Crusoe signs multi-year deal to power Perplexity's AI training and inference

Crusoe on September 15, 2026, announced a multi-year partnership with Perplexity under which Perplexity will run its full model lifecycle on Crusoe Cloud, training frontier models on dedicated NVIDIA GB300 NVL72 clusters and serving them in production through Crusoe’s Managed Inference service.

The announcement, datelined San Francisco, also commits Crusoe to adopting Perplexity Enterprise Pro and Max for its 1,800 employees. According to the release, the deployment is meant to equip staff with web and internal knowledge search, multi-step research, data analysis, and access to frontier AI models.

The release characterizes Perplexity’s products as operating in real time for millions of users, an environment in which inference latency and throughput are the product rather than theoretical benchmarks. Crusoe said its Managed Inference service is built for production-grade inference powered by proprietary optimizations and engineered for performance and cost across popular models, and that Crusoe Cloud offers the operational depth and scale to match Perplexity’s growth.

Executive Statements

Crusoe co-founder and CEO Chase Lochmiller said the fastest-moving AI companies need infrastructure that keeps pace across the entire model lifecycle and scales with them as they grow. “Perplexity is building the future of how people find answers, and Crusoe is a key partner in making that possible, from the first electron to the last token,” Lochmiller said.

Perplexity co-founder and CEO Aravind Srinivas said that running AI at Perplexity’s scale means every millisecond of latency is felt by users. He described Crusoe as built around that constraint, calling it high-throughput, low-latency inference backed by next-generation hardware.

Dion Harris, senior director of HPC and AI Infrastructure Solutions at NVIDIA, said modern AI requires a unified computing architecture from research to deployment. Powered by NVIDIA GB300 NVL72 and NVIDIA InfiniBand, Harris said, Crusoe Cloud enables Perplexity to train, fine-tune, and serve frontier models in production with the speed and efficiency agentic AI demands.

The GB300 NVL72 Platform

The dedicated training clusters named in the agreement run on NVIDIA’s GB300 NVL72, which NVIDIA describes as a fully liquid-cooled, rack-scale system integrating 72 Blackwell Ultra GPUs and 36 Arm-based Grace CPUs. NVIDIA’s published specifications list 130 TB/s of NVLink bandwidth, 20 TB of GPU memory with up to 576 TB/s of bandwidth, 37 TB of fast memory, and 2,592 Arm Neoverse V2 CPU cores. Each GPU in the system receives 800 Gb/s of network connectivity through a ConnectX-8 SuperNIC IO module, working with either Quantum-X800 InfiniBand or Spectrum-X Ethernet networking platforms.

NVIDIA claims that AI factories built on the GB300 NVL72 deliver up to a 50x overall increase in AI factory output performance compared with Hopper-based platforms. The company attributes that figure to a stated 10x improvement in user responsiveness, measured in tokens per second per user, and a 5x improvement in throughput per megawatt relative to Hopper, and notes the numbers are projected performance subject to change.

Managed Inference Service and History

The production-serving element of the agreement runs on Crusoe Managed Inference, which the company made generally available on November 20, 2025, for production inference workloads on Crusoe Cloud. The service is powered by Crusoe’s proprietary inference engine built around MemoryAlloy, a cluster-wide key-value cache that eliminates duplicate prefills by allowing GPUs to fetch prefix caches from local and remote nodes. Crusoe states the engine delivers up to 9.9x faster time-to-first-token and 5x higher throughput, benchmarked against vLLM for the Llama-3.3-70B model in a four-node deployment.

Crusoe offers the service in three deployment options, according to its Managed Inference product page. Serverless Inference offers usage-based consumption of open models through a fully managed API, aimed at early-stage workloads, low-volume traffic, and rapid experimentation. Self-Serve Deployments, which the page says are now generally available, run models optimized for throughput or responsiveness, including models customized with Crusoe’s Serverless Fine-Tuning. Tailored Deployments pair customers with Crusoe’s team for a dedicated, benchmarked endpoint, an option the page directs at proprietary models.

Developers reach the service through Crusoe Intelligence Foundry, a hub where they can generate API keys, use managed endpoints tuned to each model, monitor performance metrics, and enable provisioned throughput for production-scale deployments. Crusoe’s model hub lists pre-configured open models from labs including DeepSeek, Google, Z.ai, OpenAI, Meta, Alibaba, and NVIDIA, with stated parameter counts and context lengths for each model.