Analysis: Google opens its AI chips to rivals but won't say if it makes money

Google isn't selling silicon—it's executing a strategic loss-leader to build a CUDA alternative, structurally identical to AWS subsidizing EC2 in 2006-2010. The refusal to disclose TPU revenue is the tell: this is a $5-10B ecosystem subsidy, not a P&L line. Technically, Apple sourcing TPUs gives its Apple Intelligence inference stack a second compute path independent of NVIDIA, directly eroding the CUDA software moat. Anthropic's access accelerates PyTorch-on-TPU compiler maturity, fragmenting the training stack. On compliance, Google now occupies three roles simultaneously—design house, cloud provider, and chip supplier. US and EU antitrust authorities will likely open inquiries within 18 months. Advanced packaging still routes through TSMC in Taiwan, China, a single-point-of-failure Washington continues to pressure for diversification. Market dynamics: NVIDIA won't cut prices; it will deepen NIM and TensorRT-LLM exclusivity. AMD will position MI325X as the open alternative. The real wildcard is Apple—TPU access lets it decouple custom inference silicon from NVIDIA margins. 12-24 month outlook: TPU captures 15-20% of non-NVIDIA training workloads. The endgame isn't chip revenue—it's making JAX/TF the default framework, which makes TPU the default silicon.