Google Open-Sources TorchTPU for Native PyTorch

Google used the PyTorch Conference 2026 in San Jose to open-source TorchTPU, a software stack that lets PyTorch models run natively on Tensor Processing Units. With the release, deploying a transformer to TPU in Hugging Face Transformers v5 reduces to a single argument, device_map equal to tpu, instead of the PyTorch/XLA bridge that previously required significant engineering effort.

The project was developed with Meta, PyTorch's original sponsor, and four repositories were published together: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan. Google also published a 23-page technical report describing a hybrid eager and compiled architecture. Upstream projects including vLLM, SGLang, Hugging Face, Ray and TorchTitan shipped integration at the same time, and Google worked with Red Hat and IBM on llm-d 1.0 so that GPU and TPU inference can share one entry point on Google Kubernetes Engine.

Meta's PyTorch engineering director Jana van Greunen said the company cannot be tied to a single hardware architecture and that TorchTPU already supports key production workloads in large language models, advertising and recommendation systems. That statement matters because Meta is one of Nvidia's largest customers, and it signals that mission-critical training and serving code is being ported rather than kept single-vendor.

Software was the last real moat protecting CUDA in accounts that had already accepted TPU hardware. Anthropic has signed for up to one million TPUs, and Meta has been reported to hold multibillion-dollar TPU rental agreements. Removing the porting tax does not make TPUs equivalent to GPUs at the silicon level: the technical report concedes that shape, sharding and operator choices tuned for GPU kernels will not reach peak performance without retuning.

For buyers, the change is about second sources. A credible TPU path means accelerator procurement no longer defaults to one vendor, and it shifts value toward the parts that both architectures need: HBM stacks, advanced packaging substrates, 800V-class power delivery, liquid-cooling manifolds and 1.6T optical links. Planners should expect ASIC volumes to rise faster than GPU volumes in 2027 by unit count while GPU revenue keeps growing.

Design teams preparing dual-path boards can source the supporting silicon from our component catalogue, including the RTL8211F gigabit PHY used on management and out-of-band networks and the INA238 power monitor for per-rail telemetry.