Tpu Production Doubling

Epoch AI estimates that Broadcom shipped twice as much TPU computing power in its fiscal third quarter of 2026, May to early August, as in the quarter before: roughly 1.97 million Nvidia H100 equivalents, up from about 0.96 million. TPUs are Google's custom AI chips. Google designs the architecture, and Broadcom handles the physical design, buys the memory and works with TSMC on manufacturing. In the previous three quarters, shipments grew between 2% and 22% a quarter, according to the analysis.

The jump came from both more chips and more powerful ones. Unit volume rose about 65% to roughly 800,000, while computing power doubled, because the mix shifted toward stronger chips: the older v6e dropped out, leaving mostly TPU v7, which Epoch says delivers about 2.5 times the v6e's performance. A newer TPU v8i entered production in the quarter and made up about 11% of its computing power. The H100-equivalent unit converts each chip's peak 8-bit throughput into a count of Nvidia H100s, which Epoch says is most accurate for model training and only approximate for other uses.

These are estimates worked back from revenue, not reported shipments. Broadcom reported $16.7 billion in AI chip revenue for the quarter, and its CEO, Hock Tan, said custom accelerators made up 73% of it. Epoch's median assigns 92% of that to TPUs, then divides by an estimated price per chip. The 2026 figures include chips for Anthropic as well as Google.

Anthropic's deal is the main source of doubt. In 2025 Broadcom described it as a sale of finished racks, which a SemiAnalysis estimate values at about $25,000 per TPU, against Epoch's median chip price of about $13,000. In June 2026, Tan said, "No racks. It is all a chip business only." Epoch's central case treats the revenue as chip sales and gives rack pricing a 25% probability. If Anthropic accounted for 40% of the quarter's TPU revenue at rack prices, total shipments would be about 1.67 million H100 equivalents, 15% below the median.