THE COLOSSUS 1 DEAL — KEY FACTS
● Source: SpaceX S-1 IPO filing ahead of $1.75 trillion Nasdaq debut
● Deal: Anthropic secures all compute capacity at Colossus 1, Tennessee
● Scale: 300+ megawatts, 220,000+ Nvidia GPUs
● Price: $1.25 billion per month
● Duration: Through May 2029
● SpaceXAI retains: Colossus 2 (larger facility) for its own models
● Musk's clause: Right to reclaim compute if Claude "engages in actions that harm humanity"
● Conflict: Musk is simultaneously competing with Anthropic via Grok while renting them compute
What the SpaceX S-1 Revealed
SpaceX's S-1 SEC filing ahead of its $1.75 trillion Nasdaq debut disclosed that xAI lost $2.4 billion in Q1 2026 while spending $7.7 billion on capex, and revealed a deal renting 300 megawatts of compute capacity to Anthropic for $1.25 billion per month through May 2029. The Colossus 1 facility in Memphis, Tennessee houses 220,000+ Nvidia GPUs across more than 300 megawatts of power capacity. Anthropic secured exclusive access to all of it — not a share of it, all of it — while SpaceXAI retains its larger Colossus 2 facility for its own models.
At $1.25 billion per month, Anthropic is paying $15 billion per year for compute access. This is a significant fraction of Anthropic's total $7.7 billion raised to date — and it is a recurring monthly commitment, not a one-time capital expenditure. For context: $1.25B/month buys access to 220,000 H100-class GPUs. At spot market rates, those same GPUs would cost more. The deal is expensive but it gives Anthropic guaranteed, dedicated compute access through May 2029 — critical for training the frontier models that justify its $1.789 trillion implied valuation.
The "Harms Humanity" Clause — A Novel Supply-Chain Risk
The clause that drew immediate commentary: Elon Musk stated he was impressed after meeting with Anthropic's senior team, but reserved the right to reclaim compute if their AI 'engages in actions that harm humanity.' This is not a standard SLA clause. Standard compute agreements include termination provisions for payment failure, regulatory violations, or security breaches. A provision allowing the compute provider to unilaterally terminate access based on their own assessment of whether an AI model "harms humanity" is unprecedented in commercial cloud contracts.
The practical risk: Musk has already publicly criticised Anthropic — he called Fable 5 "editorially controlled" and "it doesn't make sense" in an internal Microsoft meeting on July 16. He is simultaneously the CEO of the company selling Anthropic compute and the CEO of SpaceXAI, which competes directly with Anthropic for developers and enterprise customers. The "harms humanity" clause gives him a subjective, unilateral mechanism to terminate $15B/year of Anthropic's critical infrastructure. Whether that risk is real or theoretical depends entirely on whether Musk would exercise it — but it exists 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. Grok does not have to beat Claude for SpaceXAI to benefit from Claude's growth. Open-sourcing the harness may expand the market for coding agents. SpaceXAI is unusually positioned to make money if that happens, even when developers choose Claude instead.
This is a structurally unusual position. In traditional markets, compute providers (AWS, Azure, GCP) do not compete directly with their largest customers in the customer's core business. Microsoft is the closest comparison — it has $5B invested in Anthropic while also building Copilot on Claude. But Microsoft does not have a clause allowing it to cut off Anthropic's compute if Microsoft decides Claude is harmful. SpaceXAI does. The deal is commercially rational for both parties right now — Anthropic needs the compute, SpaceXAI needs the revenue — but the dependency it creates for Anthropic is significant.
Colossus 1 Environmental Record
The deal also drew criticism over Colossus 1's poor environmental record. The facility has faced scrutiny for its power consumption and cooling infrastructure, including concerns about water use and grid impact in the Memphis area. Anthropic's sustainability commitments — which include carbon neutrality goals — sit in tension with using a facility that has attracted environmental criticism. Anthropic has not publicly addressed this tension.
What It Means for Anthropic's IPO Timeline
The Colossus 1 deal resolves Anthropic's most urgent constraint — compute access for training frontier models — through May 2029. This is long enough to cover at least two more major model generations (the likely successors to Fable 5 and whatever follows Mythos 5). For investors evaluating Anthropic's implied $1.789 trillion valuation: the deal confirms Anthropic has secured the infrastructure needed to remain a frontier lab for the next three years. It also confirms a $15B/year recurring compute cost that must be covered by API revenue, enterprise contracts, and future fundraising. The Colossus 1 dependency and the "harms humanity" clause are the two risk factors that any Anthropic prospectus would need to disclose — and that any serious investor should understand before the eventual Anthropic IPO.
Sources: SpaceX S-1 SEC filing · aibusiness.com · dentro.de/ai · The New Stack · Related: Nadella calls Fable 5 editorially controlled → · Fable 5 July timeline → · Grok 4.5 vs Claude Fable 5 →