Thinking Machines Lab just locked in $65 million a year of GPU compute from Crusoe.
The two companies announced a managed inference deal this week: Crusoe runs the GPU infrastructure, and Thinking Machines Lab runs its models on top of it. Myle Ott, the lab's ML infrastructure lead, said Crusoe took the company "from evaluation to production quickly" and met the same standards it holds its own systems to. The deal is framed as freeing up money that would otherwise go toward building infrastructure, letting the lab reinvest in research instead. Neither side disclosed contract length or hardware specifics.
This is the trade nearly every well-capitalized AI lab is making: rent compute instead of building data centers, then spend the savings on training and research. For Crusoe, landing a customer that builds its own frontier models is useful proof that its managed inference product works at production scale, not just in a pilot.
Whether $65 million a year buys Thinking Machines Lab a real edge, or simply keeps the lights on for a few more months of training runs, is a number neither company is sharing.