The AI Chip Sales hub estimates TPU volumes from Broadcom’s reported AI semiconductor revenue. In fiscal Q3 2026, Broadcom reported $16.7 billion in AI semiconductor revenue, and CEO Hock Tan said on the earnings call that custom accelerators made up 73% of it, about $12.2 billion. Accounting for uncertainty, our median estimate places 92% of that XPU revenue as TPUs. We model the TPU mix as being mostly TPU v7. Broadcom also began shipping TPU v8i in the quarter, and we estimate it made up about 11% of TPU compute for the quarter. We convert TPU revenue into chip volumes using an estimated per-chip price based on a bill-of-materials cost model. These derived volumes include TPUs shipped to Anthropic as well as to Google. For more details, see our AI Chip Sales documentation.

Racks vs Chips

Our analysis assumes that Broadcom’s TPU revenue comes from selling TPU packages rather than complete racks. This has been the case for Broadcom’s TPU shipments to Google. However, in Q3 2026, Broadcom also started shipping TPUs for Anthropic. Initial analysis of the deal in late 2025 suggested Broadcom deliverables would be in the form of racks.

If Broadcom booked revenue on finished racks – which would also include CPUs, system memory, networking, power and cooling, and more – its revenue per TPU would be about twice our chip price. SemiAnalysis estimated that 400,000 Ironwood TPUs for Anthropic were worth about $10 billion in finished racks, or about $25,000 per TPU, compared with our median TPU v7 price in Q3 of ~$13,000. Dividing rack revenue by a chip price would roughly double the unit count for that revenue. If 25% of Q3 TPU revenue had been booked at rack value, a chips-only estimate would then be about 14% too high. If half had been booked at rack value, our volume estimates would be about 30% too high.

Broadcom’s statements in 2025 suggested rack sales. In September 2025, Broadcom said its fourth XPU customer had placed “over $10 billion of orders of AI racks based on our XPUs.” In December 2025, it named that customer as Anthropic and called the deal “a real system sale.” The CFO added that Broadcom would be “passing through more components that are not ours,” which would lower its gross margin.

However, Broadcom leadership later gave statements that suggest they are selling only chips. Instead of producing the racks directly, Broadcom could be partnering with a third party to take the TPU chips and networking components and assemble them into racks.

• Broadcom CEO said chips only: On the fiscal Q2 2026 call in June, an analyst asked about Broadcom’s production of racks versus chips and how that affects their gross margins. CEO Hock Tan answered, “No racks. It is all a chip business only. We only [do] chips. Only chips.” The CFO added, “We’re in the chip business only.”
• The rack seller is a separate party: Broadcom’s Q3 and Q2 10-Q details that in June 2026 Broadcom arranged for a financial partner, Apollo, to take on “certain agreements to purchase AI racks based on custom AI accelerators designed by us.” Apollo also took on the related lease agreements with the customer, and Broadcom backstops up to about $29 billion of the lease obligations. The wording suggests Broadcom was originally set to buy the racks from their manufacturer and lease them to Anthropic, and then handed off both roles to Apollo. That leaves Broadcom as the chip supplier and guarantor rather than a seller of racks. Broadcom also listed “reselling the AI racks back to the seller” as a risk-mitigation option, implying the seller is a different company.
• Shipments went mostly to Asia: Broadcom notes revenue when the title transfers and reports it by shipment location. Subtracting VMware license revenue from Q3 hardware revenue leaves $20.8 billion, of which $19.9 billion shipped to Asia Pacific, where TPU racks are assembled. Americas hardware revenue was only $0.46 billion, down from $0.73 billion in Q2. If Broadcom produced and sold the racks themselves, the rack deliveries to US data centers would then raise the Americas figure beyond the reported figure.
• Margins: In September, management attributed gross margin pressure to the growing share of XPUs and their “increasing memory content.” It no longer mentioned the rack pass-through described in December.
• Growth came through Broadcom’s usual chip channel: In Broadcom’s 10-Q filing, they note that sales to one distributor rose from 42% of Broadcom’s revenue in Q2 to 50% in Q3, an increase of roughly $5.5 billion. This roughly matches the ~$5.5 billion increase in XPU revenue, which is consistent with the TPUs for Anthropic being sold through the same channel as Broadcom’s other chips.

Our central case assumes that the revenue that Broadcom books in its XPU segment is from sales of TPU chips rather than racks. However, we are still uncertain about this. Broadcom’s filings still describe its products as including “in some cases, racks.” In September, Hock Tan said customers request capacity “in the form of chips or in some cases, even in the form of racks.” Broadcom does not disclose revenue from Anthropic.

The effect size of Broadcom selling Anthropic racks rather than chips on our volume estimates depends on Anthropic’s share of TPU revenue. We estimate this by assuming that all TPU revenue before Q3 came from Google and extending Google’s recent spending trend. Over the previous four quarters, our estimate of Broadcom’s TPU revenue grew 12–26% per quarter. Applying that growth to Q2 implies Google spent roughly $6.0–6.8 billion in Q3. Subtracting this from our Q3 TPU revenue estimate of about $11.2 billion leaves around $4.4 - $5.2 billion, or around 40% of TPU revenue driven by Anthropic. We use a range of 30–50% to allow for uncertainty in both Q3 revenue and Google’s growth.

Broadcom also said it would deliver 1 GW of TPU compute to Anthropic in 2026, which roughly bounds the number of chips Anthropic receives in total. The main uncertainty is therefore timing: how much was delivered in Q3 rather than later in the year. A share near the top of our range would mean Q3 was front-loaded. A share near the bottom would mean deliveries are weighted toward later quarters.

In our modeling, we treat Anthropic’s TPUs as chip sales rather than rack sales in the central case, and assign a 25% probability that this revenue was booked at rack value. For Q3 FY26, this lowers the median estimate from 2.05 million to 1.97 million H100e. If later disclosures show otherwise, our Q3 volumes would be revised downward. If Anthropic accounted for 40% of Broadcom’s Q3 TPU revenue and that revenue was for racks at about $25,000 per TPU, then Anthropic would have received about 180,000 TPU v7s rather than the 340,000 we estimate from chip prices. Total Q3 FY26 shipments would then be about 1.67 million H100e, which is 15% below our current median.

Even if TPUs for Anthropic are sold as chips, Broadcom may charge a different price for them than the TPUs sold to Google. For example, Broadcom could charge Anthropic a premium per chip and pay Google a per-chip fee for its TPU IP. If that fee were booked in Broadcom’s cost of revenue, Broadcom’s gross profit per chip would be unchanged, but its gross margin percentage would fall. However, we haven’t found such a fee in Broadcom’s or Alphabet’s filings. SemiAnalysis’s estimated $25,000 per TPU v7 in finished racks also leaves little room for a large premium on the chip itself. A higher chip price for Anthropic would overcount Anthropic’s units in the same way rack sales would. We apply the same price distribution to both.

We convert chip computing capabilities into H100 equivalents (H100e) based on their relative FLOP/s specifications, specifically their maximum 8-bit specification. However, different chips have different strengths and weaknesses (e.g., memory capacities and bandwidths, software stability). Depending on the use case, the H100e of a chip can be more or less effective than the H100e implied by its peak performance specs. Conversions are thus approximate, and are most accurate for model training, which is typically compute rather than bandwidth-bottlenecked.

Our unit estimates also depend on how Broadcom books revenue on TPUs for Anthropic. Rack-level revenue would mean our estimates overcount units, as we detailed above.