Google
2026-04-29
Google Opens TPU Hardware to On-Prem, 8th-Gen Chips Target Nvidia
Summary
Key Takeaways
In Q1 2026 earnings, Google detailed 8th-gen TPUs: TPU 8t for training (3x processing power over Ironwood), TPU 8i for inference (80% better perf/dollar). It will also offer NVIDIA Vera Rubin NVL72. Critically, Google will deliver TPU hardware directly to customer data centers for the first time, targeting AI labs and HPC firms. The Wiz acquisition closed, enhancing AI security. Gemini Enterprise Agent Platform launched; BigQuery Gemini workflows grew 30x YoY. Enterprise AI became Cloud's primary growth driver, with gen AI revenue up ~800% YoY.
Why It Matters
This move is a strategic encirclement of Nvidia. By offering on-prem TPUs, Google locks customers into its JAX/TensorFlow stack, creating hardware-level dependency that outlasts cloud contracts. Hidden limitations: TPU's efficiency on sparse models (MoE) and general workloads lags behind GPUs. The 3x performance claim likely applies only to dense matrix ops, bottlenecked by HBM bandwidth and interconnect topology. Customers face hidden TCO from liquid cooling, RoCEv2 networking, and power infrastructure—potentially exceeding cloud costs. Once Google shifts architecture (e.g., chiplet), deployed hardware depreciates rapidly, locking buyers into a single-vendor roadmap.
PRO Decision
[Vendors]: Nvidia must reinforce CUDA irreplaceability, launch inference-optimized GPUs (e.g., B200) with lower TCO, and partner with Arista/Dell for open AI clusters to attack TPU's software lock-in. AMD should accelerate ROCm framework support and ally with HPE for open hardware. [Enterprises]: CIOs must demand independent MLPerf benchmarks for TPU 8t/8i, assess hybrid deployment with existing GPUs, and require clear EOL policies. Prioritize hardware supporting both CUDA and ROCm to preserve flexibility. [Investors]: TPU hardware sales may cannibalize Google Cloud revenue short-term and trigger price wars with Nvidia. Monitor capex for TPU manufacturing and gross margin trends. The inference market opportunity is real but unproven; current evidence favors Nvidia's ecosystem maturity.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)