic is projected to turn an operating profit of $559 million on $10.9 billion of revenue in Q2 2026 [2]. This would be the company's first ever profitable quarter.

The number is a projection, not a closed period. 5% operating margin is not the margin of a mature software business (Microsoft, Oracle, and Adobe operate north of 30%); it is the margin of a business that has just crossed the breakeven line and has not yet captured the operating leverage that scale provides. The cost structure is dominated by compute. The gross margin profile might be closer to a capital-intensive cloud business than to a pure-software business, although it is hard to anticipate today how the business model of frontier AI labs will evolve, given the pace of technological change.

2. Valuation and Funding History

The funding-round trajectory tells a story about capital availability before it tells a story about value.

Anthropic post-money valuation across funding rounds

Closed rounds hover any bar for round details

Source: Anthropic press releases (Series E, F, G, H); Bloomberg, CNBC, TechCrunch, The Information (Series B, C, D).

Funding rounds summary [10, 11, 12, 13, 14]:

Series A, May 2021: $124M raised, valuation not publicly disclosed, led by Jaan Tallinn
Series B, April 2022: $580M raised, ~$4B post-money, led by Alameda Research (Sam Bankman-Fried) with $500M of the $580M; stake later sold by FTX bankruptcy estate
Series C, May 2023: $450M raised, $4.1B post-money, led by Spark Capital — Google made a $300M strategic investment in February 2023, ahead of this round.
Series D, December 2023 to early 2024: ~$2.8B raised, $18.4B post-money, led by Menlo Ventures — Amazon committed up to $4B in convertible notes in September 2023, completed across two tranches in March and November 2024.
Series E, March 2025: $3.5B raised, $61.5B post-money, led by Lightspeed Venture Partners — Google added a third strategic tranche of approximately $1B in January 2025, around the time of Series E.
Series F, September 2025: $13B raised, $183B post-money, co-led by Iconiq Capital, Fidelity, and Lightspeed
Series G, February 2026: $30B raised, $380B post-money, led by GIC and Coatue — In April 2026 Amazon committed an additional $25B plus 5GW of compute capacity; Google committed up to $40B option-based on milestones.
Series H, 28 May 2026 (closed): $65B raised, $965B post-money, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital; co-led by Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN [57] — Series H included $15B of previously committed hyperscaler investments rolled into the round, $5B of which came from Amazon. Additional compute agreements: Amazon up to 5GW new capacity; Google and Broadcom 5GW of next-generation TPU capacity.

Peer comparison on valuation. OpenAI's most recent primary valuation is approximately $852 billion, set in its 31 March 2026 $122 billion funding round [17]. xAI was valued at approximately $250 billion as part of the February 2026 SpaceX-xAI merger, up from approximately $230 billion in its January 2026 Series E [58]. Mistral closed a €1.7 billion Series C in September 2025 at €11.7 billion post-money, led by ASML [59]. Perplexity reportedly raised at approximately $20 billion in September 2025, with later reporting placing it closer to $22 billion in early 2026 [60]. These peer figures should be read with a caveat: AI valuations have moved fast over the past twelve months, and the snapshots here are months old for everyone except OpenAI and Anthropic, both of which priced in the past two months. The Series H closed on 28 May 2026 at $965 billion post-money, making Anthropic the most valuable private pure-play AI company in the world [57].

Strategic rounds versus pure equity rounds. Three of the most consequential capital events in Anthropic's history are not standard equity rounds. Amazon has invested $8 billion across three tranches and committed an additional $25 billion plus 5 gigawatts of compute capacity in April 2026 [19]. Google has invested $3.3 billion across three tranches and committed an additional $40 billion in April 2026, option-based on Anthropic meeting performance milestones [20]. The Series H itself included $15 billion of previously committed hyperscaler investments rolled into the round, $5 billion of which came from Amazon, alongside additional compute agreements with Amazon for up to 5 gigawatts and with Google and Broadcom for 5 gigawatts of next-generation TPU capacity [57]. The headline equity rounds (Series B through Series H) understate the cumulative capital and strategic coupling that Anthropic has accepted.

