Anthropic is in early talks to raise at least $30 billion at a pre-money valuation above $900 billion, according to Bloomberg. The Financial Times put the upper bound at $50 billion. No term sheet is signed; the round could close by the end of May. If it does, Anthropic crosses OpenAI's $852 billion mark from March and becomes, on paper, the most valuable private AI company in the world.

The number is the headline. The number is not the story.

What the number is paying for is increasingly compute and distribution held on terms Anthropic does not control. The premium isn't priced in basis points. It's priced in governance authority.

What the valuation is paying for

Look at the stack Anthropic is currently buying its way through.

Compute. Up to 5 GW from Amazon, with the first gigawatt online by end of 2026. Up to 5 GW from Google and Broadcom starting 2027. $30 billion of Azure capacity with Microsoft and Nvidia. $50 billion with Fluidstack. And the deal that closed last week: the entire Colossus 1 facility from SpaceX (formerly xAI) in Memphis — 300 megawatts, 220,000 Nvidia GPUs, online within a month of signing. Inference workloads, primarily. Claude Code rate limits doubled the day it was announced.

The Colossus 1 deal carries a reclaim clause. Elon Musk has stated SpaceXAI reserves the right to reclaim the compute if Anthropic's systems "engage in actions that harm humanity," adjudicated unilaterally by him. Simon Willison flagged it the day after the announcement, calling it "a new form of supply chain risk." Akhil Rao threaded the implications the same week.

This is the first time a hyperscale compute supplier has asserted contractual judgment over a renter's outputs. Compute as enforcement, not just supply. And Anthropic — the company most identified with safety as a brand commitment — is the renter.

Distribution. The parallel JV with Blackstone, Hellman & Friedman, Goldman Sachs, with Apollo, General Atlantic, GIC, Sequoia behind it. Smaller than OpenAI's Deployment Company ($1.5B vs $4B). Different shape: bank partners as both investors and customers. Paired with a 10-agent finance library and the Opus 4.7 launch. The pitch is reasoning-traceable agents inside regulated workflows. The structure is the same as OpenAI's: pay PE to route product into portfolios.

Regulatory standing. The European Commission accepted OpenAI's offer of EU access to GPT-5.5-Cyber under the new EU Cyber Action Plan, fronted by former UK Chancellor George Osborne. Anthropic — having released Mythos a month ago — has had "four or five meetings" with Brussels without reaching the stage of discussing access. Spokesperson Thomas Regnier said the contrast in public on Monday: "With one (OpenAI), you have a company proactively offering to give access. With the other one (Anthropic), we have good exchanges though we're not at a stage where we can speculate on potential access or not."

That is the first time the Commission has explicitly compared the cooperation postures of the two labs in public. AI Office Director Lucilla Sioli noted that once the AI Act's enforcement powers activate on August 2, Anthropic will be subject to her office's jurisdiction regardless.

So: compute rented from a competitor with a reclaim clause, distribution rented through a JV with banks at PE-premium prices, and regulatory access trailing the main rival in the world's largest unified market.

That is what $900 billion is paying for.

The revenue is real

This isn't a bubble argument. Anthropic's annualized revenue went from roughly $9 billion at end of 2025 to over $44 billion by May 2026. Gross margins are reportedly above 70%. Over a thousand enterprise customers spending more than $1 million per year. Eight of the Fortune 10. The reason Colossus 1 had to come online inside a month is that Claude Code and the API were hitting capacity. The demand is the constraint.

In February, Anthropic closed a $30B Series G at a $380B post-money — at the time the second-largest private financing in history. Three months later, a round at $900B pre-money would more than double that price. The pace is what's drawing the money. The structural questions don't dent the revenue.

But the revenue lives inside a structure where the people supplying the compute can pull the lease, the people pushing the distribution have a contractual return to defend, and the regulators write the enforcement rules in August.

The shape of the week

This is the third brief in a row turning on the same axis.

