From Shubhi K | Product & Market Analysis
Broadcom AI Revenue Rose 221% to $16.7B. 6 Buyers Carry the Custom ASIC Bet
On this page
Broadcom's AI semiconductor revenue reached $16.7 billion in its fiscal third quarter of 2026, up 221% in a year. That equals 18.8% of Nvidia's $89.0 billion data center line, against about 12.7% a year earlier. The Broadcom AI revenue story is a real share gain. It also rests on very few buyers: top 5 end customers now make up about 55% of Broadcom's net revenue.
Key takeaways
- Broadcom's AI revenue is growing about twice as fast as Nvidia's data center line. Broadcom reported $16.7 billion, up 221%, for the quarter to 2 August 2026. Nvidia reported $89.0 billion, up 117%, for the quarter to 26 July 2026.
- Custom accelerators were 73% of Broadcom's AI revenue. That is about $12.2 billion by our arithmetic, from a stated share on the earnings call. Broadcom's release gives only the $16.7 billion total, with no split.
- Broadcom's top 5 end customers reached about 55% of net revenue. The share was about 40% a year earlier, per its 10-Q. In dollars, that is roughly $16.3 billion against $6.4 billion, so those 5 buyers supplied about 73% of the group's revenue growth.
- Only 25% of Broadcom's $179.2 billion of committed obligations falls in the next 12 months. That is about $44.8 billion, which equals 39% of the $115 billion AI outlook management gave for fiscal 2027, and the filing covers software as well.
Broadcom's AI revenue is about one fifth of Nvidia's data center revenue and growing about twice as fast. In quarters that ended 1 week apart, Broadcom reported $16.7 billion and Nvidia $89.0 billion. Nvidia remains 5.3 times larger. Broadcom's disclosed buyer base is far narrower, and that is the risk to price.
What Broadcom's $16.7 billion of AI revenue contains
Broadcom's AI semiconductor revenue is one line in the earnings release, and everything finer comes from the call. The release dated 2 September 2026 gives total revenue of $29.6 billion, up 86%. CEO Hock Tan adds the $16.7 billion AI figure, up 221% on the year and 54% on the prior quarter.
Custom accelerators are the bulk of it
Broadcom calls its custom accelerators XPUs. These are chips designed with one customer for that customer's own workloads. Google's TPU is the best known example. On the earnings call, Tan said XPU shipments represented 73% of AI revenue during the quarter. That is about $12.2 billion by our multiplication.
The remaining 27%, or about $4.5 billion, is mostly AI networking. Broadcom does not label the remainder, so treat networking as the inference it is. Tan said AI networking revenue rose over 2.5 times in a year, slower than XPU shipments at over 3.5 times. He also said networking should "grow just as fast as XPUs over the next few years."
AI now accounts for most of the group's growth
Total revenue rose by $13.6 billion year over year. AI semiconductor revenue rose by about $11.5 billion, from an implied $5.2 billion. So AI supplied about 84% of Broadcom's growth. The rest of the semiconductor business, about $4.1 billion by subtraction from the $20.8 billion segment figure, is a much smaller engine than it was. The infrastructure software segment still delivered $8.8 billion, up 29%.
Broadcom vs Nvidia: the AI revenue comparison, measured
Nvidia is still more than 5 times Broadcom's size in AI, but Broadcom is taking a larger share of each new dollar. Compare the two most recent reported quarters, which ended 1 week apart. Nvidia's release of 26 August 2026 reports data center revenue of $89.0 billion, up 117%.
| Measure | Broadcom (to 2 Aug 2026) | Nvidia (to 26 Jul 2026) |
|---|---|---|
| AI revenue line | $16.7B AI semiconductors | $89.0B data center |
| Growth on the year | 221% | 117% |
| Growth on the quarter | 54% | 18% |
| Largest component disclosed | XPUs, 73% of AI revenue | Hyperscale, about $49B of data center |
| GAAP gross margin, whole company | About 69.1%, our calculation | 75.0% |
| Where Nvidia still wins | Sells to 6 XPU customers, 4 discussed by name | Sells across hyperscalers, AI clouds, industrial and enterprise buyers |
Revenue and growth figures are each company's release. The 73% and the $49B are from each earnings call. The Nvidia call transcript used here is machine-generated. Broadcom's gross margin is $20,456 million divided by $29,591 million.
Growth rates flatter the smaller company. Dollars added do not. Over the year, Broadcom added about $11.5 billion of quarterly AI revenue while Nvidia added about $48.0 billion. That is 24 cents of Broadcom growth per Nvidia dollar.
The sequential number is sharper. Broadcom added about $5.9 billion on the prior quarter. Nvidia's data center line added about $13.8 billion. That is 43 cents per dollar, in a single quarter. If I had to pick one figure that captures the custom ASIC story, it would be that one.
