Broadcom Inc. (NASDAQ: AVGO)

Figures from the FY2025 10-K, the FY2026 Q3 10-Q filed 10 September 2026, and the Q3 FY2026 earnings release and call. Broadcom does not disclose AI revenue in any 10-K or 10-Q; it appears only in quarterly releases and on calls.

1. What the business is

Two very different customers have the same problem: they need infrastructure that a general-purpose product does not deliver economically at their scale. A hyperscaler running one enormous, well-understood workload can buy merchant GPUs, or it can commission silicon shaped around that workload and own the resulting cost advantage. A large enterprise running thousands of applications on its own servers needs software that turns a rack of machines into a pool of capacity, and needs it to keep working with every storage array, network card and application it already owns. Broadcom sells to both, and the two businesses are close to equal contributors: semiconductors were 58% of FY2025 revenue and software 42%, but software produced 49% of the two segments’ combined operating income.

2. Tracing one unit

Follow a single XPU from design to cash. The customer specifies the architecture and funds the non-recurring engineering. Broadcom designs the system-on-chip around its own serialiser/ deserialiser, optical and packaging intellectual property. TSMC fabricates the wafers — approximately 95% of every wafer Broadcom’s contract manufacturers produced in the first half of FY2026. Title on most products transfers in Penang, Malaysia, and the parts are sold largely through a single distributor, which accounted for 42% of net revenue in Q2 FY2026 against 32% for all of FY2025 and 21% in FY2023.

What that unit earns became visible only in the FY2025 10-K, which adopted the expanded segment-expense standard. Semiconductor segment revenue of $36,858m carried $11,740m of segment cost of revenue — a 68.1% gross margin — and produced $21,232m of segment operating income. The software segment turned $27,029m of revenue into $20,765m of operating income on $1,902m of cost. Because these segment lines exclude acquisition amortisation and stock compensation, they describe the products rather than the reported income statement.

Broadcom sells the parts of what it buys that do not clear its threshold: Brocade’s IP networking business for $800m, CA’s Veracode for $950m, Symantec’s Cyber Security Services, and VMware’s end-user computing division to KKR for $3.5bn — each described in the filings as “not aligned with our strategic objectives.” The move now running the other way deserves attention precisely because it breaks that pattern.

Racks carry memory, substrates and assembly that pass through at a fraction of silicon’s mark-up. This is the first business in two decades Broadcom has entered because a customer asked rather than because the margin was attractive.

3. Where the profit sits in this industry

The AI data-centre industry sells one thing: usable compute inside a power envelope. Everything else is an input. The profit is distributed along the chain with unusual sharpness — design, IP, foundry and memory earn 58–85% gross margins, while physically assembling the result earns 6–12%. Broadcom sits firmly on the profitable side, and the rack business is a deliberate step toward the other one.

Each figure from that company’s own quarterly release.

Merchant silicon versus custom silicon

NVIDIA sells a part it designed, owns and prices: $89.0bn of data-centre revenue at a 75.0% gross margin in the quarter ended July 2026. Broadcom sells a part its customer specified and paid to develop. It takes less margin per unit, but carries no architectural bet and less inventory risk, and keeps the socket for the platform’s life. That the model is real is now visible outside Broadcom’s accounts: Amazon disclosed a custom-chip run-rate “now over $20 billion” with “over $225 billion in revenue commitments for Trainium,” and Google has shipped seven TPU generations. The competitive question is who supplies it. Marvell is the only comparably scaled alternative at $2.171bn of data-centre revenue and 58.9% non-GAAP gross margin — roughly an eighth of Broadcom’s AI run-rate. Below that sit design houses such as Alchip at 34.8%, the commodity end of the same activity.

What would actually be hard to reproduce

Three things bind. The first is the interconnect IP that moves data between accelerators, which has become harder than the compute itself — evidenced by the fact that every disclosed deal in the field over the past year has been in interconnect rather than compute: Qualcomm bought Alphawave for $2.4bn, NVIDIA put $4bn into Coherent and Lumentum for photonics capacity rights, and Marvell, Credo, Semtech and Infineon all made photonics tuck-ins.