Monday's piece named the spine: the vertical stack is consolidating at both ends and governance authority is migrating with it. Tuesday's brief — the price of renting distribution — read OpenAI Deployment Company's 17.5% PE return guarantee as roughly 2x the standard PE preferred return: a premium paid to rent distribution fast rather than build it slow. Today is the compute mirror. Anthropic is paying its own kind of premium — not in basis points but in supplier authority — to rent inference under a clause that lets the seller pull the plug for cause the seller defines.

A reader compressed Tuesday's arc to five words: portcos are the corpus, not the customer. Substitute "GPUs" for "portcos" and the line still works. Colossus 1 isn't capacity Anthropic owns. It's a lease where the landlord is also a competitor and the eviction clause is ideological.

The pattern that's resolving: governance authority is migrating to whoever holds the leases. Compute providers get to assert it (reclaim clause). Distribution partners get to assert it (return guarantees that force volume). Regulators get to assert it (AI Act enforcement powers activate August 2 and the labs that cooperated walk in warm). The middle of the stack — the model — is where the valuation sits. The ends of the stack are where the authority sits.

On the cyber-AI side

OpenAI launched Daybreak on Monday, framed explicitly as the answer to Anthropic's Mythos. Three model tiers under the Trusted Access for Cyber framework: GPT-5.5, GPT-5.5 with Trusted Access, and GPT-5.5-Cyber for red teaming and penetration testing. Launch partners include Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, Sophos, and Zscaler. The pitch is defender-first and secure-by-design — the inverse of Mythos's "too capable to release" posture.

Independent skepticism of Mythos has hardened in the last week. Daniel Stenberg (creator of cURL) had Mythos run against his codebase via Project Glasswing. The third party reported five vulnerabilities; one was real, three were false positives, one was a minor bug. His conclusion: "primarily marketing." Heidy Khlaaf at AI Now flagged the absence of false-positive metrics and comparisons against established static-analysis tooling. AISLE showed a 3.6B-parameter open-weights model could replicate much of Mythos's flagship analysis at $0.11 per million tokens. Point Estimate's evaluation found Mythos Preview "one to two months ahead of the trend line" on capabilities and "essentially tied with GPT-5.5" on cyber benchmarks.

Three days ago, Anthropic opened a public HackerOne bug bounty for Claude and internal systems. Paired with the Mythos restricted-release posture, it reads as an admission that the model that allegedly finds critical vulnerabilities at frontier scale still needs human researchers to find them in the company's own infrastructure.

Daybreak under Trusted Access is OpenAI's bid to make defensive cooperation the regulatory entry point. Mythos's restricted release is Anthropic's bid for a different governance model. Brussels is choosing between them in real time. So far the choice is visible.

Also worth watching

- Anthropic's 10-agent finance library — announced alongside the JV, agents not yet publicly identifiable. The JV's economic logic depends on these agents slotting into bank workflows. - The Commerce Department deletion of the AI pre-release security-testing page (May 5, Microsoft/Google/xAI) still has no explanation. Eight days. Watch for congressional question or FOIA. - Cursor and SpaceX — the conditional $60B takeover bid sits in the background of every Anthropic-SpaceX story. Cursor also uses Colossus capacity.

What I'm watching for

Whether Anthropic responds publicly to the EU contrast. Whether the $30B prices at $900B pre-money or drifts toward the FT's $50B / $950B upper band. Whether anyone inside Anthropic addresses the reclaim clause as a governance matter rather than a contract detail.

The valuation is impressive. The stack underneath it is rented. Both can be true. The question is who decides when the lease runs out.

---

Sources

- Bloomberg, Anthropic in talks to raise $30 billion at $900 billion valuation (May 12) - Datacenter News, Anthropic signs SpaceX compute deal (May 8) - IDC, Anthropic, SpaceXAI, and the new compute race (May 12) - Simon Willison, Notes on the xAI/Anthropic data center deal (May 7) - Reuters via Yahoo, EU says OpenAI offers to open access (May 11) - IAPP, OpenAI grants European Commission access (May 12) - Infosecurity Magazine, OpenAI Launches Daybreak (May 12) - The Register, Anthropic's bug-hunting Mythos was greatest marketing stunt ever (May 11) - The New Stack, Anthropic puts the "myth" in Mythos (May 10) - Point Estimate, How good is Mythos? (May 5)