The peer comparison on revenue multiples produces wildly different intensities: Anthropic at ~21x annualised revenue (using the $47 billion disclosed run-rate), OpenAI at ~35x ($852 billion on $24 billion annualised), xAI at materially higher on a small revenue base. One valuation multiple is worth tracking. In 2025, a recurring valuation framework crystallised for fast-growing AI-related companies: enterprise value divided by next-twelve-month revenue (NTM) anticipations. The framework matters because it makes the implicit growth assumption explicit. The multiple is a single number that captures how much future growth the market is paying for today. OpenAI's 2025 rounds illustrate the frame. OpenAI raised at approximately $300 billion in early 2025 and at approximately $500 billion later in the year. Both rounds priced the company at roughly 20x next-twelve-month revenue. Anthropic's Series G in February 2026, at $380 billion, implied approximately $19 billion of NTM at the same 20x multiple. In February, before the Q1 inflection, that NTM projection looked reasonable. At $965 billion, a 20x EV/NTM multiple implies $48 billion of next-twelve-month revenue. The disclosed run-rate as of May 2026 is already $47 billion [57]. The 20x multiple at $965 billion therefore assumes essentially no growth from here, which sits awkwardly against the rest of the company's trajectory. Run the numbers in the other direction. The exercise is speculative, but it is the kind of speculation the multiple forces. If Anthropic's NTM reaches $65 billion, which would itself imply a dramatic slowdown from the current pace, the implied multiple compresses to roughly 15x. At $96 billion of NTM, the multiple is 10x. Both sit well below the multiples applied to OpenAI's 2025 rounds and to Anthropic's own previous rounds.

3. Services and Products

Anthropic's product surface has expanded materially over the past 18 months. The product family currently in market:

The Claude model family. Three primary tiers: Haiku for compact deployment, Sonnet for balanced inference, Opus for flagship reasoning. The most recent generation in the public record is Claude Opus 4.8, released on 28 May 2026, alongside Claude Sonnet 4.6 and Claude Haiku 4.5 [21, 65]. Opus 4.8 shipped just 41 days after Opus 4.7 and includes a Dynamic Workflows research preview in Claude Code that orchestrates hundreds of parallel subagents. Each generation has historically delivered measurable capability improvements on standard benchmarks while preserving Anthropic's distinctive safety-research positioning. The model family is the foundation; everything else is built on it.

The developer API is the largest revenue line, sold directly and through AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry on Azure to developers and enterprise customers building production AI applications. Pricing is per token, with volume discounts and committed-use agreements. The API competes with OpenAI's API and Google's Gemini API on capability, latency, pricing, and the developer-experience surface. In the Series H announcement, Anthropic stated that Claude is the first frontier model available on all three of the world's largest cloud platforms (AWS, Google Cloud, and Microsoft Azure), a distribution footprint no other frontier lab has matched [57].

Claude.ai is the consumer-facing product, available on web, iOS, Android, and desktop. The pricing ladder runs from a free tier with daily usage limits, to Claude Pro at $20 per month, to Claude Max at $100 per month for five times Pro's usage allowance and $200 per month for twenty times. For organisations, Claude Team is $25 per seat per month for the Standard tier and $125 per seat per month for the Premium tier that includes Claude Code, both with a five-seat minimum. Claude Enterprise is custom-priced for organisations with compliance, data-residency, or governance requirements, and includes the 1-million-token context window, SAML SSO, HIPAA readiness, and dedicated support. The consumer and team plans are a smaller revenue line than the API but serve as a brand-builder and a real-time feedback loop into product development.

Claude Code is a command-line tool for software developers that lets Claude act as a collaborator inside a developer's terminal environment. Released in early 2025 and updated continuously. Anthropic disclosed that Claude Code crossed approximately $1 billion in annualised revenue within six months of public launch, one of the fastest software-product growth curves on record [36]. It is also the product that has attracted the loudest user feedback, both positive (productivity gains for senior engineers) and negative (the Q1 2026 performance complaints that became a public-relations issue, covered in Section 8).

Computer Use is an agentic capability that allows Claude to operate desktop applications by reading screen state and producing input actions. Released in late 2024 and refined in 2025. The capability that defines Anthropic's positioning as a builder of agents, not only models.

Claude in Chrome is a browser-based agentic surface that allows Claude to perform tasks inside the user's Chrome browser. A consumer-product direction that puts Anthropic in direct distribution competition with Google's own Chrome-AI integrations.

Cowork is a desktop product released in early 2026 that wraps the Claude assistant for non-technical users to automate file and task management. The launch of Cowork was the immediate trigger of the February 2026 SaaSpocalypse (covered in Section 8), as the market priced in the implication that agentic automation could compress seat-licensed knowledge work [3, 4, 5].

Claude Design is a research-preview product launched by Anthropic Labs in April 2026. It turns natural-language prompts into clickable prototypes, slide decks, one-pagers, and mockups, and can apply a team's design system by reading the team's codebase and existing design files. Exports flow to PDF, PPTX, URLs, and Canva. Powered by Claude Opus 4.7 and aimed at founders and product managers without a design background, Claude Design positions Anthropic in direct competition with Figma for the design-tooling layer [61].

