Google Cloud Backs SGLang to Kill the GPU-to-TPU Migration Tax
Google and RadixArk bring SGLang's full inference stack to TPUs, letting teams swap GPU clusters for cheaper hardware without rewriting a line of serving code
- Google + RadixArk partner to bring SGLang to Google Cloud TPUs, eliminating the framework migration tax for inference teams.
- SGL-JAX is live today supporting Gemma, Qwen, DeepSeek, Grok, Flux, Wan, and more on the latest TPU generations.
- SGL-torchtpu coming later this year — a PyTorch-native TPU backend with MPMD support, so teams need zero JAX knowledge.
- Full SGLang feature parity on TPUs: Radix Cache, HiCache, speculative decoding, quantization, and all parallelism modes included.
- Day 0 TPU support promised for every new open model and every new TPU generation going forward.
- RadixArk raised $100M seed at $400M valuation from Accel, Spark Capital, NVentures, and AMD; SGLang stays Apache 2.0 and free.
Google Cloud and RadixArk have announced a formal partnership to bring SGLang , the open-source inference framework already running on hundreds of thousands of GPUs worldwide , to Google's TPU hardware. The first deliverable, SGL-JAX, is available today. A second backend, SGL-torchtpu, is coming later this year and will let teams run the exact same SGLang serving stack on TPUs using standard PyTorch tooling, no JAX knowledge required.
The framework behind a trillion tokens a day
If you haven't heard of SGLang, you've almost certainly used something it powered. SGLang runs inference for Google, Microsoft, NVIDIA, Oracle, AMD, LinkedIn, and xAI , generating trillions of tokens per day across more than 400,000 GPUs as of 2026. The framework was initially released in 2024 and developed by Ying Sheng and Banghua Zhu out of LMSYS, the Berkeley research collective behind Chatbot Arena. It has since grown to over 30,000 GitHub stars and 1,700 contributors.
SGLang's commercial backer, RadixArk, launched with a $100 million seed round at a $400 million valuation, but the framework itself remains Apache 2.0 and free for self-hosting. The round was led by Accel and co-led by Spark Capital, with participation from NVentures (NVIDIA's venture capital arm), AMD, and others. The investor list reads like a who's-who of the inference world: John Schulman (OpenAI co-founder), Soumith Chintala (PyTorch creator), and Thomas Wolf (Hugging Face co-founder) also joined the round.
What's actually shipping
The partnership delivers two distinct TPU backends, each targeting a different developer profile:
- SGL-JAX (available now): It leverages SGLang's high-performance server architecture and uses JAX to compile the model's forward pass, delivering fast, native TPU inference while maintaining support for continuous batching, prefix caching, tensor and expert parallelism, speculative decoding, and highly optimized TPU kernels. Model support already covers Gemma, Qwen, DeepSeek, GLM, Kimi, Grok, and diffusion models like Wan and Flux.