Rohan Paul (@rohanpaul_ai)T2·85 pts
- Google TPUs exit their native cloud ecosystem for the first time, supported by Oracle infrastructure with capacity ramping up from 2027
- 8th-gen TPUs deliver 12.6 PFLOPS FP4 compute per chip, offering up to 80% better performance-per-dollar for low-latency serving
- The open-source TorchTPU stack supports PyTorch, vLLM, and SGLang, enabling migration from NVIDIA GPUs without model rewrites