Open‑weight models are becoming the foundation for the next AI ecosystem, and the US must compete rather than isolate itself. I’ve seen a similar story in 2013 when I co‑founded Mesosphere, built on Apache Mesos, and later released DC/OS. Our growth was halted when Kubernetes emerged.

Kubernetes proved that a neutral, open platform can attract a full ecosystem of networking, storage, observability, and deployment tools. Vendors built enterprise features around it, and the community could extend it freely, demonstrating that no single vendor can outpace the collective innovation.

Open‑weight models, unlike fully open source‑self, let users download and modify trained parameters while keeping data hidden. This has spawned a self‑hosting stack—vLLM, SGLang, llama.cpp, Ollama, MLX—that powers Hugging Face’s 2M+ models. Developers now create quantized weights, LoRA adapters, and runtime adaptations for silicon and purpose.

The current debate over restricting Chinese open‑weight models risks cutting the US off from a growing ecosystem where 41% of downloads come from China. Instead, the US should release its own frontier models, use procurement to demand portable systems, and set independent safety standards rather than impose a blanket ban.