What happened
Epoch AI’s Nov. 4, 2026 Data Insight estimates that Broadcom’s TPU compute shipments doubled between its fiscal second and third quarters of 2026. The Q3 period runs from May to early August. Its central estimate is 1,965,979 H100-equivalents, with a 90% confidence interval from 1,567,748 to 2,350,392. Epoch AI says this was a sharp departure from the three prior quarters, when estimated growth ranged from 2% to 22% per quarter.
The estimate distinguishes chip count from aggregate compute. Epoch AI estimates that TPU unit shipments rose about 65% quarter over quarter to roughly 800,000 chips, while compute output doubled. It attributes that difference to both added volume and a new production mix. TPU v7 supplied the vast majority of estimated Q3 compute. Epoch AI says TPU v7 offers about 2.5 times the performance of TPU v6e, which had accounted for most TPU compute through fiscal Q4 2025. TPU v8i also entered production in fiscal Q3 2026; the analysis estimates it accounted for about 11% of quarterly TPU compute and delivers roughly 2.2 times the performance of TPU v7.
Epoch AI derives the figures from Broadcom’s revenue disclosures rather than a public unit-by-unit shipment report. Broadcom reported $16.7 billion in AI semiconductor revenue in fiscal Q3 2026. According to Epoch AI, Broadcom CEO Hock Tan said custom accelerators represented 73% of that total, or about $12.2 billion. The analysis places a median 92% of XPU revenue in TPUs, assumes a mix dominated by TPU v7, and converts revenue to chip volumes using estimated per-chip prices based on a bill-of-materials cost model. The calculated volumes include shipments to Google and Anthropic.
The report also explains why that conversion is uncertain. Its central case assumes Broadcom booked revenue for TPU chips rather than complete racks. Yet Broadcom’s earlier statements characterized the Anthropic arrangement as a system sale and referred to AI rack orders. On the fiscal Q2 2026 call, Hock Tan answered a question about rack production: “No racks. It is all a chip business only. We only [do] chips. Only chips.” Epoch AI also points to Broadcom filings describing Apollo taking on certain agreements to purchase AI racks and related leases. This leaves an unresolved question over the commercial path for the hardware.
Epoch AI assigns a 25% probability that Anthropic-related revenue was booked at rack value. It estimates Anthropic represented around 30% to 50% of Q3 TPU revenue, with a central reference point of about 40%. In the report’s illustrative case where 40% was booked at rack prices, estimated Q3 shipments would be around 1.67 million H100-equivalents, or 15% below the current median. The report’s central estimate after its uncertainty adjustment is 1.97 million H100-equivalents.
What the archive adds
New Runtime’s earlier coverage has followed the expanding TPU supply and availability picture, including Oracle’s accelerator expansion on Oct. 31, 2026, and Google’s TPU 8tx and TPU 8i Max announcement on Nov. 9, 2026. This is a new quantified estimate of Broadcom’s shipment scale, rather than another TPU product or cloud-availability event.
Trend signal
This supports the broader infrastructure shift visible in New Runtime’s earlier TPU coverage, but it does not yet fit a specific pattern New Runtime tracks. The estimate is about manufacturing scale, product mix, and revenue accounting; it does not establish a change in agent economics, interoperability, or the cost of verified completed work.
Why now
The apparent jump coincides with a transition in the shipment mix toward TPU v7 and the first estimated TPU v8i production. Epoch AI also includes early Anthropic deliveries in its 2026 estimates. Those factors make the quarter important, but they also make it a poor basis for assuming a stable long-term quarterly growth rate.
Connections to other events
The result meets two recent New Runtime threads: new Google TPU systems and prospective TPU availability through Oracle Cloud Infrastructure. Those events describe expanding options for obtaining or using Google-designed accelerators; Epoch AI’s analysis addresses the upstream volume that Broadcom may have delivered. The two views are complementary, but neither proves how much deployable capacity reaches external cloud customers.
Why it matters
For teams building and budgeting AI systems, the practical implication is not that an H100 can be replaced one-for-one by a TPU. It is that the supply of non-Nvidia training and inference capacity may be expanding quickly, and that hardware mix can increase delivered compute faster than chip count alone suggests. That can influence where large model providers place workloads, which cloud capacity appears, and how concentrated accelerator supply remains.
Builders should separate those strategic signals from near-term procurement claims. The analysis does not announce a new TPU service, a price reduction, or immediate capacity for ordinary teams. It does show why infrastructure planning increasingly needs to account for custom accelerators, their software environments, and the commercial arrangements that determine who can actually use them.
Limits and uncertainty
These are Epoch AI estimates, not independently confirmed shipment counts from Broadcom. The report converts chip capabilities into H100-equivalents using relative peak 8-bit FLOP/s specifications. Epoch AI cautions that memory capacity, memory bandwidth, software stability, and workload type can make real-world usefulness higher or lower than that conversion implies. It says the comparison is most applicable to compute-bound model training.
The largest uncertainty is whether Broadcom’s Anthropic-related revenue represents chip-only sales or rack-level sales that include other components. Rack-value accounting could lower inferred unit volume materially. Broadcom does not separately disclose TPU revenue from Anthropic, and the timing of Anthropic deliveries inside the quarter is also unresolved. The reported confidence interval captures model uncertainty, not an audited physical inventory.
Skill to build now
Build an accelerator-portability check into infrastructure planning this week. List the training and inference workloads that genuinely require a specific CUDA dependency, then identify what would have to change to run each one on a TPU-oriented stack: framework support, compilation path, serving runtime, observability, and performance tests. Keep a workload-level benchmark plan rather than treating H100-equivalent figures as deployment proof.
This exercise is useful even if a team never buys TPU capacity. It exposes where hardware choice is constrained by software assumptions and helps distinguish a future supply option from an immediately usable platform.
Reader takeaway
Epoch AI estimates that Broadcom’s TPU compute shipments nearly reached 2 million H100-equivalents in fiscal Q3 2026, doubling quarter over quarter as both volume and chip capability rose. It is a meaningful signal of custom-accelerator scale, but not a disclosed physical count or a direct measure of workload performance. Treat the headline as an informed estimate with important accounting and comparability limits.
Open question
Will subsequent Broadcom disclosures clarify whether Anthropic-related revenue was recognized for chips or complete racks, and how much of the estimated Q3 increase was delivered to Anthropic rather than Google?