Anthropic’s SpaceX Deal: Compute Desperation, Mythos Shock, and the New AI Power Trade
The White House Moves Toward FDA-Style AI Review
On May 6, National Economic Council Director Kevin Hassett confirmed that the Trump administration is studying an executive order to mandate pre-release government review for powerful AI models. This is not a theoretical exercise—it is a direct, scrambling policy response to Anthropic's Claude Mythos Preview. While the White House simultaneously pushes to stress-test government systems by granting federal agencies access to Mythos, the Commerce Department's Center for AI Standards and Innovation has already quietly secured early-access agreements with Microsoft, Google, xAI, OpenAI, and Anthropic. The stated focus: cyberattack risk, chemical and biological misuse, and data integrity.
The regulatory trajectory is unmistakable. Frontier models are no longer being treated as mere software products; they are being fundamentally reclassified as dual-use cyber capabilities. For investors, this is the most consequential shift of the cycle.
Anthropic's Emergency Compute Deal With SpaceX's xAI
Anthropic has struck an agreement to consume all available compute capacity at SpaceX’s xAI Colossus 1 data center, absorbing over 300 megawatts and more than 220,000 NVIDIA GPUs—a blend of H100, H200, and GB200 accelerators—within the month. The company claims this sudden injection of capacity will directly relieve service pressure on Claude Pro and Max subscribers, effectively doubling Claude Code's five-hour usage limits and eliminating peak-hour throttling.
On paper, Anthropic's broader compute stack is already staggering: up to 5 GW committed with Amazon (including nearly 1 GW by the end of 2026), 5 GW lined up with Google and Broadcom beginning in 2027, $30 billion in Microsoft Azure capacity, and a $50 billion Fluidstack-backed U.S. infrastructure war chest. The company even nodded to the future by expressing interest in developing multiple gigawatts of orbital AI compute with SpaceX.
Yet, despite that mountain of capital and vendor commitments, Anthropic still had to commandeer all of Colossus 1 on impossibly short notice. That fact alone is the tell.
Claude Mythos: A Cybersecurity Inflection Point
The catalyst for much of this urgency is Anthropic's unreleased Claude Mythos Preview. Briefly exposed via a misconfigured CMS, the company described the model as finding "thousands" of high-severity vulnerabilities across every major operating system and web browser. Critically, Mythos can chain these vulnerabilities into novel attack vectors that outpace human security researchers. At the time of disclosure, over 99% of its findings remained unpatched. Access is currently locked down to roughly 50 organizations under a heavily controlled program dubbed Project Glasswing.
The political fallout was immediate. Anthropic CEO Dario Amodei met directly with White House Chief of Staff Susie Wiles on April 17, mere days after the preview's release. Meanwhile, the Pentagon separately barred Anthropic from a classified AI deal it signed with SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle, citing supply-chain risks—an exclusion the White House has reportedly pushed back on with intense force.
Complicating matters, Reuters reported an incident of unauthorized access to Mythos through what Anthropic described as a possible third-party vendor environment. When a model is capable of autonomous zero-day exploitation, access control ceases to be an administrative footnote. It becomes the entire product.
The Revenue Signal Hidden in Plain Sight
Anthropic's run-rate revenue has eclipsed an astonishing $30 billion, up from roughly $9 billion at the end of 2025. In under two months, the number of enterprise customers spending over $1 million annually has more than doubled, jumping from over 500 to over 1,000.
That violent growth rate explains why Anthropic is currently operating like a public utility under an emergency load. However, investors must be careful not to conflate scale with margin clarity. Anthropic is now routing workloads simultaneously across AWS Trainium, Google TPUs, NVIDIA GPUs, Microsoft Azure, Fluidstack, and SpaceX xAI. While that architectural diversity is powerful, it breeds routing complexity, margin opacity, and deep supplier entanglement.
One quietly deployed detail compounds the margin picture: Anthropic's Opus 4.7 tokenizer may consume up to 35% more tokens for identical text. While the headline prices hold steady, the effective cost per workload is rising. Anthropic is pulling three levers at once—securing emergency compute, executing stealth tokenizer repricing, and enforcing rate-limit tier segmentation—all while the market looks the other way.
This Is No Longer a GPU Story
The shallow read of this week's news is simply that Anthropic secured more GPUs. The deeper, structural read is that frontier AI has entered a harsh new phase where product quality, compute access, cybersecurity risk, and government permission are inextricably linked—and the investment trade has rotated accordingly.
Anthropic is so severely capacity-constrained that it was forced to swallow politically awkward compute from a strategic competitor—one that also courts Cursor and holds the very Pentagon AI contracts from which Anthropic was excluded. SpaceX xAI's real product isn't Grok; it is deployment velocity. It is the demonstrated ability to brute-force massive infrastructure online faster than traditional hyperscalers by leveraging vertical integration and regulatory aggression. Meanwhile, Mythos has fundamentally changed how sovereign governments will classify and regulate frontier models. (And while the orbital compute language provides a strategically useful narrative for SpaceX, it remains entirely un-investable in the near term, plagued by unsolved physics constraints like heat rejection, radiation tolerance, and orbital hardware refresh economics.)
The best opportunities in this cycle are no longer found in parsing model benchmarks. They belong to whoever controls the physical and political bottlenecks: power rights, high-end NVIDIA supply, government trust, and the platforms capable of operationalizing AI-discovered vulnerabilities before adversaries exploit them. The AI trade has quietly mutated into a power and permissions trade. Most of the market just hasn't noticed yet.
not investment advice
Sources: https://x.com/KobeissiLetter/status/2052062576456949986