← Back to feed

The cost of Anthropic’s compute crisis is now on the record. The company is paying SpaceX $1.25 billion per month — $15 billion a year — under the Colossus 1 lease announced on May 6, with the contract running through May 2029. Total committed spend: roughly $45 billion.

That figure, reported by Axios and corroborated by Business Insider, makes the Colossus deal one of the largest infrastructure contracts in the history of enterprise technology. SpaceX, which merged xAI into its corporate structure in February, is expected to clear more than $2.5 billion in cash profit on the arrangement annually, per New Street Research analyst Antoine Chkaiban.

What Colossus 1 Delivered

The original announcement covered 300 megawatts of capacity at SpaceX’s Memphis, Tennessee facility — 220,000 Nvidia GPUs spanning H100, H200, and GB200 generations, all already energised and racked. Anthropic was able to bring the capacity online in weeks rather than the quarters a greenfield build would require. Claude Code rate limits doubled for all paid tiers immediately after the deal closed. Opus API limits were also raised, reflecting the 80-fold year-on-year growth in Q1 2026 that Dario Amodei called “difficulties with compute.”

Colossus 2 Is Next, With GB200

The expansion to Colossus 2 changes the story. SpaceX had retained the larger Colossus 2 facility for its own Grok model infrastructure after the Anthropic deal for Colossus 1 was announced. Now, per a signal from Tom Brown — one of GPT-3’s original authors, now at xAI — Anthropic is in discussions to expand its footprint there. Colossus 2 is understood to be a GB200 NVL72-class deployment, moving beyond the mixed H100/H200/GB200 profile of Colossus 1 to a pure Blackwell Ultra configuration.

GB200 clusters deliver substantially higher memory bandwidth and FP8 throughput per rack than the H100 hardware that still constitutes much of Colossus 1’s installed base. For Anthropic’s inference workloads — particularly the multi-step agentic chains that Claude Code runs — the upgrade path matters more than raw GPU count.

Why the Price Holds

Anthropic has made it work because the alternatives are worse. GPU lead times for H200 and B200 clusters are running at 52 weeks. Transformer substation permits take 18 to 36 months. The hyperscaler buildouts Anthropic has secured — 5 gigawatts from Amazon, a further 5 gigawatts from Google and Broadcom due in 2027 — deliver capacity in stages over years, not months.

Renting a fully energised supercomputer at $1.25 billion a month is expensive by any conventional measure. Against the revenue it enables — Anthropic is tracking $30 billion annualised, most of it from Claude Code’s token-heavy agentic workloads — the math holds. Claude Code alone was generating $2.5 billion in run-rate revenue by February.

The Longer Play

Both companies mentioned orbital compute in the original announcement. SpaceX is developing satellite-based data centres and filed plans with the FCC for an 88,000-satellite constellation. The Colossus arrangement gives Anthropic first-mover positioning in that roadmap — a hedge against terrestrial grid constraints that no competitor can replicate quickly.

The IPO calendar adds context. SpaceX’s roadshow for the xAI-integrated entity is reportedly targeting the week of June 8 at a $1.75 trillion valuation. A $15-billion-a-year Anthropic contract is the kind of revenue line that prices an infrastructure offering, not just an AI lab.

Key Numbers

• Monthly payment: $1.25 billion
• Annual rate: $15 billion
• Contract term: Through May 2029 (~$45 billion total)
• Colossus 1: 300 MW, 220,000 Nvidia GPUs (H100/H200/GB200 mix)
• Colossus 2: GB200 NVL72-class, previously reserved for Grok
• SpaceX cash profit on deal: ~$2.5 billion annually