Perplexity has handed its whole model pipeline to Crusoe, from training runs through to the inference calls behind each answer.
The multi-year agreement, announced September 15, puts frontier training on dedicated Nvidia GB300 NVL72 clusters inside Crusoe Cloud. Finished models then serve live traffic through Crusoe’s Managed Inference service. Neither company disclosed a dollar value.
Crusoe gets a customer as well as a client. Its own staff, 1,800 people, will use Perplexity Enterprise Pro and Max for web and document search, multi-step research and data analysis.
Speed is the argument on both sides. Srinivas, Perplexity’s co-founder, has said latency at his scale is something users feel rather than measure, and he pointed to newer silicon with serving tuned for it. Lochmiller, who co-founded Crusoe, frames his side as covering a model’s whole life, a span he describes as going from the first electron to the last token.
Nvidia’s Dion Harris, who leads HPC and AI infrastructure solutions, tied the deal to running research and deployment on a single architecture, with GB300 NVL72 systems and InfiniBand underneath.
Both companies are growing fast. Crusoe recently raised $3B at a $30B valuation and is expanding its Tulsa footprint, while Perplexity rents its compute rather than owning data centers.
