A Lab Bought Managed Inference Instead of Building It, in a Bite

An AI lab with a deep engineering bench rented its serving layer instead of building one. That is the story.

• Thinking Machines Lab signed a $65 million annual agreement with Crusoe to serve production inference for its open models.
• The work runs on a dedicated deployment of NVIDIA HGX B200 systems on NVIDIA Quantum-2 InfiniBand networking.
• Crusoe operates the cluster as a dedicated, benchmarked, SLA-backed endpoint. The lab pays for the throughput.
• Crusoe Managed Inference passed $100 million in contracted annual recurring revenue less than a year from launch.
• Training capacity is still bought with capital. Inference capacity is now shopped on price, latency, and reliability.

Read the full analysis here.