SpaceX's S-1 IPO filing revealed Anthropic secured exclusive access to all 220,000 Nvidia GPUs at the Colossus 1 facility in Tennessee for $1.25 billion per month through May 2029. The deal includes an unprecedented clause allowing Elon Musk to reclaim compute if Claude engages in actions that harm humanity — creating a supply-chain risk with no precedent in commercial cloud contracts. Musk simultaneously competes with Anthropic via Grok while renting them the infrastructure they need to compete.
SpaceX S-1 Filing Reveals Largest AI Compute Deal in History — With a Clause No One Saw Coming #
The SpaceX S-1 IPO filing dropped a bombshell that rewrites the AI infrastructure landscape: Anthropic pays SpaceXAI $1.25 billion per month for exclusive access to all 220,000 Nvidia GPUs at the Colossus 1 facility in Tennessee. Through May 2029. That's $15 billion per year in recurring compute costs — more than Anthropic has raised in total since its founding. And Elon Musk reserved the right to pull the plug if Claude "engages in actions that harm humanity."
The deal is unprecedented in scale and in structure. No infrastructure arrangement for a private AI company has crossed $15 billion annually. And no cloud contract in history includes a clause allowing the provider to unilaterally terminate based on their assessment of a customer's AI behavior. This is not a standard SLA termination. This is a novel supply-chain risk with no precedent.
The Numbers That Matter #
Colossus 1 runs more than 300 megawatts of power across 220,000 Nvidia GPUs. Anthropic secured all of it — exclusive access, not a shared allocation. SpaceXAI retains Colossus 2, an even larger facility, for training its own Grok models.
At $1.25 billion monthly, Anthropic is committing $15 billion per year for infrastructure before it generates $15 billion in annual revenue. The company disclosed $2 billion in API and enterprise revenue for Q1 2026. Even at continued growth rates, the Colossus deal represents a significant fraction of expected income — possibly more than half. That means Anthropic is betting its future on sustained growth to cover this bill.
For context: $1.25 billion per month buys access to 220,000 H100-class GPUs. At current spot market rates for comparable compute, that is a competitive price. Dedicated, guaranteed capacity costs more than the public cloud — but public cloud does not offer 220,000 GPUs in a single contiguous cluster. The deal is expensive, but the alternative was not being able to train frontier models at all. Anthropic traded margin for certainty.
The "Harms Humanity" Clause: Real Risk or Marketing? #
The filing includes a provision allowing Musk to reclaim compute if Claude "engages in actions that harm humanity." This is not a standard contract term. Standard compute agreements include termination for payment failure, legal violations, or security breaches. A clause based on the provider's subjective assessment of whether an AI model "harms humanity" has no precedent in commercial cloud contracts.
The practical concern: Musk has publicly criticized Anthropic repeatedly. He called Fable 5 "editorially controlled" in an internal Microsoft meeting on July 16. He said competing models "make more sense" than Claude. He is simultaneously the CEO of SpaceXAI, which sells Grok — a direct competitor to Claude. He is the CEO of the company Anthropic pays $1.25 billion per month. The "harms humanity" clause gives him a subjective mechanism to terminate $15 billion per year of critical infrastructure.
Whether Musk would actually exercise it is a different question. Terminating Anthropic's compute would eliminate $15 billion in annual revenue for SpaceXAI — revenue that funds Grok development and Colossus 2 expansion. It would also trigger immediate legal action. Anthropic is too deep into the deal for a surprise shutdown to be painless for either party. But the clause exists. It is signed. It is in the contract.
The Strategic Paradox: Competing While Supplying #
Musk is competing with Anthropic for developers while renting Anthropic the machines it needs to compete. This is structurally weird. AWS does not build its own foundation models. Azure hosts OpenAI but Microsoft does not sell a competing consumer chatbot. Google Cloud competes with Google DeepMind, but they are the same company. SpaceXAI is the first case where the compute provider directly competes with its largest customer in the customer's core business.
From SpaceXAI's perspective, the deal makes commercial sense. Grok does not need to beat Claude for SpaceXAI to benefit from Claude's growth. Every Claude API call that runs on Colossus 1 generates revenue for SpaceXAI. If developers use Claude instead of Grok, SpaceXAI still makes money. The only scenario where SpaceXAI loses is if Anthropic moves its compute elsewhere — which, at 220,000 GPUs of dedicated capacity through 2029, is not happening.
What It Means for Anthropic's Future #
The deal resolves Anthropic's most urgent constraint — compute access for frontier training — through May 2029. That covers at least two more model generations beyond Fable 5. It confirms Anthropic has the infrastructure to remain a frontier lab for the next three years.
It also confirms three major risks: the "harms humanity" clause introduces a supply-chain vulnerability that no other AI lab faces, $15 billion in annual compute costs must be covered by API revenue and enterprise contracts that are not fully realized yet, and Colossus 1's criticized environmental record sits in tension with Anthropic's sustainability commitments.
For investors evaluating Anthropic's next funding round or eventual public listing: the Colossus 1 deal is both the reason Anthropic can remain competitive and the single largest dependency risk in its corporate structure. No other AI lab's compute is controlled by a direct competitor with a subjective termination right. The deal was signed before the filing. The clause is in the contract. The check clears every month. And every month, Anthropic's survival depends on Elon Musk not deciding Claude has crossed a line only he can define.
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
Anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei.
Chatbot An AI system designed to have conversations with humans through text or voice.
Claude Anthropic's family of AI assistants, including Claude Haiku, Sonnet, and Opus.
Compute The processing power needed to train and run AI models.