On September 2, 2026, Broadcom announced that AI semiconductor revenue reached $16.7 billion in the third quarter of fiscal 2026. This figure was more than triple the year-ago quarter, up 54% from the previous quarter, and represented 56.4% of total company revenue of $29.591 billion. The growth is driven by custom XPUs co-designed with six companies, including Google and OpenAI, and by the Ethernet products that connect them.
While expanding its chip supply, Broadcom also signed an agreement to backstop lease obligations tied to a deal in which a third party purchases AI racks and leases them to customers. However, the financing support for OpenAI and Anthropic that the company discussed on September 2 remains under consideration and is not a new contract. Similarly, the long-term outlook of roughly $115 billion in AI semiconductor revenue for fiscal 2027 and $230 billion for fiscal 2028 is a company forecast that accounts for customer facility readiness rather than confirmed orders.
Custom XPUs Drove Three-Quarters of the $16.7 Billion
AI semiconductor revenue of $16.7 billion in Q3 FY2026 represented 56.4% of total company revenue of $29.591 billion and 80.1% of semiconductor segment revenue of $20.839 billion. Applying the 73% product mix disclosed during the earnings call, custom XPUs generated an estimated $12.191 billion, while AI networking products accounted for approximately $4.509 billion.
| Revenue Category | Q3 FY2026 | YoY Change | Share of Total Revenue |
|---|---|---|---|
| AI Semiconductors | $16.7 billion | +221% | 56.4% |
| Total Semiconductor Segment | $20.839 billion | +127% | 70% |
| Infrastructure Software | $8.752 billion | +29% | 30% |
| Total Company | $29.591 billion | +86% | — |
Two growth trends overlapped within AI semiconductors. Custom XPU shipment volume grew more than 3.5 times year-over-year, while AI networking revenue grew more than 2.5 times. However, growth rates for shipment volume and revenue are not identical, since per-unit pricing and product mix across generations are not disclosed. The $16.7 billion figure is not custom chip revenue alone—it also includes networking products such as switches and optical interconnects.
AI semiconductor revenue in the previous quarter was $10.8 billion. Since Broadcom had guided for $16 billion in Q3 back in June, the actual result exceeded guidance by $700 million, or about 4.4%. Consolidated non-GAAP gross margin fell 2.1 percentage points from the prior quarter to 75%, while non-GAAP operating margin rose 2.4 percentage points year-over-year to reach 67.9%. This disclosure alone does not allow us to isolate how much product mix contributed to each of these two metrics.
Customer-Specific Roadmaps Across Six Companies
While Broadcom counts six companies as custom XPU customers, it has disclosed the names and deployment plans of only four. The remaining two are undisclosed, and their identities should not be filled in through speculation.
| Customer | Current Status | Company Outlook for 2027–2028 | Contractual/Planning Boundaries |
|---|---|---|---|
| Mass-shipping TPU v7 "Ironwood"; began production shipments of TPU 8i | Plans to supply tens of billions of dollars worth of TPUs annually over the coming years | Contracts for future TPU generations and networking extend through 2031 at the latest | |
| Anthropic | Deploying 1GW of Ironwood in 2026 | 5GW of TPU 8i in 2027, plus an additional 10GW in 2028 | Outlook has expanded from the roughly 3.5GW disclosed in April, though usage depends on commercial success |
| OpenAI | Began shipping first-generation Jalapeño | 1.3GW in 2027; over 5GW including successor chips by 2028 | The 10GW figure represents a multi-generation plan through the end of 2029 |
| Meta | Co-developing MTIA for inference and recommendation workloads | Three generations by end of 2027; total of 3GW by 2028 | Plan to scale from an initial 1GW+ to multiple GW by 2029 |
This table does not represent "contracted revenue." While the long-term agreement with Google forms a foundation for supply, the GW figures for Anthropic, OpenAI, and Meta include forward-looking projections. The 10GW plan with OpenAI does not mean the first-generation Jalapeño will reach 10GW of operational capacity within 2026—rather, it represents a goal of deploying multiple generations by the end of 2029.