The second is advanced packaging, which is the industry’s binding physical constraint and is not Broadcom’s to control. TSMC’s chief executive said on the 2Q 2026 call that “our packaging capacity is so tight that now it’s limiting my customers’ growth,” that “the gap is very big,” and that the shortage runs through 2029–2030; TSMC raised FY2026 capital expenditure to $60–64bn, approved a further $29.4bn of capital appropriations in August covering advanced technology and advanced packaging, and said in September it is building 25 fabs and packaging facilities this year. Broadcom’s advantage here is priority born of volume and of money committed in advance — section 6 sets out the mechanism.

The third is the design win. Once a customer has committed a rack architecture, a compiler stack and a year of software work to a specific accelerator, switching means redoing all of it. That is why Broadcom can name six XPU customers — Google, Anthropic, OpenAI and Meta among them — and treat the number as the business rather than as a concentration problem. It is both.

Networking is more contested than the AI numbers suggest. Cisco announced its Silicon One G300 at 102.4 Tb/s in February 2026, matching Broadcom’s Tomahawk 6, and NVIDIA reported $14.8bn of data-centre networking revenue in its April 2026 quarter, up 199%. No independent body publishes Ethernet switching-silicon market share, so the size of Broadcom’s lead is unknown rather than established.

The software position, tested against outside evidence

Broadcom asserts that its infrastructure software is mission-critical and sticky. Outside evidence supports the mechanism and is more specific than the company is. When the European Commission cleared the VMware acquisition in July 2023 it required Broadcom to commit to open access to VMware’s APIs and certification programme and to grant an open-source driver — remedies aimed at the I/O Vendor Partner certification programme, because hardware makers depend on it for their adapters to work at all. A competition authority mandating access is the strongest available proof that the certification matrix, not the hypervisor, is the lock.

The durability claim also holds so far. Software ARR grew 15% year on year in Q3 FY2026 with revenue up 29% to $8.8bn. Nutanix, the most direct beneficiary of any exodus, reported FY2026 revenue of $2.85bn — about a twelfth of Broadcom’s software run-rate — growing 12% with 106% net dollar retention and FY2027 guidance implying further deceleration. No competitor has published a quantified VMware displacement rate. The Japan Fair Trade Commission, the only authority to have adjudicated the licensing conduct, closed its investigation on 3 July 2026 finding that “facts sufficient to constitute a violation of the Act were not found.” Where the evidence is weak is on price: CISPE alleges cumulative increases “exceeding 1,000 percent” in a pending Article 102 complaint and is separately seeking annulment of the merger clearance, but no primary source establishes a verified like-for-like price series.

4. Growth: what is real and what was bought

Reported growth over the last three years has three different explanations, and conflating them produces the wrong business.

VMware contribution from the FY2024 10-K acquisition note. 9M FY2026 as filed in the Q3 FY2026 10-Q; the implied year adds Broadcom’s own $34.8bn Q4 guidance and is therefore part reported, part guided. FY2024 growth is flattered twice — by an acquisition and by a 53-week year.

Strip VMware from FY2024 and the underlying business grew 9.4%. FY2025 and FY2026 are different: Broadcom has made no acquisition since a $600m Seagate asset purchase in April 2024, so the 24% of FY2025 and the 55% of the first nine months of FY2026 are organic. This is the first period in a decade in which Broadcom has grown substantially without buying anything.

Net revenue by fiscal year, $bn, split by segment. FY2017–FY2025 from the FY2019, FY2022 and FY2025 10-Ks, using recast segment figures where a filing restated them. FY2026 is the dashed bar: three reported quarters plus Broadcom’s own Q4 guidance, so it is part result and part forecast. The VMware step in FY2024 is visible in the blue.

