Broadcom shipped TPU computing power equal to about 1.97 million Nvidia H100 GPUs in its fiscal third quarter of 2026 (May to early August), roughly double the 0.96 million of the quarter before, according to estimates published by Epoch AI. These are Epoch’s modelled estimates, built from Broadcom’s reported AI revenue, not unit figures Broadcom discloses. Epoch’s 90% confidence interval for the quarter runs from 1.57 million to 2.35 million H100-equivalents.
By our arithmetic from Epoch’s median estimates, the quarter’s growth was 104%, after 22%, 2% and 10% in the three quarters before it. Over six quarters, from 0.33 million in fiscal Q1 2025, the quarterly figure is about six times larger, also by our arithmetic.
Business Pill · Vertical Integration
An evergreen Business Engineer explainer on vertical integration: owning more of the stack you depend on. It teaches the general idea only and does not discuss Broadcom, Google or Anthropic.
The key insight: Epoch AI estimates Broadcom shipped TPU compute equal to about 1.97 million H100s in fiscal Q3 2026. By our arithmetic from its medians, that is 104% more than the quarter before, after growth of 22%, 2% and 10% in the three prior quarters.
Epoch’s median estimates of TPU compute shipped per Broadcom fiscal quarter, in H100-equivalents, with quarter-on-quarter change by our arithmetic:
Q1 FY25: 326,753
Q2 FY25: 514,633 (+57%)
Q3 FY25: 696,508 (+35%)
Q4 FY25: 851,340 (+22%)
Q1 FY26: 872,392 (+2%)
Q2 FY26: 961,591 (+10%)
Q3 FY26: 1,965,979 (+104%), of which TPU v7 1,752,773 and TPU v8i 213,870
Epoch says unit volumes rose “by about 65% quarter over quarter to roughly 800,000 chips, while compute output doubled,” which it attributes to a shift toward more powerful chips. It says the TPU v7 delivers about 2.5 times the performance of the v6e, and that the TPU v8i, which entered production in the quarter, delivers about 2.2 times the v7. Epoch estimates the v8i made up about 11% of TPU compute in the quarter.
What Was Posted on X
Broadcom TPU compute shipments doubled from fiscal Q2 to Q3 2026 (May to early August), reaching the equivalent of nearly 2 million Nvidia H100 GPUs, according to our estimates.
This is a sharp increase from the previous three quarters, when shipments grew 2โ22% per quarter.
Epoch AI posted its estimate on X on 5 November: Broadcom TPU compute shipments doubled from fiscal Q2 to Q3 2026, reaching the equivalent of nearly 2 million Nvidia H100 GPUs, after growth of 2-22% per quarter in the previous three quarters. This piece is built on the data insight the post links to.
Epoch AI’s median estimates of TPU compute shipped per Broadcom fiscal quarter, in millions of H100-equivalents. Quarter-on-quarter growth under each bar is our arithmetic from Epoch’s medians. Estimates, not company-reported unit figures.Animated: Epoch AI’s median estimates of TPU compute shipped per Broadcom fiscal quarter, in millions of H100-equivalents, Q1 FY25 to Q3 FY26. Estimates, not company-reported figures; quarter-on-quarter growth is our arithmetic.
Epoch defines an H100-equivalent by dividing a chip’s peak dense 8-bit operations by the H100’s, then multiplying by the number of chips; on its AI chip sales hub it gives the example that a TPUv7 performs about 2.3 times as many operations as an H100. Epoch says the conversions are approximate and rest on peak 8-bit specifications. For Google TPUs, its hub says the estimates “are based on Broadcom revenue and thus likely reflect chips produced and delivered to Google.”
The input is Broadcom’s own number. In its third-quarter release, for the quarter ended 2 August 2026, Broadcom reported AI semiconductor revenue of $16.7 billion, up 221% year on year and 54% on the quarter, and guided to $21.7 billion for the fourth quarter. Epoch cites CEO Hock Tan saying on the earnings call that custom accelerators made up 73% of it, about $12.2 billion. Its median estimate places 92% of that XPU revenue as TPUs, about $11.2 billion, and converts revenue into chips using a per-chip price from a bill-of-materials costmodel; its median TPU v7 price in the quarter is about $13,000.