In addition to custom XPUs, Broadcom counts Ethernet switches and optical interconnects as AI networking revenue. Based on the Q3 product mix, this can be estimated at approximately $4.509 billion, though the company has not disclosed adoption rates by customer or revenue breakdowns by product. The fact that both custom XPUs and networking grew should be viewed separately from adoption status for individual products.
Preparing Racks and Financing Before the Sale
Broadcom's support for AI labs advanced on June 8, 2026, into a backstop agreement covering up to $29 billion in five-year rack lease obligations. During the September 2 earnings call, the company stated it would evaluate future deals with OpenAI and Anthropic on a case-by-case basis, but did not announce any newly finalized contracts regarding residual value guarantees.
In this arrangement, investor Apollo took on both the purchase contract for AI racks equipped with Broadcom-designed custom XPUs and the lease agreements with customers. If a customer fails to make lease payments over the five-year term, Broadcom provides backing. The $29 billion figure is neither a current loss nor an expenditure—exposure grows as racks are deployed and shrinks as customers make payments. In the event of default, Broadcom also has options to assume the lease or sell the racks.
SEC filings do not disclose the identity of this leasing customer, so it should not be equated with OpenAI or Anthropic. What management newly indicated during the earnings call is a policy of considering similar investor-mediated financing support for future deals with both companies. If residual value guarantees are added, the risk of asset devaluation upon future rack resale could also fall on Broadcom, though contract amounts and terms have not yet been determined.
For Broadcom, this adds a credit risk separate from chip sales. The 10-Q explicitly states that losses could arise depending on customers' ability to pay and the resalability of custom-designed products. While the $29 billion figure is not a loss amount, it represents contingent liability that accumulates as racks are deployed—going forward, it will be necessary to track contracted amounts and actual exposure separately.
Land and Supply Chain Constraints on the $115 Billion Outlook
Broadcom projects Q4 AI semiconductor revenue of $21.7 billion and full-year fiscal 2026 revenue of $58 billion. This full-year guidance was raised by $2 billion from the $56 billion figure given in June. The company further indicated an outlook of approximately $115 billion for fiscal 2027, doubling to $230 billion in fiscal 2028.
This long-term outlook is not a mechanical conversion of customers' desired GW capacity into revenue. While the six companies' desired computing capacity totals roughly 30GW across fiscal 2027–2028, CEO Hock Tan explained that the company does not assume all of this capacity will become operational within two years. Broadcom stated that it checked land, power, and building readiness on a customer-by-customer basis and estimated revenue based on the amount that could realistically be deployed.
Multiple bottlenecks also exist on the chip side. Beyond advanced logic, racks cannot be completed without sufficient HBM, system memory, substrates, and optical communication lasers. Broadcom has built substrate production capacity in Singapore and expanded indium phosphide optical component factories in the U.S. and Singapore to more than three times their capacity from the previous year. Even when the company states it has "secured supply," this does not guarantee customers' power infrastructure, financing, or procurement of every component.
Management estimates Broadcom's product revenue at $20–30 billion per GW. However, this company benchmark represents neither total data center construction costs nor AI lab revenue, and the calculation method for different product mixes has not been disclosed. To assess the $115 billion figure for fiscal 2027, it will be necessary to track actual operational GW capacity at completed facilities, rather than order requests.
Competition with NVIDIA Won't Be Decided by Chips Alone
Broadcom's custom XPUs are premised on co-design with Google, Anthropic, OpenAI, and Meta. However, this earnings report does not indicate that NVIDIA's software foundation, including CUDA, has been displaced, nor that overall GPU demand in the broader market has declined.
The scope of competition has expanded from custom XPUs to networking and rack financing deals. However, it has not been disclosed that all customers have adopted the same products or financing support. In the first half of fiscal 2026, the top five end customers accounted for approximately 45% of total company revenue—a structure in which delays or credit deterioration at a small number of companies could readily impact results.
At the full-year earnings report scheduled for December 9, the first thing to verify will be whether the company achieved its Q4 targets of $21.7 billion in AI semiconductor revenue and $34.8 billion in total company revenue. The key indicators for gauging the credibility of the $115 billion fiscal 2027 figure will be completed data centers and actual operational GW capacity. Separately regarding financing support, it will be necessary to track whether the counterparties, amounts, and guarantee terms for deals with OpenAI and Anthropic become finalized.