Google's TPU program represents the most mature and vertically integrated custom silicon threat to NVIDIA's data center GPU dominance. The 7th-generation TPU v7 (Ironwood), announced April 2025 with limited availability by late 2025, delivers 4,614 FP8 TFLOPS per chip -- slightly exceeding NVIDIA B200's 4,500 TFLOPS -- with 192 GB HBM3E and 7.4 TB/s bandwidth. Ironwood's defining advantage is pod-scale: 9,216 chips interconnected via ICI in a single superpod delivering 42.5 ExaFLOPS, compared to NVLink's 72-GPU ceiling at 0.36 ExaFLOPS.
The Anthropic deal (Oct 2025) -- up to 1M TPUs, $10B in Broadcom-manufactured Ironwood racks plus an $11B follow-on, with remaining capacity rented via GCP totaling ~$52B -- is the largest cloud compute deal ever. In January 2026, Google confirmed TPUs outshipped GPUs by volume for the first time. Google and Meta's TorchTPU collaboration (announced Dec 2025) directly targets CUDA switching costs by enabling native PyTorch execution on TPUs, though production readiness is 12-18 months away. Key limitations: Ironwood is cloud-only (cannot be purchased), ICI bandwidth per chip (1.2 TB/s) trails NVLink (1.8 TB/s), no FP4 support vs Blackwell's FP4 advantage, and the software stack -- historically JAX-only with limited external tooling -- remains inferior to CUDA's two-decade ecosystem. For NVIDIA investors, the TPU threat is most acute in inference (where Google claims 4.7x price-performance vs H100) and in capturing frontier lab spend (Anthropic, potentially Meta), but less threatening for training where NVLink's low-latency interconnect and CUDA's flexibility remain advantages..
Competitive pressure is real but bounded
Custom ASICs and AMD offer cheaper alternatives for specific workloads, but only a handful of companies can afford multi-billion-dollar chip programs. The competitive threat is structural but limited in scope.
What is the actual MLPerf benchmark performance of Ironwood vs Blackwell for common LLM inference workloads (Llama 3, Gemini, Claude)? Google has not submitted Ironwood MLPerf results, preventing apples-to-apples comparison.