Elon Musk Called Anthropic Evil. Then He Leased It 220,000 GPUs.

In February 2026, Elon Musk posted on X that Anthropic "hates Western civilization." On May 6, he announced that xAI no longer exists as a standalone company, dissolved it into SpaceX under a new sub-brand called SpaceXAI, and leased the full computing capacity of the Colossus 1 supercomputer — 220,000 NVIDIA GPUs and more than 300 megawatts of power — to Anthropic. The same week, Musk was in a San Francisco courtroom testifying that xAI had partially trained Grok by distilling OpenAI's models, while suing OpenAI for betraying its mission to develop AI for the benefit of humanity. If you are keeping score: Musk is suing the company he says is evil for training on other companies' models, while leasing his GPU cluster to the company he said hates Western civilization, whose models his company trained on.

The deal is real. Anthropic confirmed that all 220,000 GPUs housed in the Colossus 1 facility in Memphis, Tennessee — including NVIDIA H100, H200, and next-generation GB200 chips — will be online within a month. The Elon Musk Anthropic deal is either the most cynical transaction in recent AI history or the most rational one, depending on what you think motivation is.

What Actually Happened

xAI was founded in 2023 with a stated mission to defeat ChatGPT and build an alternative to what Musk described as the ideologically compromised AI labs. It built Colossus 1 in Memphis in record time — a factory shell converted into one of the world's largest GPU clusters in under a year. It trained Grok 3 and Grok 4 there. It expanded to Colossus 2, now housing 555,000 GPUs at an estimated cost of $18 billion.

Then, according to data from DataCenter Dynamics, xAI's Colossus model compute utilization sat at approximately 11 percent — against the roughly 40 percent achieved by its rivals. The world's most powerful supercomputer was running at one-ninth capacity. Musk's explanation was direct: SpaceXAI had already migrated its training workloads to Colossus 2, which made Colossus 1 available. The Anthropic deal was a logical use of idle hardware.

What the 11 percent figure actually reflects is that Grok failed to generate developer demand at scale. A training cluster at 11 percent is not a scheduling problem. It is a demand problem. Anthropic's problem is the opposite: its annualized run-rate revenue surged from $9 billion at the end of 2025 to more than $30 billion by March 2026 — a trajectory that CEO Dario Amodei described as far exceeding internal projections. The company projected 10x growth and got significantly more. Compute became the binding constraint on how fast Claude could grow.

Musk had 220,000 idle GPUs. Anthropic needed 220,000 GPUs urgently. The transaction required both parties to set aside approximately three months of mutual public contempt.

The Irony Is the Story

The Elon Musk Anthropic relationship has been, until this week, defined almost entirely by hostility. Musk co-founded OpenAI in 2015, departed in 2018, sued it in 2024, and has spent much of 2025 and 2026 in active litigation against Sam Altman. His stated objection is that OpenAI abandoned its nonprofit mission and now operates in the interest of Microsoft and its own profit.

Anthropic was founded by Dario Amodei, Daniela Amodei, and other former OpenAI researchers in 2021. It has positioned itself as the most safety-conscious major AI lab — the company with Constitutional AI, detailed system cards, and deliberate restrictions on capability deployment. Musk's public posture toward Anthropic has alternated between dismissal ("Anthropic is just OpenAI 2.0") and hostility ("hates Western civilization").

In the courtroom during the OpenAI trial, when pressed on which AI company was most capable, Musk ranked Anthropic first — ahead of OpenAI, ahead of Google. The same week, he signed the compute lease.

His explanation after the deal was announced: "No one set off my evil detector." In February, he had written that Anthropic "hates Western civilization." The gap between those two statements is approximately 90 days.

The less satisfying but more accurate reading is that the Elon Musk Anthropic deal is simply a business transaction between a company with stranded compute and a company with stranded demand. The ideological framing was always more useful as marketing than as strategy. When the economics align, the framing adjusts.

What This Means for Anthropic

The compute constraint was real and documented. Anthropic had been publicly managing Claude rate limits for months, throttling API access and developer usage despite demand that continued to accelerate. The mismatch between what users wanted and what the infrastructure could deliver was costing Anthropic revenue at the precise moment it was trying to establish enterprise market share against OpenAI and Google.

Colossus 1 is not a permanent solution. New infrastructure investment — Amazon's Trainium2 and Trainium3, Anthropic's own buildout — takes 12 to 24 months to translate into available capacity. Colossus 1 bridges that gap. The 220,000 GPUs coming online within a month represent a near-term capacity surge that no amount of infrastructure planning could have delivered faster.

The secondary element of the deal — an expressed interest in developing orbital AI compute capacity together — is the more speculative but potentially more consequential detail. SpaceX's Starlink infrastructure gives it a deployment path for space-based data centers that no other company has at comparable scale. If orbital compute becomes viable in the 2027-2029 window, the company that controls that infrastructure will have an advantage that terrestrial competitors cannot easily replicate. Anthropic signing onto an exploration of that possibility, even informally, suggests it is thinking about compute supply chain beyond the current GPU land grab.

What This Means for xAI and Grok

The dissolution of xAI as an independent brand is the more quietly significant half of this announcement. SpaceXAI is not a rebranding exercise — it is a structural admission that the independent AI lab strategy did not produce the differentiation Musk wanted.

Grok has users. By early 2026, it had reached approximately 64 million monthly active users, and its U.S. market share grew substantially through 2025. But developer adoption — the metric that determines whether a model becomes infrastructure — lagged. The 11 percent Colossus utilization is the number that explains why xAI is now a SpaceX sub-division rather than a standalone company competing with OpenAI.

TechCrunch asked the relevant question directly: is xAI becoming a neocloud? A neocloud is an infrastructure provider that rents GPU capacity to AI companies rather than building its own model products. CoreWeave, Lambda Labs, and similar companies operate this way. The Anthropic deal could be read as SpaceX moving in that direction — using Colossus 1 as a revenue-generating asset while Colossus 2 handles its own AI development.

If that is the trajectory, it represents a significant narrowing of ambition. Musk launched xAI to beat OpenAI. The outcome, so far, is a GPU rental agreement with Anthropic and a sub-brand inside a rocket company.

The Week That Summarized Everything

This particular story landed in a week that also included Musk testifying under oath that xAI trained Grok by distilling OpenAI's models — the same practice he has cited as evidence of OpenAI's bad faith toward the open-source AI community. The simultaneous suing-OpenAI-while-copying-OpenAI-models position is difficult to reconcile on its own. Adding the lease of 220,000 GPUs to a company he called evil three months earlier completes a picture of AI industry incentives overriding stated principles at every turn.

None of this means the Anthropic-SpaceX deal is bad. For Claude users and developers who have been hitting rate limits for months, additional compute capacity is welcome news regardless of who owns the hardware. For Anthropic, Colossus 1 provides the fastest path to closing the gap between demand and supply. For SpaceX, idle infrastructure earns revenue.

The deal is rational. The irony is that rationality is the only framework that makes it coherent. The ideological positions that Musk and Amodei have both staked out in public — about safety, about mission, about what kind of AI development is legitimate — none of those positions determined the outcome. The 11 percent utilization rate did.

For anyone tracking how AI infrastructure actually gets distributed, the Elon Musk Anthropic compute deal is the clearest signal yet that the GPU supply chain is becoming its own geopolitical layer — one where the lines between competitors, customers, and ideological opponents blur completely. The companies that can capture compute when it's available, from wherever it's available, will determine which models reach users at scale. Tools that help capture and synthesize knowledge across fast-moving events like this one are increasingly necessary to follow what's actually happening beneath the headlines.