The Anthropic deal covers 5 gigawatts of TPU capacity and is worth up to $40 billion, marking Google's largest commercial commitment of its custom silicon to a single external customer. The expansion will provide Anthropic with multiple gigawatts of TPU capacity, expected to come online starting in 2027, via Google Cloud services as well as access to Google-built TPUs supplied through Broadcom.
Each Ironwood chip delivers 4,614 TFLOPs of FP8 compute with 192 GB of HBM3e memory and 7.4 TB/s of memory bandwidth, with Google claiming a 10x peak performance improvement over TPU v5p. The volume signals Google's conviction that custom silicon, not just rented GPU clusters, will define the economics of AI inference at scale. Semianalysis argues Google is the newest and most threatening merchant silicon challenger to Nvidia, noting that Anthropic will deploy TPUs in its own facilities, positioning Google to compete directly with Nvidia.
For infrastructure architects evaluating AI accelerator procurement, Google's TPU capacity deals represent a viable cost/performance alternative to NVIDIA GPUs for inference-heavy workloads. Training large language models on TPUs can be 40–60% cheaper than equivalent GPU workloads, according to Google Cloud's estimates, though supply is now committed through major partners through 2027–2028.