‹ rbtfl

Anthropic locks ~10GW across Amazon, Google/Broadcom and SpaceX

Up to $25bn with Amazon, a multi-gigawatt Google-TPU/Broadcom deal, and a monthly lease of xAI's Colossus 1 output

AI· active Whose Money·The Long Game ·5 takes · ·rbtfl upd Jun 25, 2026
post

The split

The same story, as told by newsrooms in different countries. Their words, attributed and linked.

United States

TechCrunch

“With dual hyperscaler deals, Anthropic now sits on pledged capital and roughly 10 gigawatts of reserved AI training power.”

tech / financialread the original ↗

United States

MindStudio (compute timeline)

“Anthropic agreed to pay roughly $1.25 billion a month to rent the entire output of xAI's Colossus 1 data center.”

infrastructure analysisread the original ↗

post

Summary

Alongside its $65bn Series H, Anthropic disclosed compute deals totalling roughly 10 gigawatts of reserved training power. Amazon expanded its commitment to as much as $25bn for up to 5GW of dedicated capacity, ~1GW online by end-2026. Google and Broadcom locked a further 5GW of next-generation TPU capacity beginning 2027, reported as a ~$200bn-over-five-years commitment. Separately, Anthropic agreed to rent the entire output of Spacex/xAI's Colossus 1 data centre at ~$1.25bn/month through May 2029, 220,000 H100-class GPUs and 300MW immediately. The pledges underwrite the near-$1tn valuation and rival Openai's Nvidia-anchored build-out.

By the numbers

  • ~10GW, total reserved training power.
  • up to $25bn, Amazon commitment, up to 5GW.
  • 5GW, Google/Broadcom TPU capacity from 2027.
  • ~$1.25bn/month, SpaceX Colossus 1 lease through May 2029.
  • 220,000, H100-class GPUs in the Colossus 1 deal.

Why it matters

Compute reservation has become the real moat: Anthropic's valuation rests as much on locked gigawatts as on revenue. The Google/Broadcom TPU tilt diversifies away from Nvidia and gives Google a strategic supply hook into a rival; renting xAI's Colossus output ties two competitors together through SpaceX.

What to watch

  • Whether the 2027 TPU capacity arrives on schedule.
  • Power and grid constraints on the Amazon 5GW build.
  • Whether Nvidia's share of frontier-lab compute keeps eroding to custom silicon.

The briefing, by email