What is driving it, ranked

5. Margin, cash and where the money goes

Broadcom converts revenue to cash better than almost any company of its size because it owns designs rather than plant. Capital expenditure over FY2018–FY2025 totalled $4,020m against $126,722m of operating cash flow — 3.2%. The capital intensity of the industry sits on TSMC’s balance sheet, which is why TSMC is spending $60–64bn this year while Broadcom spends around $2bn. What Broadcom pays for that arrangement is dependence, not cash.

Years chosen to show the shape of the change, not the full series. Comparability warnings: FY2019 was the first year of CA and used a three-segment presentation later recast; FY2024, not shown, had 53 weeks and the VMware close; FY2025 reclassified $7,800m of upfront licence revenue between the products and subscriptions lines with prior years restated; share counts before July 2024 are restated for a ten-for-one split. 9M FY2026 revenue, margin and cash flow are as filed in the Q3 FY2026 10-Q, the share count is derived from earnings and EPS, and the fiscal-year debt figures are principal while the Q3 FY2026 figure is carrying value.

GAAP gross and operating margin by fiscal year, from the FY2019, FY2022 and FY2025 10-Ks and the FY2026 quarterly filings. The FY2024 dip is VMware acquisition amortisation, not a demand trough; the FY2019 low is the last real semiconductor downturn. Both measures are now above any prior peak.

What actually moves the margin

Gross margin is a mix question, not a cost question. In FY2025 the semiconductor segment carried 68.1% and software 93.0%, so every point of mix that shifts toward semiconductors mechanically lowers the blend. That is what is happening: semiconductors were 58% of FY2025 revenue and 70% in Q3 FY2026. Broadcom’s own explanation is identical — gross margin improved “partially offset by a higher mix of semiconductor solutions net revenue, which has a lower gross margin than infrastructure software.” Non-GAAP gross margin fell 210 basis points sequentially in Q3 FY2026 to 74.9% for this reason, while non-GAAP operating margin rose.

The largest non-cash cost is stock compensation, $7,568m in FY2025 — 11.8% of revenue, up from 6.1% in FY2023 — because Broadcom granted two-year equity awards in Q2 FY2025 in place of the usual annual grant, front-loading the expense. $20,106m remained unrecognised at May 2026 over roughly three years.

Where the cash went, FY2018–FY2025

From the cash flow statements in the FY2020, FY2023 and FY2025 10-Ks. Debt principal went from $17,609m at FY2018 to $67,120m at FY2025; 64.5% of it matures after FY2030. No 10-K in the archive states a leverage target, a covenant ratio or a deleveraging commitment.

The vendor financing, which is now the second balance sheet

Two arrangements that appear in no annual report now sit in the quarterly filings, and both lend Broadcom’s credit to its own customers. The first is the Backstop: Broadcom arranged for a financial partner to take on agreements to purchase AI racks and the related customer leases, and guaranteed that customer’s lease obligations over five-year terms. The Q3 filing discloses the mechanism for the first time — on a default, Broadcom owes the difference between 85% of the amounts outstanding and the resale value of racks recovered — with maximum exposure unchanged at about $29bn and nothing yet paid. Management called the structure the “XPV platform” on the Q3 call, with a first tranche of $35bn closed in June for one gigawatt for Anthropic.

The second is new this quarter and larger. Under the same lease agreements the customer “may, under certain circumstances and if needed, issue to us convertible promissory notes up to an aggregate principal amount of $42 billion,” usable solely to meet its obligations to Broadcom. None have been issued. The structure is circular by construction: a customer that cannot pay Broadcom may instead hand Broadcom paper convertible into its own equity, and Broadcom books the revenue either way. Together the two facilities put roughly $71bn of customer credit behind revenue that runs through the income statement at full value.

6. Cycle position, constraints and what to monitor

Broadcom is at the top of its own range on every measure that exists. Q3 FY2026 GAAP gross margin of 69.1% and operating margin of 54.0% compare with prior peaks of 68.9% and 45.2% in FY2023 and troughs of 55.2% and 15.2% in FY2019. Revenue in that single quarter, $29.6bn, exceeded the whole of FY2019. A record margin at a peak and a record margin at a trough are opposite facts, and this is the first kind.

