Crusoe and Perplexity ink a multi-year deal for train-to-serve AI cloud

Crusoe said Perplexity will run its full model lifecycle on Crusoe Cloud — training frontier models on dedicated NVIDIA GB300 NVL72 clusters with InfiniBand, then serving them in production through Crusoe Managed Inference. Crusoe will also roll Perplexity Enterprise Pro and Max to its 1,800 employees.

Same-day HARDWARE primary that a major AI answer engine is putting train and serve on one energy-first GPU cloud built around NVIDIA’s latest GB300 racks, not just renting generic GPUs.

Crusoe announced a multi-year partnership with Perplexity on 15 Sep 2026 to power Perplexity AI workloads on Crusoe Cloud. Crusoe calls itself a vertically integrated AI infrastructure provider — it builds the energy, the data centers, and the cloud that trains and runs models, rather than only renting someone else’s hall. The company newsroom post, titled “Crusoe and Perplexity Announce Multi-Year Partnership Across Full Model Lifecycle,” is dated September 15, 2026, and datelined SAN FRANCISCO, CA. The same-day GlobeNewswire carry is the same company text. Those company pages are the filing event. These are company claims. This desk did not see a contract.

Scope, as Crusoe tells it: Perplexity will run its full model lifecycle on one platform — training frontier models on dedicated NVIDIA GB300 NVL72-powered clusters with NVIDIA InfiniBand, then serving those models in production through Crusoe Managed Inference. A frontier model here is one of the newest, largest general-purpose AI models a lab is training. GB300 NVL72 is NVIDIA’s high-end multi-GPU rack for large AI training — many of NVIDIA’s latest chips wired together in one cabinet. InfiniBand is a fast private network between those racks, so the chips can talk without waiting on ordinary ethernet. Managed Inference is Crusoe’s hosted service that runs finished models for live user traffic — the answers people see, not the overnight training job. File that train-then-serve picture as Crusoe’s. This desk did not sit on a cluster.

Crusoe will also adopt Perplexity Enterprise Pro and Max for its 1,800 employees, the company said, for web and internal knowledge search, multi-step research, data analysis, and access to frontier AI models. File the 1,800-employee roll-out and those two product names as Crusoe’s. This desk did not count staff or watch a seat provision.

Named voices on the company page: Chase Lochmiller, co-founder and chief executive of Crusoe; Aravind Srinivas, co-founder and chief executive of Perplexity; and Dion Harris, Senior Director of HPC and AI Infrastructure Solutions at NVIDIA. HPC here means high-performance computing — the big clustered machines used for training. File the names and titles as theirs, via Crusoe. Their quotes are color only and stay in Sources. The NVIDIA quote is attributed commentary on Crusoe’s page, not a separate NVIDIA press release, and this desk is not inventing an extra NVIDIA commercial commitment.

Perplexity About blurb on the same page, company — attribute, do not independently certify: founded in 2022; answers more than 1.5 billion questions monthly; flagship product Perplexity Computer described as a model-agnostic orchestrator of AI agents across tools, files, chips, and devices. Model-agnostic here means it can call more than one lab’s model, not only Perplexity’s own. File the 2022 founding, the 1.5 billion monthly questions, and that Computer line as Crusoe’s About box for Perplexity. This desk did not count queries.

Plain English for the rest of the card: GB300 NVL72 = NVIDIA’s high-end multi-GPU rack for large AI training. InfiniBand = a fast private network between those racks. Managed Inference = Crusoe’s hosted service that runs models for live user traffic. Frontier model = one of the newest, largest general-purpose AI models a lab is training. Vertically integrated here = Crusoe builds the energy, the halls, and the cloud, not only the rental GPUs.

PRIMARY here: Crusoe’s 15 Sep 2026 company newsroom — Tier A PRIMARY company source, the original record — plus the same-day GlobeNewswire company wire of the same release. The multi-year partnership, the dedicated NVIDIA GB300 NVL72 clusters and InfiniBand training path, Crusoe Managed Inference for production serve, the Perplexity Enterprise Pro / Max roll-out to 1,800 Crusoe employees, the Lochmiller / Srinivas / Harris names, the 2022 founding, the more-than-1.5-billion-questions-monthly line, and the Perplexity Computer About blurb are company-attributed. The NVIDIA quote stays attributed commentary, not a second NVIDIA newsroom. NOT claimed: a contract dollar value, a GPU count, exclusivity, independently verified query counts or employee counts, that this desk sat on a cluster or saw a contract, a stock tip, or investment advice. Distinct from the already-filed branch-energy-33m-series-b, google-cloud-singapore-engineering-center, salesforce-google-unify-agents, and openai-perplexity-gpt6-astra-systems.