Claude Agent SDK is a developer toolkit for building custom agents on top of the Claude model family. The infrastructure underneath Claude Code and Computer Use, exposed for third-party developers.

The connector layer. Connectors and integrations form the surface that lets Claude act inside the user's other tools. The current connector library covers Gmail, Google Calendar, Google Drive, Slack, Notion, Asana, GitHub, Figma, Canva, Webflow, Salesforce, and Adobe's creative applications, among many others, and is open for third-party publishers through Anthropic's MCP plugin marketplace. The strategic importance of this layer is that an AI that can read and write inside existing workflows is harder to displace than one that lives in its own chat window.

The trust dividend. Anthropic has built its public identity around safety more deliberately than any other frontier lab. The company was founded in 2021 by former OpenAI researchers who left over disagreements about deployment pace, and that founding posture continues to shape how Anthropic engages the outside world. Dario Amodei testified to the U.S. Senate Judiciary Subcommittee on Privacy, Technology, and the Law in July 2023, and Anthropic is a founding member of the Frontier Model Forum alongside OpenAI, Google, and Microsoft. The pattern carries through to two high-stakes events in the past two months. On 7 April 2026, the same day Anthropic announced Claude Mythos Preview, a frontier model strikingly capable at discovering software vulnerabilities, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened a closed-door meeting with the chief executives of the largest U.S. banks to discuss the cybersecurity implications for the financial system [62]. Rather than release Mythos publicly, Anthropic stood up Project Glasswing, a controlled-access programme for roughly fifty defensive-security organisations including Amazon, Apple, Google, JPMorgan Chase, Microsoft, and Palo Alto Networks. On 25 May 2026, Pope Leo XIV presented his encyclical on artificial intelligence, Magnifica Humanitas, alongside Anthropic co-founder Christopher Olah, the first time a pope has personally unveiled an encyclical [63]. Anthropic had previously consulted Catholic scholars and Vatican representatives, including Bishop Paul Tighe of the Vatican's Dicastery for Culture and Education, while developing the Claude Constitution that guides the company's ethical framework. The strategic value of this posture is not abstract. Regulated buyers, governments, and large enterprises tend to prefer vendors that can be defended internally if something goes wrong, and Anthropic has spent four years building exactly that defensibility. The mythos is itself a commercial asset.

The product question that matters for valuation is which product carries which revenue and which product provides the moat. The revenue answer is mostly the API. The moat question is less obvious: the Claude models are competitive at the frontier but not unambiguously ahead of OpenAI's frontier or Google's Gemini. Benchmark leadership rotates between labs every few months.

4. Main Investors and Partners

Anthropic's cap table and partnership structure are unusual for a venture-stage company. The cap table reflects three layers, and a fourth layer of strategic coupling that sits alongside but outside the equity rounds.

Founders and team. Dario Amodei and Daniela Amodei founded the company in 2021 together with several former senior OpenAI researchers. The founding team's combined stake, after eight priced rounds of dilution, is not publicly disclosed at the share level. The founders retain control through governance arrangements that go beyond pure equity.

Venture syndicate. Across Series A through Series G, the lead investors and meaningful co-investors include Jaan Tallinn (Series A), Alameda Research and Sam Bankman-Fried (Series B, stake later sold by the FTX bankruptcy estate to a consortium including Abu Dhabi's sovereign wealth fund), Spark Capital and Menlo Ventures (Series C, D), Lightspeed Venture Partners (Series E), Iconiq Capital, Fidelity, and Lightspeed (Series F), GIC and Coatue (Series G with D. E. Shaw, Dragoneer, Founders Fund, Iconiq, MGX). The Series H closed on 28 May 2026, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, and co-led by Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN [57]. Significant participants in the round include Baillie Gifford, Blackstone, Brookfield, D. E. Shaw Ventures, DST Global, Fidelity, General Catalyst, Insight Partners, Jane Street, Lightspeed, MGX, T. Rowe Price, and Temasek.

Strategic partners. Amazon and Google hold positions that combine equity, convertible debt, and compute commitments.

Amazon: ~$8B equity across September 2023, March 2024, November 2024. Additional $25B + 5 gigawatts of compute capacity committed April 2026 [19]. AWS is named Anthropic's primary cloud and training partner. AWS Bedrock is the largest enterprise distribution channel for Anthropic's API.
Google: ~$3.3B equity across February 2023, October 2023, January 2025. Additional up to $40B committed April 2026, option-based [20]. Google is simultaneously Anthropic's investor and Anthropic's most direct competitor at the frontier through Gemini.

