Broadcom in the AI stack

The business and what the filings show

Broadcom is the main design partner for hyperscalers building their own accelerators, and the leading supplier of the Ethernet switch chips used in AI networks. Its semiconductor segment earned $59.4B of revenue over its latest four quarters, +75%; the company as a whole earned $89.1B at a 48% operating margin, helped by its infrastructure software business.

Custom silicon is a business of very few, very large customers. Broadcom says its top five end customers accounted for about 55% of net revenue in its latest quarter, and one distributor accounted for half. Each hyperscaler that commits to its own chip is a long design relationship, but also a concentration risk.

For investors, Broadcom is the main listed way to own the custom-silicon side of the accelerator contest. It wins whether Nvidia’s share falls or not, as long as the buyers keep designing their own chips, and its networking business grows with the shift to Ethernet in AI clusters.

How AI shows up in the numbers

AI shows up at Broadcom as custom accelerator revenue from a few hyperscalers and as Ethernet switching and connectivity for AI clusters. Its infrastructure software business, built through acquisitions, adds steady high-margin revenue that cushions the lumpiness of chip programs, so the semiconductor segment is the part to read for AI.

For scale, Broadcom’s revenue over the latest four quarters equals 16% of the $566B the five largest buyers spent on capital projects over their latest four quarters. Their capex rose +81% on the year before, while Broadcom’s revenue grew +49%: slower than the spending, so it is capturing a smaller share of each new capex dollar, or more of its business sits outside the build-out.

How it talks about AI

In its latest earnings release, furnished to the SEC on September 2, 2026, Broadcom mentioned AI 3 times in 1,011 words of narrative text, about 3.0 mentions per 1,000 words. Of its 3 sentences about AI, 2 included a number: a revenue figure, a growth rate, a customer count or another measurable result. Across the 8 releases we read, back to December 12, 2024, AI mentions per 1,000 words went from 2.7 to 3.0.

Specific claims are easier to check against the numbers than general enthusiasm, which is why the share of AI sentences that carry a figure is worth watching alongside how often AI is mentioned. The sentences below are quoted from the release as published, after removing safe-harbor text and company boilerplate.

• “Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter," said Hock Tan, President and CEO of Broadcom Inc.”
• “In Q4 the momentum continues, and we expect AI semiconductor revenue to accelerate to $21.7 billion, up 236% year-over-year.”
• “Demand for our custom AI accelerators and networking continues to be very strong.”

Source: earnings release (Form 8-K, Exhibit 99.1), September 2, 2026.

The trend in the numbers

Since the start of 2023, Broadcom’s largest quarter by revenue was Q2 2026 at $29.6B, and its highest gross margin was 70% in Q1 2023. The latest quarter set the revenue record for the period.

Quarter by quarter, Broadcom’s revenue grew +28% in Q3 2025, +29% in Q4 2025, +48% in Q1 2026, +86% in Q2 2026 on the same quarter a year earlier. Growth has been accelerating.

Gross margin in the latest quarter was 69%, against 67% a year earlier, so it is keeping more of each dollar as it grows. Operating margin moved from 37% to 54% over the same span. Over the latest four quarters it turned 46% of its revenue into operating cash flow and spent 1% of revenue on capital projects, against 1% a year earlier.

Segments that matter for AI

The parts of Broadcom’s reported results most tied to the AI build-out, from the segment and product disclosures in its 10-K and 10-Q filings.

Where it sits in the stack

Broadcom sits in the chips layer of the AI stack. It is paid by the AI clouds and servers and networking layers; its own spending flows to foundry & tools, memory.

Among the 7 companies in this layer on our map, it ranks number 2 by revenue and number 2 of 6 by revenue growth. Its gross margin of 69% compares with 66% for the layer as a whole, close to the layer average. The full layer analysis is on the chips page.

Latest four quarters for 10-Q filers; latest fiscal year for 20-F filers. Companies under $100 million of revenue are left out.

Against the 6 US-listed companies in its layer, Broadcom grew +49% over the latest year versus +52% for the layer combined, and converted 44% of revenue into free cash flow versus 35% for the layer. It is growing roughly in line with its layer.

What has to go right, and what could go wrong

What has to go right. Hyperscalers have to keep committing to their own accelerators, and those programs have to keep scaling, while the shift to Ethernet in AI clusters keeps its switch chips in demand.

What could go wrong. A custom-silicon business with a few customers can lose a large share of revenue when one program ends or moves to another design partner, and buyers could slow chip programs if capex flattens.

These describe the conditions the business depends on, as the filings show them. They are not a view on the stock.

Through the lens of Investing in AI

Rob May’s book offers six mental models for telling AI moats from traps. The one most useful for Broadcom’s layer is “AI changes markets and corporate strategy”: When creation gets cheap, choke points move. Like the internet shifting newspaper profits to discovery, AI pushes moats toward coordination and control points. The winners sit at the new choke point, not the best AI story.

The question it asks of any company is: What becomes scarce in this value chain as AI makes the work abundant? Applied to Broadcom, the filings on this page are the place to start: the margin trend shows whether it holds scarce value or passes it through, the customer and financing disclosures show who it depends on, and the commitments show what it has promised before the revenue arrives.

Customers and concentration

US companies must disclose any customer that accounts for 10% or more of revenue, and foreign filers usually describe their largest customers. Concentration matters because a company’s growth is only as durable as the handful of buyers behind it, and in the AI stack those buyers are often the same few companies.

• “We believe aggregate sales to our top five end customers, through all channels, accounted for approximately 55% and 50% of our net revenue for the fiscal quarter and three fiscal quarters ended August 2, 2026, respectively, and approximately 40% of our net revenue for each of the fiscal quarter and three fiscal quarters ended August 3, 2025.” 10-Q, September 10, 2026

Commitments and backlog

• Backlog. Broadcom reports $179B of remaining performance obligations, 2.0 times its revenue over the latest four quarters: contracted revenue not yet earned.
• Purchase commitments. $127B of unconditional purchase obligations as of August 2, 2026 (10-Q): what it has contracted to buy from its own suppliers.

What to watch

Broadcom’s next quarterly report will cover the quarter ending around November 1, 2026. Large companies must file a 10-Q within 40 days of a quarter’s end and a 10-K within 60 days of the fiscal year’s end, so the new figures should arrive by about December 31, 2026 at the latest. This page, the layer pages and the next quarterly report all update from that filing.

• Semiconductor segment growth and what the company says about AI revenue, the read on custom accelerator demand.
• Customer concentration disclosures, which show how much rides on a few hyperscaler programs.
• Remaining performance obligations, an early signal of future custom-chip and networking orders.
• Its place in the next quarterly AI Stack Report, which updates every company on the map from the new filings.

Figures from Broadcom’s SEC filings, latest period ended August 2, 2026. Other companies: all company pages.