Perplexity just moved its entire AI model lifecycle, training through production inference, onto Crusoe, a cloud provider still mostly unknown outside AI infrastructure circles. The company was already running workloads on AWS and other major providers. It went looking for Crusoe specifically because Crusoe could deliver something AWS apparently couldn’t fast enough: access to Nvidia’s newest GB300 chips.
That’s really the whole story here, even though the announcement dresses it up with plenty of other detail. Perplexity will train models on Crusoe’s dedicated GB300 NVL72 clusters and serve them through Crusoe’s Managed Inference service, consolidating work that currently spreads across multiple vendors. Neither company disclosed the deal’s value, so there’s no way to compare it directly against Crusoe’s roughly $13 billion Jane Street contract signed earlier this month. What’s clear is that scarce, cutting-edge Nvidia hardware is now dictating which cloud provider AI companies choose, not brand recognition or existing vendor relationships.
Perplexity CEO Aravind Srinivas made the priority explicit: at Perplexity’s scale, every millisecond of latency reaches users directly, and he described Crusoe’s infrastructure as built specifically around that constraint. Crusoe CEO Chase Lochmiller framed it as supporting Perplexity across the model’s full lifecycle, training through the last token served.
The partnership runs both directions, too. Crusoe is rolling out Perplexity’s Enterprise Pro and Max tools across its own workforce of more than 1,800 employees for internal research and knowledge search, while Crusoe takes on all hardware maintenance and operational oversight for Perplexity’s infrastructure, reducing what Perplexity’s own team has to manage directly.
Crusoe built its name constructing large data centers, including the Stargate project for OpenAI in Abilene, Texas, before pivoting toward leasing AI compute directly to enterprise customers. That shift puts it in direct competition with CoreWeave, Nebius and the major hyperscalers. The path hasn’t been entirely smooth. OpenAI and Oracle both passed on expanded Abilene capacity, and Google removed Crusoe from a Wyoming data center project. None of that seems to be slowing demand, though. Advanced Nvidia hardware remains scarce enough that access to it, not brand loyalty, is deciding where AI companies actually put their workloads.