The fourth layer: compute counterparties. SpaceX (operating xAI's Colossus 1 and 2 facilities) is now a major commercial counterparty through the $45 billion compute deal covered in Section 5. NVIDIA is Anthropic's GPU supplier and a strategic player at the upstream end of the value chain.

Strategic infrastructure partners. Anthropic added Micron, Samsung, and SK hynix as named strategic partners in the Series H announcement, recognising the role memory, storage, and logic chips play in the scaling of frontier inference and training capacity [57].

The reader considering exposure should know that what looks like a venture cap table is in fact a hybrid. A material share of Anthropic's economic future is committed to compute providers and strategic partners through structures that sit alongside the equity rounds, not within them. This connects directly to the void-shares story in Section 6.

5. The SpaceX Compute Deal

On May 20, 2026, SpaceX disclosed in its S-1 IPO filing that Anthropic will pay it $1.25 billion per month through May 2029 for access to the Colossus 1 data centre in Memphis, Tennessee, and an expansion to Colossus 2 [22, 23, 24]. The deal is approximately $45 billion in total contractual value. Colossus 1 holds over 220,000 NVIDIA GPUs across H100, H200, and GB200 generations, with 300 megawatts of capacity. Either party may terminate the contract with 90 days' notice. Anthropic also expressed interest in working with SpaceX to develop multiple gigawatts of compute capacity in space.

The structural facts of the deal are worth dwelling on. Colossus 1 is operated by xAI, Elon Musk's frontier AI company, which is one of Anthropic's most direct competitors. xAI had reportedly been running Colossus 1 at approximately 11% utilization, having migrated its primary training operations to the newer Colossus 2 facility. Anthropic is, in effect, filling the idle capacity of a competitor's underutilised data centre [22, 25]. SpaceX recognised the payment in its IPO filing, which suggests the commercial structure flows through SpaceX rather than directly through xAI.

The deal must be read in the context of OpenAI's April 9, 2026 investor memo, which described Anthropic as “operating on a meaningfully smaller curve” and projected Anthropic at only 7 to 8 gigawatts of compute by end of 2027 versus OpenAI's planned 30 gigawatts by 2030 [26]. The Colossus deal is, in part, the answer to that critique: Anthropic's CFO Krishna Rao said the company was “making our most significant compute commitment to date to keep pace with this unprecedented growth” [26]. Six weeks later the Colossus deal was signed.

If all the disclosed compute commitments are honoured, Anthropic has bound itself to approximately $130 billion of compute and infrastructure spend over the next three to five years [22, 19, 20], against the disclosed May 2026 run-rate of $47 billion [57]. The implied gross margin available after compute is the central operating question for the company.

Three questions worth holding open. Who actually captures the economics of frontier AI? Historical precedent in semiconductors and cloud infrastructure suggests the picks-and-shovels suppliers often outperform the model-builders. What happens if the relationship with xAI deteriorates? The 90-day termination clause is the only structural protection. Has Anthropic reduced its risk by going multi-cloud, or increased its operational complexity? Both, probably; the question is the weighting.

6. The Void-Shares Controversy

On May 11, 2026, Anthropic published a notice that reads, in the company's own words [27]:

Any sale or transfer of Anthropic stock, or any interest in Anthropic stock, that has not been approved by our Board of Directors is void and will not be recognized on our books and records.

The notice extended explicitly to special purpose vehicles, forward contracts, beneficial interests, and tokenized securities. The company initially named eight platforms it considered unauthorized: Open Door Partners, Unicorns Exchange, Pachamama Capital, Lionheart Ventures, Hiive, Forge Global, Sydecar, and Upmarket [28]. After public pushback from several of the named firms, Anthropic quietly revised the list down to four: Open Door Partners, Unicorns Exchange, Pachamama, and Upmarket. Lionheart Ventures, Hiive, Forge Global, and Sydecar were removed [66]. Hiive's CEO publicly criticised the original notice for damaging his platform's reputation and noted Anthropic had not contacted them before publishing; Forge Global stated it had been included erroneously. The notice for the remaining four platforms stands.

Three points of legal substance. First, the choice of the word void rather than voidable is, in Delaware corporate law, the most aggressive position a company can take against an unauthorized transfer [27]. A voidable transfer can be undone but may be salvaged by good-faith doctrines or equitable defenses available to a downstream buyer who paid value without notice. A void transfer is treated as having never legally occurred. The crypto lawyer Gabriel Shapiro flagged the word choice as the most consequential element of the announcement [29].

Second, the notice specifically forecloses the SPV path that has been the dominant structure for outside investors in late-stage private companies for the past decade. The exact wording: “We do not permit special purpose vehicles to acquire Anthropic stock and any transfer of shares to an SPV are void under our transfer restrictions.” Anthropic's position is that no SPV qualifies unless the company has consented to it directly. That consent has been granted only to a small number of cap-table general partners.