The industry is genuinely cyclical and the record is specific: global semiconductor sales fell 32% in 2001, roughly 11.5% peak-to-trough across 2008–09, 12.0% in 2019 and 8.2% in 2022–23. Each resolved within 12 to 24 months, and each happened. Against that, the current expansion is larger than any prior contraction — sales reached $403.3bn in Q2 2026, up 35.1% sequentially, and WSTS forecasts 2026 industry revenue of $1.51 trillion, roughly double 2025.

Broadcom’s demand curve is four companies’ capital budgets. In the June 2026 quarter Amazon spent $54.2bn, Alphabet $44.9bn, Microsoft $35.8bn and Meta $31.1bn on property and equipment — about $166bn in one quarter, with Meta guiding FY2026 to $130–145bn. That spending is discretionary in a way a handset cycle is not, and it can be deferred by a board in a single meeting. The downside case is not a demand collapse but a deferral: hyperscalers slowing deployment by two or three quarters to digest what they have installed, landing on shipment volumes against a supply chain Broadcom has already committed to.

What is committed, and what that commitment buys

Purchase commitments — obligations to buy goods, primarily inventory, which the filing defines so as to exclude anything cancellable without penalty — went from $132m at November 2025 and $54m at February 2026 to $128,110m at May 2026. In the quarter to August 2026 they fell slightly, to $126,821m, with $52,674m falling in FY2027 and $72,952m in FY2028. The FY2027 tranche declined by $2,540m even as Broadcom told the market on 2 September that it expects to roughly double AI revenue to about $115bn in FY2027. A company that had to buy more capacity for next year would have added to this line. It did not.

That is the best available evidence that the order book was already complete, and it cuts against reading the commitment as an unreciprocated obligation. Across the industry in 2026, binding purchase commitments are how allocation is actually secured: Micron has take-or-pay agreements with roughly $18bn of customer cash deposits behind them, NVIDIA’s supply and capacity commitments went from $119bn to $279bn in a single quarter, AMD carries $29–30bn, and Marvell is prepaying about $1bn. Broadcom’s $126.8bn is not money pledged without a counterparty — it is the reservation itself.

The disclosure moved the same way. The Q2 filing said Broadcom does “not generally have long-term capacity commitments” with its contract manufacturers. That sentence is absent from the Q3 filing, replaced by “We may enter into long-term capacity commitments with our CMs”, and the phrase “capacity constraints” was dropped as well. This was established by controlled phrase-testing against the EDGAR full-text index rather than read in continuous text, so confirm it against the document before relying on it. On its face Broadcom has stopped describing itself as holding no capacity rights — though “may” is not “has”.

Which constraint actually binds is worth stating plainly, because management does not put it where an investor might expect. Asked at a conference on 8 September, Hock Tan placed it downstream: “The main constraint is not chip output. The bigger issue is whether data centers can be built fast enough” — power access, permitting, transformers and turbines, with roughly two years needed to prepare sites for 2028. If that is right, the risk in these numbers is not that Broadcom cannot obtain wafers. It is that Broadcom has bought inventory it cannot cancel without penalty, against deployments its customers may not be able to energise on schedule. Inventory has already risen from $2,270m at FY2025 year-end to $4,523m at August 2026, and the filing offers no explanation of the build anywhere in its pages.

The order book tells a subtler version of the same story. Committed remaining performance obligations went $27.5bn in August 2025, $33.3bn in November, $45.0bn in February 2026, $164.6bn at May 2026 on a single long-term custom accelerator contract, and $179.2bn at August 2026. But the share expected within twelve months fell from about 30% to about 25% — so the near-term slice shrank from roughly $49bn to roughly $45bn even as the total grew. The backlog is lengthening, not converting faster. The company cautions these “are not indicative of revenue for future periods.” Concentration of this magnitude in one agreement remains the central fact about Broadcom’s forward visibility, and it is simultaneously the strongest and the most fragile thing in the filings.