It covers one quarter, and Broadcom's jump was unusually large. Do not extrapolate it. For the other half of this comparison, the analysis of what Amazon discloses about Trainium shows how little most custom silicon programmes reveal at all.
Custom AI chips carry a lower gross margin than the headline suggests
Broadcom's custom AI chip mix is pulling its gross margin down, and the headline 75% hides how far. The company reported non-GAAP gross margin of 75.0% for the quarter, against about 78.4% a year earlier by our calculation from its reconciliation tables. CFO Amie Thuener guided the fourth quarter to about 73%.
The stated reason is mix. She pointed to "the increasing mix of XPUs with their increasing memory content." Memory that Broadcom buys and passes through adds revenue with little margin attached. So the fastest growing product is also the one that dilutes the ratio.
The semiconductor segment sits near 67%
The transcript contains a correction worth catching. In prepared remarks, the CFO's semiconductor gross margin was given as about 76%. In the question session, she stated it as approximately 67%. The arithmetic supports the lower figure. With semiconductors at 70% of revenue and software at about 94% margin, 67% blends to 75%.
Nvidia's margin is not on the same basis
Nvidia guided its next quarter to gross margin of 74.0%, plus or minus 50 basis points. Its non-GAAP measure has included stock-based compensation since the first quarter of fiscal 2027. That makes GAAP the fairer comparison: 75.0% for Nvidia against about 69.1% for Broadcom. Tan's own response was "stop focusing on gross margin," since he expects to sustain operating margin at about 66%. That is a defensible position. It is also what a seller says when gross margin is falling.
Broadcom's customer concentration rose as its AI revenue rose
Broadcom's customer concentration climbed sharply in a single year, and the rise tracks AI revenue almost exactly. Its 10-Q for the quarter ended 2 August 2026 estimates that sales through all channels to its top 5 end customers were about 55% of net revenue. A year earlier the share was about 40%.
In dollars, that is roughly $16.3 billion against $6.4 billion. The difference, about $9.9 billion, is 73% of the group's $13.6 billion revenue increase. Those figures are our multiplication of rounded percentages, so read them as approximate. Across the 3 fiscal quarters to August the top 5 share was about 50%, so the quarter is running above the year to date.
A distributor is not a buyer
The same filing says direct sales to 1 semiconductor solutions customer, a distributor, were 50% of net revenue for the quarter, up from 32%. That number looks alarming and it is the less important of the two. A distributor is a routing path. The top 5 end customer estimate is the one that describes dependence, and it is notably an estimate, not a ledger figure.
The named gigawatts point to 2 frontier labs
Tan said "Our 6 XPU customers," and that "4 of them are just simply going to be huge." He discussed 4 by name. He did not say the 4 huge ones are those 4, though the gigawatts suggest it.
| Customer | Stated deployment | Tan's words |
|---|---|---|
| Anthropic | 1 GW in 2026, 5 GW in 2027, an incremental 10 GW for 2028 | "on track to become our largest XPU customer in 2027" |
| OpenAI | 1.3 GW in 2027, over 5 GW for 2028 | "would make OpenAI our second largest XPU customer" |
| Meta | 3 GW through 2028 | "line of sight" |
| No gigawatt figure given | "multi-tens of billions of dollars of TPUs annually" | |
| 2 others | Not named or sized | Not discussed individually |
Source: Motley Fool transcript of the call, 9 September 2026. Gigawatt phrasing differs by customer, with "line of sight" being a weaker commitment than a planned deployment. Timing refers to calendar deployment as described by management.
Stacy Rasgon, the analyst on the call, summed these to about 10 GW in 2027 and about 20 GW in 2028. On those totals, Anthropic alone is half of the listed capacity in both years. Anthropic and OpenAI together are 63% in 2027 and 75% in 2028. I think this is the real concentration number, and it is worse than the 55% headline, because neither lab is a self-funding hyperscaler yet.
The buyer's filing shows the other side. Anthropic's Form S-1 describes 5-year master equipment leases with Broadcom, with aggregate rent of about $161.2 billion excluding interest, non-cancelable except on default. Its loss profile, covered in the reconciliation of Anthropic's reported losses, is the credit behind that rent.
Nvidia's concentration is not directly comparable. It discloses direct customers, not end customers. In the same quarter one direct customer took 16% of revenue and 5 direct customers held 70% of receivables. Most of those direct buyers are server assemblers.
The $179.2 billion backlog and what it leaves out
Broadcom's remaining performance obligations are large, but the filing warns against reading them as future revenue. The 10-Q puts remaining performance obligations at about $179.2 billion on 2 August 2026, and it says those commitments include "a long-term contract for custom AI accelerators" signed in the quarter ended 3 May 2026.
