Nvidia shares fall 3% on report Meta will use Google AI chips (www.cnbc.com)

The Information reports Meta is considering deploying Google’s tensor processing units (TPUs) in its data centers in 2027 and may start renting TPUs from Google Cloud as soon as next year. The news knocked Nvidia shares down ~2.5% in premarket trade while Alphabet rose similarly, reflecting investor sensitivity to any move that could shift AI hardware demand away from Nvidia’s GPUs. Google’s TPUs, first built for internal use in 2018 and since iterated into more advanced generations, are specialized ASICs optimized for dense tensor operations (matrix multiplies, convolutions) that power large-scale model training and inference. If Meta adopts TPUs it would validate Google’s custom-chip approach and intensify competition in AI infrastructure. Technically, TPUs can deliver higher throughput and energy efficiency on certain ML workloads compared with general-purpose GPUs, but moving major PyTorch-centric pipelines to TPU hardware involves software changes (XLA/JAX/TPU toolchains or interoperability layers). Short-term rental from Google Cloud would let Meta experiment without full stack migration. Strategically, the move could diversify hyperscaler procurement, pressure Nvidia on performance-per-watt and software ecosystem, and accelerate an industry trend toward heterogeneous, ASIC-accelerated AI datacenters rather than GPU-only deployments.