THE HOOK
The company that built PyTorch is now running production AI on Google's chips, and Nvidia should be paying attention.
THE GIST
A new PyTorch-native version of the SGLang inference engine lets Google's TPUs serve models without a JAX port, and it's running in production at Meta today. A matching open-source vLLM build landed too, built with startup RadixArk.
THE TAKE
Nvidia's real lock-in was never the silicon, it was the software, and stripping out the painful JAX port turns a TPU into just another box you can rent. Meta gets cheaper inference plus a bargaining chip in every GPU negotiation, and everyone else gets proof the switch works. CUDA just sprang a leak.
OPINION
The telling detail is the roadmap: day-zero TPU support for new open models by early 2027, starting with DeepSeek V4.1 Flash, per Tom's Hardware. If that lands, picking hardware for inference stops being a religious decision and becomes a spreadsheet one, which is bad news for whoever has the fattest margins.
- Story type: Important story
- Impact: High
- Confidence: High (official announcement)
- Audience: Developers
- What changed: Google TPUs now serve AI in production at Meta through PyTorch-native SGLang, with a matching open-source vLLM.
- What to watch: Whether day-zero TPU support for new open models arrives by Q1 2027, and which other big PyTorch shops follow Meta.
- Reading time: 1 minute