Broadcom expects about 25% of the total to become revenue in the next 12 months. That is about $44.8 billion by our multiplication. Compare it with the $115 billion AI outlook for fiscal 2027 and the coverage is 39%, at most.
The exclusions are the point
The filing excludes contracts with an original duration of 1 year or less. It excludes software and services contracts where customers can terminate for convenience. It also excludes royalty-based licensing. It then states that these obligations "are not indicative of revenue for future periods." So do not read the 61% gap as unordered product. Short-dated purchase orders sit outside this number. They are simply invisible to you.
One contract may dominate the balance
Here is the inference I would test first. Broadcom does not name the customer behind the long-term XPU contract. Anthropic's S-1 describes a March 2026 arrangement to procure AI accelerators, in the same fiscal quarter, with rent of about $125.2 billion excluding interest. That filing also gives the $161.2 billion aggregate for the master leases. Against $179.2 billion, those are 70% and 90%. If the contract is Anthropic's, then most of the disclosed backlog is one customer. Neither filing says so, and Broadcom has assigned part of its rights to a third party. It remains the most useful open question in this filing.
For comparison across the industry, the side-by-side of cloud backlog disclosures shows how differently each company defines the same line.
Broadcom and Nvidia guidance language, read side by side
Broadcom's guidance is more specific about AI than Nvidia's, and looser about what would change it. The two companies now describe their futures in different grammar, and the difference matters for anyone relying on either.
| Item | Broadcom | Nvidia |
|---|---|---|
| Next quarter | AI semiconductor revenue of $21.7B, up 236%. No range. | Total revenue of $108.0B, plus or minus 2%. No data center guide. |
| Longest horizon | Fiscal 2028 AI revenue of $230B | Fiscal 2028 total revenue growth of about 70%, "preliminary" |
| Supply wording | "we have secured the supply" | "This is a supply-constrained outlook" |
| Demand claim | "Our demand actually exceeds this outlook" | "Our demand is much higher than that" |
| Update policy | "we do not intend to update it on a quarterly basis" | Updated each quarter through the normal guide |
| Where Nvidia's wording is better | Calls the figure "a judged number" | Gives a dated quarterly number with a stated error band |
The $115 billion for fiscal 2027 sits only 32% above the fourth quarter guide annualised, which is $86.8 billion. The bigger claim is the second doubling to $230 billion. Tan described the outlook as "a carefully structured outlook," and CFO Thuener framed it as a trajectory rather than a forecast to be maintained.
Tan put Broadcom's content at $20 billion to $30 billion per gigawatt. Nvidia's Colette Kress cited $25 billion with Blackwell and $40 billion with Vera Rubin. At Broadcom's range, $115 billion implies about 3.8 to 5.8 GW of revenue-equivalent shipments in fiscal 2027, against about 10 GW on the customer list. Rasgon's own estimate was $11 billion to $12 billion per gigawatt. Tan said directly, "we're not saying" 30 GW goes into production across the 2 years. A customer list is not a shipment schedule, and you should not value it as one.
Broadcom is now financing part of its own demand
Broadcom has begun standing behind some of the AI racks its customers lease, which ties its credit exposure to the same few buyers. The 10-Q describes a residual value backstop: on a lease default, Broadcom covers any gap between 85% of the amounts owed and what the racks fetch on sale. The filing says the fair value of the backstop was not material.
The ceiling is another matter. Broadcom's 10-Q for the May quarter put maximum exposure at $29 billion on a June 2026 arrangement. An 8-K of 26 October 2026 adds a new vehicle with a maximum potential liability of about $43 billion, undiscounted, on 72.5% of the amounts owed. Added together that is about $72 billion, our sum and not a figure Broadcom publishes. Broadcom calls it a contingent liability with a low probability of occurring.
Nvidia's 10-Q for the July quarter reports maximum gross guarantee exposure of $108.5 billion, including guarantees capped at $105 billion tied to an affiliate of OpenAI. The route money takes through these structures is traced in the filings trace of Nvidia's OpenAI investment. Both chip suppliers are now underwriting some of their largest buyers.
On the call, Tan said only Anthropic and OpenAI receive financing vehicles, and "Long term, we see these guys to be no different from a hyperscaler." That may prove right. Today it is a forecast about 2 companies' future credit. Broadcom's own risk factors warn that top customers are pushing for leases and deferred payment, and that customers moving to their own tooling could reduce demand for its products.
Where this reading of Broadcom is weakest
This post leans on a concentration argument, and the concentration argument has real holes. State them before you act on it.
