Meta runs Google AI chips in its own data centers

Big news landed in the AI world this week. Google’s custom chips are now running live workloads for Meta. This is a significant move for the tech giants.

The change comes through a new software update. Google announced a PyTorch native version of its SGLang inference engine. This engine now works directly with Google’s Tensor Processing Units. A similar update for the vLLM engine was also released as open source. Previously, these tools required Google’s JAX framework to function on TPUs. Now, they work natively with PyTorch. This means developers do not need to rewrite their code. The hardware is now a direct option for many AI teams.

Meta confirmed that the new setup is already in use. Zongwei Zhou, who leads Meta’s PyTorch TPU team, said the system is running in production today. He noted a close collaboration between Meta, Google, and the SGLang team. His team works at RadixArk, a startup founded by the creators of SGLang. Zhou described the new engine as a remarkably mature serving stack. Jana van Greunen, a director at Meta, added that the company cannot be limited to one hardware type. She said the new tool allows Meta to deploy its workloads on Google’s chips. Google confirmed it built this version with Meta’s PyTorch team.

This is not the first time these companies have worked together. Google calls the older PyTorch XLA bridge a predecessor to the new system. Reports from last year indicated that Meta research teams used TPUs through this older bridge from 2020 to 2023. Meta eventually canceled those contracts. The current push by Google was aimed specifically at Meta. The goal was to provide a path for Meta to use TPUs without switching to JAX.

The location of these chips remains a topic of discussion. It is not entirely clear where Meta’s production TPUs are housed. Talks last year included plans for Google Cloud TPU rentals in 2026. There were also mentions of purchases in 2027. Meta reportedly signed a multi year rental agreement. This suggests a long term commitment to the hardware.

The timing of this announcement is notable. It came one day after Amazon made a similar move. Amazon open sourced its own PyTorch backend for its Trainium chips. Both actions appear designed to reduce reliance on Nvidia’s software. Nvidia has long held a dominant position in the AI market. These new tools offer competitors more flexibility.

Performance numbers are impressive. A tuned vLLM setup served a large open weight model called Kimi K3. The model has 2.8 trillion parameters. On the seventh generation Ironwood TPU, the speed topped 700 tokens per second for a single user. This data comes from Inferact, a startup that helped build the new software. The test used 16 Ironwood chips with a method called speculative decoding. This technique uses a smaller model to speed up the main one. The same test on 16 Nvidia GB200 chips reached 452 tokens per second. Without that extra technique, the new kernels delivered 1.4 to 2 times the baseline throughput at small batch sizes.

The development team has more plans. SGLang TorchTPU will add more models. DeepSeek V4.1 Flash and Reflection Beam are next on the list. The team also aims to support new open models on their release day. This goal is set for the first quarter of 2027. More details will be shared at the SGLang Summit 2026. That event is scheduled for November 12 and 13 in San Francisco.

This shift signals a changing landscape. For years, Nvidia’s chips and software were the default choice. Now, other options are maturing quickly. Meta’s move shows that major tech companies are willing to diversify their hardware. This could pressure Nvidia to adjust its strategies. The race to build faster and more efficient AI infrastructure is heating up.

Readers in Texas and the South Central states should note the broader impact. Data centers are expanding in the region. More hardware options mean more competition. This can lead to better services and lower costs for businesses. The tech world is moving fast. Keeping up with these changes helps local businesses make informed decisions. The next few months will reveal if this trend continues to grow.