Third, the notice is the company's position, not a court ruling. Whether a Delaware court would enforce the void designation against a good-faith arms-length buyer who paid value to an apparent owner of the underlying shares is a question that has not, to public knowledge, been litigated against Anthropic.

The market context: at the time of the announcement, implied valuations of Anthropic on token markets such as Ventuals and PreStocks had reached approximately $1.6 trillion, roughly double the price paid in the company's most recent priced round [30]. On Forge Global, the implied valuation sat near $1 trillion. One offer through Saints Capital implied $1.15 trillion.

A potential investor should know, before deploying capital, exactly which side of the company's transfer-approval line their structure falls on. The single-layer SPV from an approved cap-table GP, with the company's written consent in the file, is on the recognised side. The forward purchase agreement traded on an unapproved platform, the multi-layer feeder, the tokenised representation acquired on Ventuals: all of these are, by the company's stated position, void. Ownership is a legal status, not a brokerage screen.

7. The Department of Defense Dispute

Since January 2026, Anthropic has been in open conflict with the United States Department of Defense over the use of Claude for military and surveillance purposes. The dispute is unusually public for a private company's relationship with the Pentagon, and it has consequences for both Anthropic's revenue base and its strategic positioning.

The substance of the dispute. The DoD asked Anthropic to allow Claude to be used for “all lawful purposes,” a formulation that would have included surveillance and the development of lethal autonomous weapons. Anthropic refused. The DoD then designated Anthropic as a “supply chain risk” and an estimated $200 million in contracts was terminated [47]. Anthropic sued Defense Secretary Pete Hegseth and the Department of Defense in March 2026 [48].

The legal status. In March 2026, U.S. District Judge Rita F. Lin granted a preliminary injunction against the government, writing that “The Department of War's records show that it designated Anthropic as a supply chain risk because of its hostile manner through the press. Punishing Anthropic for bringing public scrutiny to the government's contracting position is classic illegal First Amendment retaliation” [49]. In April 2026, the D.C. Circuit Court of Appeals denied Anthropic's motion to lift the designation [50]. On May 19, 2026, the appeals court heard arguments on the broader merits; a decision is pending.

What this means for the investor. Three things. First, Anthropic's safety-research positioning is not marketing copy; the company has refused $200 million of contractual revenue rather than compromise it. That is a costly signal of values consistency. Second, regulatory risk is real and active: the largest customer in the US federal government has, at least temporarily, classified Anthropic as a supply chain risk, with implications for other federal agency procurement. Third, the legal posture is unresolved: the article cannot say how the litigation will end, only that the litigation exists and that the company's compliance posture is not the same as that of competitors who have accepted military-use clauses.

If Anthropic's commercial moat depends in part on regulatory alignment (the regulatory capture regime in Section 9), the DoD dispute is both validation of the alignment (the company is genuinely safety-first) and a counterexample (the regulator is not, in this instance, aligning with the company). The investor needs to know which way the regulatory question cuts.

8. AI-flation, AI-fatigue, and the return-on-AI question

The growth story is unambiguous. The customer-satisfaction story is not. Four named voices have, over the past four months, said the same thing from different vantage points: Anthropic's growth has outrun its capacity, its prices have outrun what customers will pay for the value delivered, and the products themselves have shown signs of strain.

The customer voice. In March and April 2026, Claude Code users mounted the loudest public-relations backlash Anthropic has yet faced [31, 32]. An AMD executive described the product as “unusable for complex engineering tasks.” Cybersecurity professionals warned of “dangerously degraded code quality.” The Register and Fortune covered the complaints; Anthropic publicly acknowledged the issues, called the situation “top priority,” and Fortune published a piece on April 24 in which Anthropic explained the engineering missteps that produced the performance decline [32].

The bill-payer voice. Microsoft cancelled its internal Anthropic Claude licenses with a June 30, 2026 cutoff [33, 34]. The pilot had launched in December 2025 and burned through its annual budget in months. Token-based billing exposed the true cost that flat seat licensing had hidden. Microsoft still resells Claude externally at a 65% premium; the cancellation was for free internal access. Uber's CTO sent an internal memo stating that Uber had burned through its entire 2026 AI budget in four months [35]. Procurement-data tracker Tropic recorded American AI software prices jumping 20% to 37% in the year ending Q1 2026 [35].

