Crusoe is moving its AI infrastructure pitch beyond rented GPU access by adding managed fine-tuning and inference services for open-weight models.

Data Center Knowledge reported that the new Serverless Fine-Tuning and Self-Serve Deployments services will sit inside Crusoe Intelligence Foundry.

The services are aimed at teams that want to adapt open-source foundation models, deploy managed inference endpoints or export fine-tuned model weights without provisioning GPU clusters directly.

Crusoe Intelligence Foundry Adds Serverless Fine-Tuning

The new fine-tuning service lets customers bring data to open-weight foundation models and receive completed weights in the open .safetensors format.

Data Center Knowledge reported that customers can deploy those weights on Crusoe or move them to another platform, making portability part of the product claim.

Erwan Menard, Crusoe's senior vice president of product, said in the report that enterprises are moving towards model ownership instead of relying only on proprietary APIs.

He said AI-native companies are feeding production data back into open-weight models regularly as they try to improve performance and reduce inference costs.

Menard said demand for continuous fine-tuning is accelerating faster than Crusoe expected, particularly among teams building production AI agents where model predictability and data ownership affect procurement decisions.

The platform currently supports a curated library of open-weight models including Qwen, DeepSeek, Gemma and GPT-OSS.

Data Center Knowledge Cites IDC On Competition Beyond GPU Access

Dave McCarthy, research vice president at IDC, told Data Center Knowledge that raw GPU access was the dominant story for about 18 months but is no longer enough by itself.

He said enterprise buyers are looking at fine-tuning pipelines, evaluation, deployment tooling and inference optimisation as one system.

McCarthy also said providers that only sell chips risk becoming interchangeable.

His comments framed the Crusoe launch as part of a wider shift in AI data centre competition from capacity supply towards full model-lifecycle platforms.

Data Center Knowledge described portability as another procurement issue.

McCarthy said portability is no longer optional for enterprise buyers, while Menard said organisations using open-weight models increasingly expect to keep their fine-tuned weights rather than stay locked to one inference platform.

The inference side of the launch uses Nvidia H100 and H200 GPUs for managed endpoints.

Data Center Knowledge reported that customers will be able to deploy inference services without sourcing and configuring hardware directly, while fine-tuning jobs will be scheduled dynamically across Crusoe's AI infrastructure.

Crusoe Services Are Scheduled For General Availability Next Week

Data Center Knowledge reported that Crusoe plans to open both services to general availability next week through Intelligence Foundry.

The same report said the fine-tuning charge will use a per-million-token model and that managed inference will be charged by GPU hour.

Those billing details are tied to Crusoe's claim that customers can avoid reserving clusters in advance, although the company still did not provide production-load benchmarks or customer savings data.

Crusoe described automatic job restarts, checkpoint saving during training and billing that stops when a model stops improving.

The article also said traditional reserved-capacity models can leave expensive hardware idle because fine-tuning workloads are spiky.

Crusoe did not disclose named customers, per-million-token prices, GPU-hour rates, utilisation targets, benchmark methodology, service-level terms or customer-verified cost savings for the new fine-tuning and inference services.