Google’s TPUs land inside Meta’s data centers, shaking up the AI chip landscape

Google’s in‑house AI accelerator, the Tensor Processing Unit (TPU), has quietly taken a major step into the broader market: Meta is now running AI workloads on Google TPUs inside its own data centers. According to reporting on the deal, this is the first time Google has sold TPUs as a large‑scale solution to an external hyperscaler, turning the chips into a direct alternative to Nvidia’s GPUs for training and serving big AI models.

Meta is using the seventh‑generation “Ironwood” TPUs for some of its generative AI tasks, such as large language models and image systems. Benchmarks cited in the report show an optimized setup delivering high token‑generation throughput for massive open‑weight models, underlining that TPUs can compete at the frontier of model serving. Until now, TPUs were largely confined to Google’s own services and its cloud platform.

The move comes as surging demand for AI compute makes Nvidia hardware costly and scarce, pushing large platforms to diversify their accelerator options. Just a day earlier, Amazon opened up TorchNeuron, a native PyTorch backend for its Trainium AI chips, underscoring how major cloud providers are building full hardware‑plus‑software stacks. For developers, the trend points toward a multi‑vendor future where AI workloads may be spread across GPUs, TPUs and custom ASICs, depending on cost, availability and performance needs.