The company's own voice. At a developer event, Dario Amodei said: “We planned for a world of 10x growth per year. In Q1 2026, we saw 80x annualized growth per year in revenue and usage” [36]. New infrastructure takes 12 to 24 months to translate into capacity. The SpaceX compute deal lifted Tier 1 input tokens per minute by 1,500% and output tokens per minute by 900%, but the lift arrived after the customer complaints, not before [37].

The competitor voice. In an April 9, 2026 memo to investors, OpenAI described Anthropic as “operating on a meaningfully smaller curve” and characterised the company as compute-constrained. OpenAI projected Anthropic at 7 to 8 gigawatts of compute by end of 2027 versus OpenAI's planned 30 gigawatts by 2030 [26]. The memo was a competitive attack, but it landed on a factual base: every observable signal in the same period pointed to the same constraint. When customers complain, the company itself admits, the bill-payers walk, and the largest competitor publicly attacks on the same axis, the bear case stops being one perspective. It becomes a converging signal.

The SaaSpocalypse context. Some of the AI-flation pressure on Anthropic is a consequence of its own success. Anthropic's launch of Claude Cowork in early February 2026 was the immediate trigger of the largest AI-driven repricing event in software history. Approximately $285 billion was wiped from SaaS valuations in 48 hours; cumulative losses reached approximately $2 trillion [3, 38]. Salesforce, Monday.com, Workday, and Adobe fell 25 to 38%. Hubspot, Atlassian, and Figma fell 70 to 80% from their 52-week highs. Intuit fell 46%. Atlassian reported its first ever decline in enterprise seat counts. Workday cut 8.5% of its workforce [4, 5]. Klarna replaced its Salesforce and Workday installations with an in-house knowledge stack on Neo4j plus an OpenAI-built customer-service AI that automated approximately 700 agent roles [39, 40]. NVIDIA's Jensen Huang dismissed the SaaS-replacement narrative as “the most illogical thing in the world” [38]; the counterpoint deserves to be named even when the market is not yet pricing it.

The Socratic question this section forces on the reader is whether Anthropic's $965 billion valuation has internalised the AI-flation, the customer complaints, the competitor attack, and the supply constraint, or whether the price is built on the assumption that the next 18 months will look like the last 18 months. The historical base rate is that 18-month linear extrapolations of unusual growth phases are usually wrong.

9. Competition, Commoditization, and the Future

This section asks the hardest question in the article. The frontier AI race has had a different leader every twelve to eighteen months. OpenAI was the AI company of 2023 to 2024. Google was, by consensus, the laggard whose response to ChatGPT was botched in early 2023; by late 2025, Gemini had executed one of the most remarkable comebacks in technology history. 2026 is, by every measure, Anthropic's year. What about 2027 and beyond?

Google's 2025 comeback, quantified. Gemini's share of generative AI web traffic rose from 5.4% in late 2024 to 18.2% in late 2025 [41, 42]. ChatGPT's share fell from 87.2% to 68.0%, a 19-point drop. Gemini surpassed 750 million monthly active users by Q4 2025. Google AI Overviews reach approximately 2 billion monthly users, a distribution advantage no other lab can match. 75% of Google Cloud customers are using AI services; 120,000 enterprises are on Gemini; 95% of the top 20 global SaaS companies use Google AI products [41]. Gemini 3 and the Ironwood AI chip launched in late 2025; the combination of frontier-capability model plus proprietary silicon plus Google-scale distribution is the strongest structural position any frontier lab has built. The fact that an incumbent that was widely written off in 2023 became the dominant distribution force in 2025 should give any 2026 entry-price investor pause.

Open-weight pressure, quantified. DeepSeek and Qwen combined market share rose from 1% to 15% globally between January 2025 and January 2026, the fastest adoption curve in AI history [43, 44]. Qwen 3.5 shipped in February 2026 as a 397 billion parameter mixture-of-experts model under the Apache 2.0 licence, with 256K native context, 201-language support, and vision capabilities that beat GPT-5.2 on math-vision benchmarks [44]. DeepSeek V4-Pro at 1.6 trillion total parameters and 49 billion active parameters uses approximately 27% of V3.2's per-token inference FLOPs and 10% of the KV cache at 1 million context [45, 46]. Open-weight quality has closed to within 5 to 15 points of closed frontier on standardised benchmarks; the procurement question has shifted from “is open-weight good enough” to “which open-weight, which hosting stack, which workloads.”

Meta's strategic admission. Meta has paused open-weight Llama releases and pivoted frontier attention to a closed Muse line [43]. This is a striking move: the company that defined open-weight at the frontier has stepped back from it, which suggests either that Meta has concluded open-weight cannot capture the economics of frontier capability, or that the strategic value of withholding capability has risen. Either reading is consequential.

The four-regime question. There are roughly four credible regimes for the frontier AI industry over the next decade.

