Broadcom TPU compute shipments double to nearly 2 million H100-equivalents
Summary
Broadcom's shipments of Google-designed Tensor Processing Units (TPUs) roughly doubled in compute terms between its fiscal Q2 and Q3 2026, the period from May to early August. Epoch AI estimates the quarter's output at close to 2 million Nvidia H100 equivalents. That marks a break from the prior three quarters, when quarterly growth ran between 2% and 22%. Unit volumes rose by about 65% to roughly 800,000 chips. Compute grew faster than units because the mix moved toward stronger parts.
TPU v7, which delivers about 2.5× the performance of the v6e, now accounts for the vast majority of shipments. The v6e had led TPU compute through fiscal Q4 2025. TPU v8i, rated at about 2.2× the v7, entered production during the quarter and made up about 11% of TPU compute. The estimate covers chips bound for both Google and Anthropic. Google sets the architecture. Broadcom handles physical design, memory procurement and fabrication work with TSMC.
Epoch derives its volumes from Broadcom's reported results. AI semiconductor revenue was $16.7 billion in the quarter, and CEO Hock Tan put the custom accelerator share at 73%, about $12.2 billion. Epoch's median places 92% of that as TPUs, or about $11.2 billion. By projecting Google's spending trend forward to roughly $6.0–6.8 billion, Epoch attributes around 40% of TPU revenue to Anthropic, within a 30–50% range. Broadcom has said it will deliver 1 GW of TPU compute to Anthropic in 2026.
The main uncertainty is whether Anthropic revenue was booked as finished racks rather than chips. SemiAnalysis valued 400,000 Ironwood TPUs for Anthropic at about $10 billion in racks, or about $25,000 per TPU. Epoch's median v7 chip price is ~$13,000. Epoch assigns a 25% probability to rack booking, which trims its median from 2.05 million to 1.97 million H100e. In a full rack case, Anthropic would have received about 180,000 v7s rather than 340,000, and total shipments would fall to about 1.67 million H100e. MediaTek is expected to ramp TPU production from late 2026 into 2027.
Why it matters
A step change of this size in custom accelerator output shows that non-Nvidia silicon is reaching data centers in volume. Its customer base also now extends beyond Google to Anthropic. The shift toward v7 and v8i parts concentrates more compute per chip, which raises power and cooling demands at each deployment site. Broadcom's commitment of 1 GW of TPU compute to Anthropic in 2026 also turns a chip contract into a power-delivery question. The unresolved split between rack and chip revenue determines how many units are actually landing. That count is what capacity planners, colocation providers and investors need to size downstream demand for power, networking and integration.
Expert Lens
A 1 GW delivery commitment inside a single calendar year likely hinges less on fabrication than on whether host sites have energized capacity and liquid-cooled halls ready to accept dense v7 and v8i racks. If Broadcom is shipping chips while a third party handles rack integration, that assembly step may become a hidden schedule risk between wafer output and ready-for-service dates. The source also suggests deliveries to Anthropic could be weighted toward later quarters. If so, demand for powered shell and interconnection capacity may bunch up near year-end rather than spread evenly. The next signal to watch is whether Broadcom's coming disclosures clarify rack versus chip booking for Anthropic. That would show how much of this ramp is already installed and how much is still waiting for power.
Sources
This brief was drafted with AI assistance from public sources; every number is attributed to the sources above. See the AI-assisted Disclosure and Methodology.