Broadcom's TPU Output Nearly Doubles as Google Chip Mix Gets Beefier

Broadcom's custom silicon machine is running hot. According to an Epoch AI estimate published Nov. 4, 2026, the company's TPU compute shipments roughly doubled from fiscal Q2 to fiscal Q3 2026, landing at nearly 2 million Nvidia H100 equivalents in a single quarter. That's a hefty slice of global AI compute, and it's coming from a chip designer rather than the GPU king.

Here's the twist: the number of actual chips didn't double. Unit shipments climbed about 65% quarter over quarter to roughly 800,000 chips, per the estimate. The gap between 65% more chips and about 100% more compute comes down to what's inside them — Broadcom pushed production toward higher-performance parts, so each unit carries more muscle.

The headliner was TPU v7, which now accounts for the majority of shipments and packs around 2.5 times the performance of its predecessor, TPU v6e. Early production has already started on TPU v8i, which delivers roughly 2.2 times the performance of v7. Chain those multipliers together and a v8i chip lands near five times a v6e.

The wider numbers tell the same story. Broadcom's fiscal Q3 2026 XPU shipments — its catch-all term for custom accelerators built for specific customers — rose more than 3.5 times year over year, and XPUs made up 73% of the company's AI revenue. AI semiconductor revenue, drawn from six XPU customers, surpassed $16.7 billion.

Google sits at the center of it all. Broadcom has a long-term agreement with the search giant covering future TPUs and AI networking, a deal projected to be worth multi-tens of billions of dollars annually over several years. That relationship is doing a lot of heavy lifting for Broadcom's AI story.

The risk is concentration. With XPUs at 73% of AI revenue and that revenue tied to just six customers, one client changing course would sting. Next up: Broadcom's quarterly results, how fast TPU v8i production ramps, whether that customer base widens, and whether Epoch AI's next estimate shows another doubling or a one-time mix shift.