The Anthropic Question: Chips or Racks
Epoch says 2026 estimates cover TPUs shipped to Anthropic as well as Google, and that the size of the quarter depends on how Broadcom booked the Anthropic revenue. Its central case treats it as chip sales. Extending Google’s recent spending trend, Epoch estimates around 40% of Q3 TPU revenue was driven by Anthropic, and uses a range of 30% to 50%.
If that revenue was booked as finished racks at about $25,000 per TPU, which Epoch takes from a SemiAnalysis estimate, Epoch says total Q3 shipments would be about 1.67 million H100-equivalents, 15% below its median.
Epoch assigns a 25% probability to rack booking; it says that lowers its median from 2.05 million to 1.97 million.
Epoch quotes Hock Tan on the fiscal Q2 2026 call in June: “No racks. It is all a chip business only. We only [do] chips. Only chips.”
Epoch reads Broadcom’s 10-Q as showing that Apollo took on agreements to purchase AI racks based on Broadcom-designed accelerators, with Broadcom backstopping up to about $29 billion of the related lease obligations.
Epoch notes that $19.9 billion of $20.8 billion of Q3 hardware revenue, after subtracting VMware license revenue, shipped to Asia Pacific, where TPU racks are assembled.
Two Ways to Count a Quarter
Our 30 October piece used Epoch’s calendar-quarter data, which put Google’s TPU additions at 1.61 million H100-equivalents in the second calendar quarter of 2026. This new insight is cut by Broadcom fiscal quarter, and its third quarter runs from May to early August, so the two series do not line up quarter for quarter.
As we read Epoch’s numbers, chips rose about 65% while compute doubled: much of the jump comes from the mix moving to TPU v7 and the new v8i, not from unit volume alone.
As we read the method, the series is Broadcom’s reported AI revenue divided by a modelled chip price, which is why Epoch spends most of its note on whether Anthropic’s TPUs were booked as chips or as racks.
Epoch AI, 4 November 2026
“We estimate Broadcom TPU compute shipments doubled from fiscal Q2 to Q3 2026 (May to early August), reaching the equivalent of nearly 2 million Nvidia H100 GPUs.”
Three Implications
FOR BROADCOM Broadcom guided AI semiconductor revenue to $21.7 billion for its fourth quarter, up from $16.7 billion in the third.
FOR ANTHROPIC Epoch estimates around 40% of Q3 TPU revenue was driven by Anthropic, in the quarter Broadcom started shipping TPUs for it.
FOR CHIP COMPARISONS Epoch’s H100-equivalents compare peak 8-bit specifications; it says the conversions are approximate.
The framework says: “Memory is becoming the new compute. HBM growth is outpacing accelerator demand, making memory the rate-limiting factor in the AI scaling equation.”
As we read it, Epoch’s note points the same way: it reports that in September Broadcom’s management attributed gross margin pressure to the growing share of XPUs and their “increasing memory content”, and that Broadcom procures the memory for TPUs. The framework does not discuss these estimates.
What Is Not Established
Broadcom does not disclose TPU unit volumes or revenue from Anthropic; every unit and H100-equivalent figure here is Epoch’s estimate. Epoch says its Q3 volumes would be revised downward if later disclosures show the Anthropic revenue was booked as racks. H100-equivalents compare paper specifications, not delivered performance, and delivered chips may take time to be installed.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Epoch AI estimates Broadcom shipped TPU compute equal to about 1.97 million H100s in fiscal Q3 2026, double the quarter before, after 2% to 22% growth in each of the three prior quarters. The figure is a model built on Broadcom’s $16.7 billion of AI revenue, and how Anthropic’s chips were booked could move it.
A note on sourcing. Unit and H100-equivalent figures are Epoch AI’s estimates (CC-BY), from its data insight of 4 November 2026 and its AI chip sales hub; Broadcom does not report them. AI revenue figures are from Broadcom’s third-quarter release of 2 September 2026. Quarter-on-quarter growth and the six-quarter multiple are our arithmetic from Epoch’s medians.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
