Meta and Google Discuss Deploying TPUs in Meta Datacenters Starting 2027 (finance.yahoo.com)

Google is shifting strategy by pitching its custom tensor processing units (TPUs) for installation inside customer data centers, and Meta is reportedly in talks to integrate TPUs into its facilities starting in 2027 while renting TPU capacity from Google Cloud as soon as next year. Historically confined to Google’s own cloud, TPUs are specialized accelerators for tensor-heavy AI workloads; offering them on-premises would be a big vote of confidence if Meta — today primarily running on Nvidia GPUs — commits “billions” to the rollout. Alphabet’s stock ticked up after the report, while Nvidia dipped, underscoring the market significance. Technically and strategically this matters because on-prem TPUs give organizations data locality, lower-latency inference, and easier compliance for sensitive workloads that some customers won’t place in external clouds. Google’s pitch could blunt Nvidia’s dominance by targeting enterprise customers with a hybrid-cloud + on-prem TPU model; Google execs estimate TPU expansion could capture a meaningful slice (up to ~10%) of Nvidia’s annual revenue. Practical implications include model-porting and software-stack changes (XLA/TPU runtimes vs. GPU CUDA ecosystems), procurement and integration timelines (multi-year deployments), and increased competition in AI compute supply chains as demand for large-scale accelerators surges.