SpaceX's amended S-1 reveals Anthropic is paying $1.25 billion per month for access to 325,000 NVIDIA GPUs — the largest disclosed AI compute deal in history. That number reframes how much frontier AI infrastructure actually costs at scale, and who controls access to it.
- SpaceX disclosed Anthropic is paying $1.25B per month through May 2029 for access to COLOSSUS and COLOSSUS II compute infrastructure
- Water scarcity was added as a formal risk factor in the amended S-1 — a new disclosure not present in the original confidential filing
- SpaceX reserved 5% of IPO shares for employees and friends of executives, exempt from lock-up restrictions
- The AI segment spent $7.7 billion in capital expenditure in Q1 2026 alone, more than double the Space and Connectivity segments combined
What Happened
SpaceX filed an amended registration statement with the SEC on June 1, 2026, disclosing material changes ahead of its planned IPO. The filing added new risk language around water access for data center cooling, confirmed a directed share program reserving 5% of IPO stock for employees and persons selected by executive officers, and introduced a warning that future share issuances could dilute existing shareholders.
The detail that carries the most weight for AI buyers and enterprise decision-makers is buried in the related party transactions section. In May 2026, SpaceX entered into Cloud Services Agreements with Anthropic, granting access to approximately 325,000 NVIDIA GPUs across COLOSSUS and COLOSSUS II.
Anthropic has agreed to pay $1.25 billion per month through May 2029. Either party can terminate after the initial three-month period with 90 days’ notice.
That single contract puts a public number on what frontier AI compute costs at scale. For context, SpaceX’s entire Connectivity segment, which runs 10.3 million Starlink subscribers across 164 countries, generated $3.26 billion in revenue in Q1 2026. Anthropic is paying nearly that amount every quarter for GPU access alone.
Why It Matters
The water disclosure is new. SpaceX’s original confidential draft stated that data center growth was constrained by power availability and long construction timelines. The amended filing adds water access alongside power as a structural constraint, noting that drought conditions or regulatory restrictions could limit expansion. Financial disclosure analysts at The Dig flagged the addition as consistent with the type of disclosure the SEC typically prompts through comment letters during confidential review.
The directed share program grants participants shares at the IPO price with no lock-up restriction, meaning they can sell immediately after trading begins. All other shareholders face restrictions ranging from 180 days to over a year. SpaceX is targeting a $2 trillion valuation and plans to raise more than $75 billion, which would make it the largest IPO in US history if completed at those figures.
For AI teams evaluating compute strategy, the SpaceX S-1 is now the most detailed public disclosure of what hyperscale AI infrastructure costs to build and rent. SpaceX brought COLOSSUS online in 122 days and COLOSSUS II’s first cluster in 91 days at construction costs below industry benchmarks per megawatt. No public cloud provider has disclosed comparable build speed or unit economics at this scale.
“Water availability has become a critical consideration in data center site selection, development and operations.” — SpaceX, S-1/A Amendment No. 1
The $1.25 billion monthly figure runs through May 2029 in theory, but either party can exit after the first three months with 90 days’ notice.
Anthropic retains ownership of its models and data under the agreement. This is a capacity lease, not a long-term infrastructure commitment. Enterprise buyers benchmarking AI vendor stability against this number should read the termination terms first.
Bottom Line
The AI segment spent $7.7 billion in capital expenditure in Q1 2026 alone, up from $2.6 billion in Q1 2025. Anthropic’s own IPO filing is moving through SEC review in parallel, meaning both companies will soon be disclosing financials that have never been visible to the market.
For enterprise buyers currently benchmarking AI vendors, the compute cost structure behind frontier models is now partially visible for the first time. Teams building vendor assessments or negotiating API contracts should treat the $1.25 billion monthly figure as a floor reference, not a ceiling, for what it costs to run at this scale.
Relve is an AI tools and trends intelligence platform tracking what these disclosures reveal about the true cost structure of frontier AI as both S-1s progress toward public offering.
