anthropic leases musk's colossus 1 supercomputer for claude

musk hates anthropic. anthropic just leased his supercomputer anyway

anthropic and spacex signed a deal on may 6 for anthropic to lease all compute capacity at colossus 1 in memphis – 300+ mw, 220k+ nvidia gpus, online within the month. mostly going to claude pro and max capacity

the awkward part: elon musk has spent the past year publicly attacking anthropic. he called it "doomed to become the opposite of its name" and wrote in february that "anthropic hates western civilization." his tone softened wednesday after meeting with the anthropic team. musk also kept a kill-switch clause – spacex reserves the right to reclaim compute if anthropic's ai "engages in actions that harm humanity"

why it works for both sides anyway:
• anthropic gets operational hardware – not a future buildout – within weeks, plus relief on its compute crunch
• spacex monetizes a data center xai had just vacated (training already moved to colossus 2) and, weeks before its reported $1.75–2t ipo, repositions itself as an ai infrastructure provider, not just a launch company

anthropic hunting for compute
sixth major compute deals in roughly seven months. not coincidental – an openai investor memo in april called anthropic's compute strategy a "strategic misstep" and said it was "operating on a meaningfully smaller curve."
anthropic spent much of 2025 capacity-constrained, with reliability issues and stricter peak-hour limits.

three things stand out about the strategy:

1. multi-vendor. nvidia gpus, google tpus, aws trainium, all in parallel. hedges supply risk and pricing, but expensive to maintain three training stacks.

2. equity-for-compute swaps. amazon put up $25b for a $100b+ aws spend commitment. google put up $40b for $200b. capital in as equity, out as cloud spend over a decade. same structure microsoft pioneered with openai.

3. speed over orthodoxy. fluidstack (a london gpu startup) over hyperscalers for $50b. now spacex/colossus over building new. both prioritize time-to-online above everything else.

the gap. ~12 gw contracted on paper across all partners, but openai's memo pegs anthropic's actual operational capacity at 7–8 gw by end of 2027 vs openai's 30 gw target by 2030. today the gap is narrower: 1.4 gw running for anthropic vs 1.9 gw for openai. amodei thinks the industry needs 100 gw by 2028, 300 by 2029.