Commoditization. Frontier capability becomes a commodity; differentiation moves up the stack to distribution, application, and customer relationship; pricing on raw API access compresses to the cost of compute plus a thin margin. In this regime, Anthropic's $965 billion price is supportable only if the company successfully migrates up the stack into a moat that does not depend on capability differentiation. The agentic products (Cowork, Claude in Chrome, Claude Code) are consistent with that migration; whether the migration succeeds before the commoditization completes is open.
Winner-take-most. One or two labs reach a capability cliff that competitors cannot cross within an investable timeframe. The lab on the right side of the cliff captures the bulk of the economic value of artificial general intelligence; everyone else captures fragments. In this regime, $965 billion is cheap, perhaps by orders of magnitude.
Specialization. Multiple labs survive, each serving a different segment. Anthropic occupies the enterprise-safety segment; OpenAI occupies consumer and general developer; xAI occupies social and political; Google occupies search-integrated; the open-weight tier occupies self-hosted and regional. In this regime, $965 billion is supportable if Anthropic's enterprise segment is large and durable enough.
Regulatory capture. Compute thresholds and government oversight create barriers that favour incumbents. Anthropic's safety-research positioning aligns it with the regulatory community in a way few other frontier labs match. Whether that alignment converts into regulatory protection is open. The DoD dispute in Section 7 cuts both ways on this question.

10. The Criticism of Reported Metrics

The revenue and profitability figures in Section 1 are not undisputed. They are the company's own headline numbers, repeated by the financial press, and they have been challenged in four distinct ways: by OpenAI in a leaked internal memo, by independent investigative reporting from Ed Zitron at wheresyoured.at, by PitchBook's research desk, and by the Wall Street Journal's own reporting on the sustainability of the profitability projection. None of the critiques have been resolved at the time of writing. All of them matter for what the $965 billion valuation is actually pricing.

ARR is a soft metric. Annualised recurring revenue, the headline number both Anthropic and OpenAI report, is not a standardised accounting concept. It typically means taking a single month or thirty-day window of revenue and multiplying by twelve. It is not audited for private companies. It does not have to factor in customer churn or in pricing changes that arrive after the snapshot. Two companies reporting "the same" ARR may be measuring very different things. Harrison Rolfes, PitchBook's senior research analyst covering Anthropic and OpenAI, characterised the current disclosure landscape as "an ARR accounting arms race ahead of their IPOs," and added that "neither is reporting on a basis that would survive a Big Four audit" [53].

The gross-versus-net dispute. The substantive accounting question is how Anthropic recognises revenue from cloud-reseller channels, principally AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry on Azure. Anthropic reports the full end-customer spend on a gross basis and books the partner's cut as an expense. This is permissible under ASC 606 if Anthropic is the principal in the transaction, and Anthropic asserts that it is. OpenAI, by contrast, recognises revenue from its Microsoft Azure channel on a net basis, after Microsoft's roughly 20% cut, because OpenAI's position is that Microsoft is the principal of the Azure OpenAI service. OpenAI's chief revenue officer Denise Dresser sent an internal memo, leaked to The Verge, arguing that the apples-to-apples comparison would shrink Anthropic's headline by approximately $8 billion [54]. Both companies' treatments are defensible under GAAP; both cannot be apples-to-apples. The reader comparing the two companies' revenue should know which methodology produced each number.

The affidavit and the public ARR. In March 2026, in the course of the DoD litigation covered in Section 7, Anthropic's chief financial officer Krishna Rao filed an affidavit stating under oath that Anthropic had generated "over $5 billion" in revenue across the company's entire history since 2021. The public ARR figures imply a very different total: Q1 2026 alone was reported at approximately $4.8 billion of revenue, and Q2 is projected at $10.9 billion. Reconciling the two requires believing that roughly 85% of Anthropic's lifetime revenue arrived in the most recent 100 days, or that the public ARR figures are not measuring what readers reasonably assume they measure. Ed Zitron, writing at wheresyoured.at, has documented the discrepancy at length and pressed it as evidence that the headline ARR has drifted further from booked revenue than the public reporting acknowledges [55]. The story has not yet been picked up by Reuters, the Wall Street Journal, or Bloomberg as of the time of writing. Treat it as serious investigative work that has not yet been mainstream-corroborated.

The operating margin and the temporary cost advantage. The Q2 2026 "first profitable quarter" claim, projected at $559 million of operating profit on $10.9 billion of revenue, coincides with the start of the SpaceX/Colossus 1 contract covered in Section 5. The $1.25 billion per month is paid at a reduced ramp rate during the early months and steps up to the full pricing thereafter. The Wall Street Journal noted in its own reporting on the profitability projection that the company "may not remain profitable for the full year" because planned expense growth, including the full SpaceX run-rate and continued infrastructure expansion, absorbs the temporary cost advantage [2].