Concentration may be the wrong risk to fear
Every large AI chip supplier is concentrated. Buyers able to spend this much are few by definition. Tan's argument is that 4 of his customers are "financially secure, stable enough to be able to fund it themselves," and a large share of the demand is contracted under non-cancellable leases. Google has been a customer for years. Anyone calling this fragile has to explain why a multi-year, non-cancelable lease is weaker than Nvidia's purchase orders.
The source material has gaps
The two earnings transcripts used here are third-party products, and the Nvidia one is machine-generated. The Broadcom transcript carries its own warning about errors. Broadcom does not publish an XPU or networking revenue line, so the $12.2 billion and $4.5 billion rest on one spoken percentage. Nvidia does break out networking, but we could only find a dollar figure in secondary write-ups, so this post does not use one.
The comparison of AI lines is also imperfect. Nvidia's data center line includes complete systems and software, and Broadcom's AI line excludes the memory-free content of its racks only as far as its reporting allows. Treat 18.8% as a directional ratio. It is the right order of magnitude and not a market share.
Frequently asked questions
How much AI revenue did Broadcom report in fiscal Q3 2026?
Broadcom reported AI semiconductor revenue of $16.7 billion for the quarter ended 2 August 2026, up 221% year over year and 54% on the prior quarter. Total revenue was $29.6 billion. Management said custom accelerators were 73% of AI revenue. For the fourth quarter, Broadcom guided AI semiconductor revenue of $21.7 billion, up 236% year over year.
How does Broadcom's AI revenue compare with Nvidia's?
Broadcom's $16.7 billion of AI semiconductor revenue equals about 18.8% of Nvidia's $89.0 billion data center revenue for quarters ended 1 week apart. Broadcom grew 221% year over year and Nvidia 117%. Nvidia is still more than 5 times larger. The two lines contain different products, so the ratio is directional rather than a precise market share.
Who are Broadcom's custom AI chip customers?
Broadcom says it has 6 XPU customers. On its fiscal Q3 2026 call, management discussed Google, Anthropic, OpenAI and Meta by name and did not name the other 2. Hock Tan said Anthropic is on track to become the largest XPU customer in 2027 and that OpenAI's planned deployments would make it the second largest.
How concentrated is Broadcom's revenue?
Broadcom estimates that its top 5 end customers, through all channels, were about 55% of net revenue in the quarter ended 2 August 2026, up from about 40% a year earlier. One distributor accounted for 50% of net revenue, up from 32%. The distributor is a sales channel, so the end customer estimate is the better measure of dependence.
What is Broadcom's AI revenue outlook for 2027 and 2028?
On its September 2026 call, Broadcom outlined about $58 billion of AI revenue for fiscal 2026, about $115 billion for fiscal 2027 and $230 billion for fiscal 2028. Management said supply is secured for those levels and that it does not plan to update the outlook quarterly. Its filed obligations cover only a part of that amount.
Where to start if your forecast assumes this growth
If your plan depends on custom silicon pricing or AI compute supply, pick up one number before the next report on 9 December 2026. Open Broadcom's next filing and go straight to the top 5 end customer estimate. If it rises again while AI revenue rises, the custom ASIC business is narrowing as it scales.
Then check the remaining performance obligations and the share due in 12 months against the outlook. Ask your own compute vendor which silicon your contract runs on, and whether its price survives a change in that supply. A supplier who cannot answer is exposed to the same 6 buyers you just read about.
References
- Broadcom, Form 8-K exhibit 99.1, fiscal Q3 2026 results, 2 September 2026. Used for revenue, segments, AI revenue and fourth quarter guidance.
- Broadcom, Form 10-Q for the quarter ended 2 August 2026. Used for concentration, remaining performance obligations, risk factors and the backstop. The $29 billion exposure is from its 10-Q for the quarter ended 3 May 2026.
- Broadcom, Form 8-K, 26 October 2026. Used for the $43 billion backstop.
- The Motley Fool, Broadcom Q3 2026 earnings call transcript, 9 September 2026. Used for XPU share, customer statements, outlook and margin. Partly LLM-assisted, per the publisher.
- Nvidia, second quarter fiscal 2027 results, 26 August 2026. Used for data center revenue, gross margin and third quarter outlook.
- Nvidia, Form 10-Q for the quarter ended 26 July 2026. Used for receivables concentration and guarantee exposure.
- Nvidia Q2 fiscal 2027 call, machine-generated transcript, August 2026. Used for hyperscale revenue, fiscal 2028 language and revenue per gigawatt.
- Anthropic, Form S-1, 2026. Used for the Broadcom equipment leases.
Last verified 8 November 2026. Weakest point in the source base: the XPU split, customer gigawatts and outlook rest on third-party call transcripts, and several 10-Q passages were read as excerpts rather than the complete filing. Refresh after Broadcom's report on 9 December 2026.
Related reading