Google TPU expansion to Oracle and OpenAI

Oracle and Google have agreed to deploy Google TPU capacity on Oracle Cloud Infrastructure starting in 2027, the first time Google's TPUs will operate on a rival hyperscale cloud. At its October 2026 Investor Day, Oracle disclosed the financial structure of its large-scale AI infrastructure contracts: a 1GW deployment generates approximately $80 billion in contract value, requires roughly $37 billion in capital expenditure, and yields about $16 billion in annual revenue. Contracts average five years, carry approximately 35% gross margin and 34% operating margin, and imply roughly 28% ROI with a cash payback period of about 3.6 years. The contracts are structured on a cost-plus basis. On the technical side, Oracle's Acceleron network ties TPU clusters directly to Oracle databases and storage, so enterprises can train and serve models without moving data across cloud boundaries. Google claims its 8th-generation TPUs deliver up to 12.6 PFLOPS of FP4 compute per chip, sustain over 97% utilization on large training runs, and provide up to 80% better performance per dollar for low-latency serving. The open-source TorchTPU stack supports existing PyTorch code and frameworks such as vLLM and SGLang, allowing workloads to move from NVIDIA GPUs without model rewrites.