Anthropic locks in $9.1B compute deal with Bitcoin miner Riot
Anthropic is buying compute from an unlikely landlord: Riot Platforms, the Bitcoin miner that is repurposing its Texas power sites for AI — and the market is cheering.
Anthropic has struck a $9.1 billion deal with Riot Platforms, the Bitcoin mining company that recently began selling AI data center capacity, people familiar with the matter told Bloomberg. Riot disclosed Monday that it had signed a 20-year agreement to supply 191 megawatts of computing — enough to power roughly 143,000 homes at any moment — from its Rockdale, Texas campus to a "leading frontier AI" company; sources say that company is Anthropic. Riot shares jumped about 25 percent in after-hours trading, and the miner said the contract runs through June 2048 with two five-year extension options that could lift total revenue to $16.1 billion. Neither company has officially confirmed the counterparty: Riot declined to comment and Anthropic did not respond to requests.
The deal is the latest stop on Anthropic's compute-sourcing spree — it recently signed a $10 billion agreement with infrastructure startup Volta Infra and, in May, agreed to pay SpaceX nearly $45 billion over three years — and it follows the Theseus Infrastructure platform the lab launched with Macquarie and GIC, which we covered yesterday — Anthropic, Macquarie and GIC launch Theseus Infrastructure. What makes this one different is the seller: Bitcoin miners own power, land and substations, the three scarce inputs of the AI buildout, and Riot is the biggest name yet to pivot from hashing to hosting. Anthropic is effectively signaling it will buy capacity from whoever can plug it in — even a company whose core business is crypto — because at this stage of the buildout, access to power beats brand-name cloud.
The creator of Redis just shipped a native Metal inference engine for MiniMax's H3 multimodal model, and it runs on a Mac. Salvatore Sanfilippo's h3.c loads the 33-billion-parameter model's components into unified memory one phase at a time, so the transformer, vision encoder and decoders never have to coexist at once — a trick that keeps peak memory near 40 GB on Apple Silicon. On an M5 Max, the fast preset renders a 512×512 image in about 17 seconds, and an aggressive four-step path finishes in roughly 3.5 seconds. MiniMax itself celebrated the release, noting that open weights are what let "the creator of Redis" show up and port H3 to a laptop. It is a reminder that single-developer, no-framework inference engines — not just the big clouds — are how frontier open-weight models reach consumer hardware, and that the cheapest distribution channel in AI is still a talented programmer with an MIT license.
What to watch: whether Riot's Corsicana expansion follows the Rockdale template, and which miner signs the next big AI lease.
Are Bitcoin miners becoming the AI landlords of the decade — or is this the top of the pivot cycle? Tell us in the comments.
Sources: antirez/h3.c (GitHub)
