Google TPUs Head to Oracle Cloud in 2027, TorchTPU Open

Google's Tensor Processing Units will be available through Oracle Cloud Infrastructure starting in 2027, the companies said at Oracle AI World, marking the first time Google's custom AI silicon is hosted by a hyperscaler outside Google. Oracle will supply the power, cooling and datacenter networking needed to deploy TPUs at scale, turning Google's in-house accelerator into a sold product line.

The move follows Google's open-source release of TorchTPU, which adds native PyTorch support to TPUs and lowers the barrier for teams considering a migration away from NVIDIA GPUs. Together the two announcements point to a faster diversification of AI compute, and with it a broader base of demand for the packaging, memory and power components behind custom ASIC accelerators.

For procurement teams, the significance is the supply chain behind the TPU. The accelerator family is designed with Broadcom, so expanded cloud availability implies higher-volume orders for advanced substrate, HBM and networking silicon across the Broadcom ecosystem. Oracle's investment in TPU-ready power and cooling also creates design-in opportunities for liquid-cooling and power-distribution vendors.

Buyers weighing GPU versus ASIC strategies now have a cleaner benchmark: AWS offers Trainium, Microsoft runs Maia, and OCI will market TPUs openly. The shift from proprietary captive accelerators to merchant availability widens sourcing options for AI server components and gives second-source strategies more credibility in 2027 planning cycles.

Expect TPU-driven demand to show up first in lead times for high-layer-count substrates and HBM modules. Teams that hedged purely on GPU availability last cycle should revisit allocation assumptions for the coming TPU ramp.