Google and Meta are making it easier to run familiar AI software on Google’s chips. On October 20, Google open-sourced TorchTPU, a native PyTorch stack co-developed with Meta. The competitive prize is the engineering work that makes switching hardware expensive.
PyTorch is a framework developers use to build and run AI models. TorchTPU makes Google’s Tensor Processing Units a native device option alongside CPUs and GPUs. Google says existing workloads can move with minimal code changes, retaining familiar APIs and debugging tools.
TorchTPU reaches beyond a compatibility promise
Google also announced native PyTorch backends for vLLM and SGLang, engines that serve model responses. These already had TPU paths through JAX; the new step is letting developers customize and debug the serving stack in PyTorch.
There is a deployment behind the announcement. In the joint SGLang technical account, Meta’s PyTorch-TPU lead Zongwei Zhou says SGLang-TorchTPU is running in production at Meta. That establishes a concrete use case, without establishing that Meta has replaced all its Nvidia systems.
Familiar code still needs hardware-specific work
The SGLang plugin keeps the scheduler and API server while supplying TPU-specific attention, memory-cache and computation components. Its support matrix distinguishes experimental features from tested ones; function calling, for example, is marked experimental. Same-day support for new open models is a goal from the first quarter of 2027.
These are disclosures from the collaborating companies, rather than independent proof that TPUs are cheaper or faster for every workload. A working deployment does not remove the need to check model support, output quality, latency and operating costs.
Our reading: this challenges the effort required to try an alternative, a barrier NeonPulse covered in Amazon’s TorchNeuron push. Google’s announcement adds another supplier and a named production user. Nvidia’s rivals need customers to bring their software with them; making that journey easier could matter as much as another impressive chip specification.