Two additional concerns are flagged by Zitron and echoed by other critical coverage [56]. First, Anthropic's "operating profit" definition appears to exclude stock-based compensation, which is one of the largest real expenses at scaled technology companies and which public-market accounting would require to be included. Second, The Information reported in January 2026 that Anthropic had missed its own gross margin projection because inference costs ran approximately 23% above plan; Q1 2026 compute consumed approximately 71 cents of every revenue dollar. Both facts are consistent with a company whose margin trajectory is more fragile than the headline profitability suggests.

What this means for the reader. The publication does not take a position on whether Anthropic's methodologies are or are not defensible. Most of them are defensible under GAAP. The reader, however, should know that the headline numbers in Section 1 carry accounting choices, that the choices favour the headline, and that the same revenue and margin figures translated into public-market accounting standards would look materially smaller. The IPO disclosures, when they arrive, will resolve much of this. Until then, every number in Section 1 should be read with the question "measured how, against which definition" attached to it.

11. Closing: the Questions Worth Answering

This article is not a recommendation. The publication does not recommend. What follows is the set of questions the reader who is offered access to this round should answer for themselves before deciding.

On the structure of the claim. Am I buying an approved single-layer SPV from a cap-table general partner, with the company's written consent in the file? Or am I buying a structure that Anthropic's May 11, 2026 notice considers void?

On the reported metrics. Have I read Section 10 and decided whether the methodology behind Anthropic's headline ARR and operating margin changes what those numbers actually mean for me?

On revenue and the growth rate. What fraction of Anthropic's $47 billion disclosed annualised run-rate is enterprise API (sticky, high-margin), what fraction is consumer (commodity-priced), what fraction is reseller through AWS Bedrock, and what fraction is strategic-deal-driven? Is the 47x growth in 18 months evidence of opportunity capture or evidence of durable enterprise?

On profitability and operating leverage. The Q2 2026 projected $559 million operating profit is the first ever positive quarter. Is the trajectory of operating margin pointing toward the level that supports $965 billion, or is the price implicitly assuming several more years of growth at the current rate?

On compute commitments and capacity. Has Anthropic committed approximately $130 billion of compute spend over the next three to five years? What happens to training continuity if the xAI/SpaceX 90-day termination clause is exercised at an inopportune moment? Is the company still compute-constrained relative to OpenAI?

On the cap table. Who holds effective control of the company today, and how does that change at the Series H? What is the dilution exposure for a Series H investor at IPO, given the strategic-partner overhang and the compute commitments?

On the valuation history. The 241x increase in valuation from Series B to Series H is a chart of capital availability and growth conviction combined. Is my view of the future at least as optimistic as the marginal investor in Series H?

On peer comparison. When I value Anthropic against OpenAI, xAI, Google DeepMind, or the open-weight tier, am I comparing companies that sell the same product to the same customer with the same moat, or am I borrowing a multiple that flattens distinctions that matter?

On AI-flation and customer signals. Have I priced in the convergent signal from customers (AMD, cybersecurity professionals), bill-payers (Microsoft, Uber), the company itself (Dario's admission), and the competitor (OpenAI's investor memo)? Or am I treating those signals as noise around a positive trend?

On regulatory exposure. Has the DoD dispute increased or decreased my conviction in Anthropic's regulatory positioning? What is the downside if the appeals court rules against the company on the broader merits?

On the competitive future. Do I believe Anthropic's lead at the frontier is durable over a multi-year hold, given that OpenAI led in 2023 to 2024, Google led in 2025, and the open-weight tier has closed to within 5 to 15 points on benchmarks? Which of the four regimes (commoditization, winner-take-most, specialization, regulatory capture) do I think most likely, and is the $965 billion price compatible with my view?

On myself. What is my personal liquidity, time horizon, tax position, and risk tolerance? Does a private-market position with a multi-year lock-up, no information rights beyond the GP's disclosure, no voting rights, and exposure to a transfer-restriction regime I do not fully control fit my circumstances?

There are no right answers to these questions in this article. There are right answers for each reader, given each reader's circumstances and convictions. The publication's role is to make sure the questions are asked. The decision is yours.

Socrates on Investing is an editorial publication, not investment advice. Nothing in this article constitutes a recommendation to buy, sell, or hold any security or fund. The author may have a position in the assets discussed; specific positions, where material, are disclosed within each article. Past performance does not predict future results. Investing involves risk, including the risk of partial or total loss of principal.