The left column survives a downturn; the right column is what a downturn removes.

What to monitor, and where it is published

7. Risks that cyclicality does not capture

Revenue and receivables depend on a handful of counterparties whose own funding is uncertain. One distributor accounted for 42% of net revenue in Q2 FY2026, against 32% for FY2025 and 21% for FY2023, and one customer represented 44% of net accounts receivable at November 2025 against 18% a year earlier. Of the four named XPU customers, two — OpenAI and Anthropic — fund purchases from raised capital rather than operating profit. Broadcom’s own filing states that AI customers “may have constrained resources or capital and may be unable to pay for their required AI infrastructure.”

The financing structures convert customer credit risk into Broadcom’s risk while leaving the revenue on Broadcom’s income statement. On a default the Backstop obliges Broadcom to pay the gap between 85% of the outstanding lease balance and whatever the recovered racks fetch — so the loss is a function of used-accelerator resale values in exactly the conditions that caused the default. Broadcom would be selling depreciating, customer-specific silicon into a market where demand had just failed. The exposure grows as racks deploy.

The FY2028 commitment is far short of what the FY2028 target would consume. Broadcom has $72,952m committed for FY2028 against a stated ambition of roughly $230bn of AI revenue that year. Scaling FY2026 product cost of revenue for that volume implies a materially larger requirement, so either the commitment line rises sharply over coming quarters or the FY2028 number is not yet supplied. That inference uses assumed margins — a flag, not a finding — but the commitment table is where it gets settled, quarter by quarter.

The rack and leasing model degrades the economics that made the business worth owning. It passes through low-margin content, consumes working capital, introduces credit exposure, and by the company’s own statement lowers gross margin and may hurt free cash flow. A single foundry relationship carries the entire semiconductor franchise — approximately 95% of wafers from TSMC, with an acknowledgement that TSMC “could choose or be required to materially prioritize capacity for other customers.” Whether Broadcom now holds contractual capacity rights is unresolved — the Q3 filing says only that it “may” enter long-term capacity commitments. And the balance sheet has no tangible cushion: goodwill and intangibles of $130,074m are 76.0% of total assets against $81,292m of equity, a tangible book value of roughly negative $48.8bn, with no goodwill ever impaired.

Management’s pay is levered to the one number an investor most needs to evaluate independently. The September 2025 CEO award pays nothing below $60bn of AI revenue and 300% above $120bn, measured on the best four consecutive quarters in FY2028–FY2030, using a company-defined “AI Revenue” figure disclosed on earnings calls and audited nowhere. The stated FY2028 target of approximately $230bn sits well above the maximum. The incentive to maximise disclosed AI revenue, including through rack sales that dilute margin, is direct.

What the sources could not answer

Each is a gap in the evidence rather than a judgement, and each would need resolving before forming a thesis: the split of AI revenue between XPUs and networking, never disclosed, which determines whether the growth is low-margin pass-through or high-margin silicon; revenue by end market, where Broadcom names five semiconductor end markets and discloses revenue against none of them in any year; how far through the VMware perpetual-to-subscription conversion the installed base is, last quantified over a year ago; whether Broadcom has actually entered long-term capacity commitments with its foundries or merely reserved the right to; whether Broadcom or its customers buy the high-bandwidth memory in an XPU, which decides whether the tightest link in the chain lands on its books or theirs; the penalty for cancelling the $126.8bn of purchase commitments; the commercial terms of the custom accelerator contracts — who owns the IP, who bears a cancellation, whether pricing is fixed across the platform life, none of it disclosed by any party in the industry; the identity of the distributor representing 42% of revenue; and a verified like-for-like VMware price series, where alleged increases run from tenfold to 1,200% in filed complaints and no primary source establishes the actual magnitude.

8. Investor takeaways