Broadcom’s TPU compute shipments double to nearly 2 million H100 equivalents

Broadcom’s custom AI chip business just had the kind of quarter that makes spreadsheet columns run out of room.

According to an Epoch AI estimate published Nov. 4, 2026, Broadcom’s TPU compute shipments roughly doubled from fiscal Q2 to fiscal Q3 2026, reaching nearly 2 million Nvidia H100 equivalents.

What actually doubled

First, a quick translation. A TPU, or tensor processing unit, is a chip built specifically for the math behind machine learning.

Comparing different chips is tricky, so Epoch AI converts their output into “H100 equivalents.” Think of it like expressing a car’s power in horsepower. It doesn’t matter what’s under the hood, you get a common unit.

Here’s where it gets interesting. The number of physical chips did not double.

Unit shipments rose about 65% quarter over quarter, landing at roughly 800,000 chips, per Epoch AI’s estimate. The gap between 65% more chips and roughly 100% more compute comes from what those chips are.

Broadcom shifted its production toward higher-performance variants. Each chip now carries more muscle, so the same factory output delivers far more computing capacity.

The generational upgrade doing the heavy lifting

The star of the quarter was TPU v7. According to the research, it now makes up the majority of shipments.

TPU v7 offers around 2.5 times the performance of its predecessor, the TPU v6e. Swap enough old chips for new ones and your compute totals jump, even if your unit count grows more modestly.

And the next step is already rolling. Early production has begun on TPU v8i, which delivers roughly 2.2 times the performance of v7.

Stack those multipliers and the trajectory gets steep. A v8i chip would offer something in the neighborhood of five times a v6e, if you chain the two estimates together.

The broader numbers back up the momentum. Broadcom’s fiscal Q3 2026 XPU shipments were up more than 3.5 times year over year.

“XPU” is Broadcom’s catch-all term for custom accelerators it builds for specific customers. Those XPUs accounted for 73% of the company’s AI revenue in the quarter.

Broadcom’s AI semiconductor revenue, drawn from six XPU customers, surpassed $16.7 billion, according to the research findings.

Background: the Google connection

The biggest name in Broadcom’s TPU story is Google. Broadcom has a long-term agreement with the search giant covering future TPUs and AI networking.

That deal is projected to be worth multi-tens of billions of dollars annually over several years.

What this means

For investors watching the AI hardware trade, the takeaway is that custom silicon is no longer a side project. Nearly 2 million H100 equivalents in a single quarter is a meaningful slice of global AI compute, and it is flowing through a chip designer rather than the GPU leader.

Because compute doubled while units rose about 65%, a large share of the gain came from better chips, not just more of them.

There are clear risks to weigh, though. Concentration is the obvious one. With XPUs making up 73% of AI revenue and that revenue tied to just six customers, any single client changing course would be felt.

What to watch next: Broadcom’s upcoming quarterly results, the pace of TPU v8i production, and whether the six-customer base widens. Epoch AI’s next estimate will also show whether the generational upgrade keeps doubling compute, or whether this quarter’s leap was a one-time mix shift.