Anthropic Leases xAI Data Center: A Controversial Deal Driven by Compute Hunger

Deal Background: Anthropic Secures All Compute from Colossus 1

At yesterday's Code w/ Claude event, Anthropic announced a deal that stunned the industry — an agreement with SpaceX/xAI to lease the "entire compute capacity" of its Colossus data center. It's one of the most high-profile infrastructure partnerships in AI today, and also one of the most controversial.

To be clear, Anthropic is getting Colossus 1, not the entirety of xAI's compute. xAI retains the larger Colossus 2 for training its own Grok models. Many initially misread this as xAI abandoning its in-house models, but that's not the case.

About Colossus's scale: The Colossus data center is a hyperscale AI training facility built by xAI in Memphis, Tennessee in 2024, completed at a staggering pace — going from groundbreaking to partial operation in roughly 122 days. The facility initially deployed approximately 100,000 NVIDIA H100 GPUs, later expanding in the Colossus 2 phase to roughly double that compute capacity. The construction speed was widely regarded as an engineering marvel, but it was precisely this "move fast" building philosophy that led to systemic environmental compliance issues.

Environmental Controversy: Why Colossus Is So Sensitive

The Colossus data center has an extremely poor environmental track record. The gas turbines installed at the facility initially operated without Clean Air Act permits or pollution control equipment, circumventing regulations by classifying them as "temporary equipment." Credible reports have linked the facility to increased hospitalizations in surrounding areas due to declining air quality.

Regulatory context: The U.S. Clean Air Act requires industrial facilities exceeding certain emission thresholds to obtain New Source Review (NSR) permits before operating and to install Best Available Control Technology (BACT). Classifying gas turbines as "temporary equipment" is a common regulatory workaround — temporary equipment is typically exempt from permanent emissions permit requirements. However, if equipment is effectively running continuously, regulators can reclassify it as a permanent installation and retroactively enforce compliance. Communities around Memphis already bear a disproportionately high industrial pollution burden, and Colossus's unpermitted emissions have further exacerbated local environmental justice concerns.

Even Andy Masley — a commentator known for pushing back against misleading claims about data centers (he's written pieces arguing that "AI water usage is a non-issue" and "data center land use is a non-issue") — was critical of Colossus:

I would absolutely not choose to run my compute at this particular data center.

At a time when the very existence of "AI data centers" is an extremely sensitive political issue (the recent protests in Utah being the latest example), Anthropic's decision to partner with this specific data center is undeniably a terrible PR move. For a company whose core brand identity revolves around AI safety and responsible development, the tension between this choice and its stated values is glaring.

xAI Model Retirement Drama: A Telling Timeline

An interesting detail: the night before Anthropic's announcement, xAI issued deprecation notices for multiple models, including Grok 4.1 Fast and others, giving users just two weeks to migrate.

SpeechMap developer @xlr8harder expressed strong frustration:

This is terrible @xai. I just spent time and money migrating to grok 4.1 fast, and you're sunsetting it less than half a year after launch with under two weeks' notice, with no fast/low-cost alternative migration path. I will never depend on your products again.

This raises an obvious question: were these retired models running on Colossus 1? If so, xAI sacrificed its own existing users' interests to lease compute to Anthropic — a serious blow to the credibility of its API ecosystem.

Elon Musk's Statement: Compute With Strings Attached

Elon Musk, who previously delighted in calling Anthropic "Misanthropic," posted after the deal was announced:

I spent a lot of time last week talking to the senior Anthropic team about how they ensure Claude is good for humanity, and I was impressed. After that, I agreed to lease Colossus 1 to Anthropic, since SpaceXAI has moved training to Colossus 2.

He then added:

Just as SpaceX launches hundreds of satellites for competitors on fair terms and pricing, we will provide compute to AI companies that take the right steps to ensure AI is good for humanity. We reserve the right to reclaim compute if their AI engages in behavior that harms humanity.

Supply Chain Risk: The Dangerous Precedent of Depending on a Competitor for Compute

That last sentence is the most significant part of the entire deal. Who gets to define "harms humanity"? If that judgment rests with Elon Musk personally, it creates an entirely new category of supply chain risk for Anthropic.

The deeper logic of compute weaponization: "Supply chain weaponization" is not a new concept in the semiconductor and tech industries — U.S. chip export controls on China, Dutch ASML lithography machine export bans, and other cases demonstrate that critical technology infrastructure can become leverage in geopolitical competition. Bringing this logic into intra-industry AI competition means that whoever controls compute can influence competitors' strategic decisions through terms of service, pricing strategies, or even threats to cut off supply. Academics call this kind of dependency "platform power" — when one player is simultaneously an infrastructure provider and a market competitor, other players face not just commercial risk but a loss of strategic autonomy. The "reclaim compute if it harms humanity" clause Anthropic has accepted effectively places its operational continuity at the mercy of a competitor's subjective judgment.

Consider this scenario: if Claude's response to a sensitive topic displeases Musk, or Anthropic takes a policy stance that conflicts with Musk's views, could compute supply be suddenly cut off? This dependency puts a company that sells itself on independence and safety under the potential control of a direct competitor.

From a broader perspective, this deal reflects a core contradiction in today's AI industry: the explosive growth in compute demand and the massive gap in infrastructure supply are forcing companies into compromises they would never consider under normal circumstances.

The structural imbalance of compute supply and demand: The root cause of the AI industry's compute supply-demand crunch lies in the exponential growth of compute requirements for large model training. Take GPT-4-class models as an example: the compute needed for training increased roughly 100x compared to GPT-3, and next-generation models demand even more. NVIDIA H100/H200 GPU production is constrained by TSMC's advanced process node capacity bottleneck, leaving virtually every top AI lab in the world facing "compute hunger." Although Anthropic has established deep partnerships with Google and Amazon and received substantial cloud compute commitments, dedicated training clusters under its own control remain scarce. This structural shortage is the fundamental reason Anthropic had to accept compute supply with political strings attached.

Anthropic's compute hunger has reached the point where it's willing to accept compute from a direct competitor — with political conditions attached.

This may also foreshadow a trend in the AI industry: compute is evolving from a mere commodity into a strategic weapon. Whoever controls compute infrastructure will wield influence over the entire AI ecosystem that extends far beyond technology itself.

Key Takeaways

• Anthropic reached a deal with xAI to lease all compute from the Colossus 1 data center, but the facility is mired in environmental violations controversy
• xAI abruptly retired multiple Grok models on the eve of the deal, giving users just two weeks to migrate and sparking strong developer backlash
• Elon Musk claims to reserve the right to reclaim compute if AI "harms humanity," introducing a new supply chain risk for Anthropic
• The deal reflects a severe compute supply-demand imbalance in the AI industry, forcing companies into painful compromises between their values and practical needs
• Compute is evolving from a simple commodity into a strategic weapon, with infrastructure controllers wielding influence that transcends technology itself