Infract

Infract is the company around [[VLLM|vLLM]] discussed in 148. 对游凯超3小时访谈：开源Infra、和模型Co-design 、“如果vLLM失败，我们会后悔一辈子”. [[YuKaichao|游凯超]] describes it as the organization needed to sustain open-source inference work once vLLM became important production infrastructure.

The source gives a practical reason for the company form. A volunteer community can review code and accept contributions, but it cannot reliably allocate full-time people, sign NDAs with strategic users, secure large cluster resources, do quarterly planning, or support customers under production pressure. Infract is therefore framed as a commercial support layer around Open Source AI Infrastructure, not as a replacement for vLLM’s open governance.

The episode says Infract explores endpoint service, BYOC, and strategic-customer cooperation. Its preferred business logic is to charge against token value created or costs saved, rather than simply selling engineering hours.

Connections

• [[VLLM|vLLM]] — core open-source project around which the company operates.
• [[YuKaichao|游凯超]] — co-founder and chief scientist in the source.
• [[PyTorchFoundation|PyTorch Foundation]] — governance home for vLLM’s trademark and community ownership.
• Open Source AI Infrastructure and Large Company Open Source Strategy — open-source sustainability context.
• AI Inference Cost Structure, Model-Infra Co-Design, and AI Infrastructure Full-Stack Moat — technical and commercial problem space.