Business Overview

Broadcom Inc. (NASDAQ: AVGO)

7 September 2026

Evidence base

Broadcom SEC filings FY2018–FY2025 (10-K), 10-Q through Q2 FY2026, proxy statements through 2026 and selected 8-Ks, read from the analyst’s local archive; the Q3 FY2026 earnings release and call of 2 September 2026; and independent industry sources — competitor filings, TSMC disclosures, SIA/WSTS, the European Commission, the Japan Fair Trade Commission and BIS.

This is an investor-grade description of the business. It is not a valuation and not a recommendation.

1. Executive Snapshot

2. What the Company Does

Two very different customers have the same underlying problem: they need computing infrastructure that a general-purpose product does not deliver economically at their scale. A hyperscaler running one enormous, well-understood workload — training and serving its own models — can buy merchant GPUs, or it can commission silicon shaped precisely 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 computing capacity, and it needs that software to keep working with every storage array, network card and application it already owns. Broadcom sells to both.

The unit, and what it earns

The unit of economics in the semiconductor business is one custom AI accelerator — an XPU — together with the Ethernet switching, network-interface and optical content that ships beside it. Trace one 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 all wafers Broadcom’s contract manufacturers produced in the first half of FY2026 (FY2026 Q2 10-Q); 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 (FY2026 Q2 10-Q) against 32% for all of FY2025 (FY2025 10-K).

What that unit earns is disclosed for the first time in the FY2025 10-K, which adopted the expanded segment-expense standard. Semiconductor segment revenue of $36,858 million carried segment cost of revenue of $11,740 million — a 68.1% segment gross margin — and produced $21,232 million of segment operating income, a 57.6% operating margin. Because the segment expense lines exclude amortisation of acquired intangibles and stock compensation, these are the economics of the products themselves rather than of the reported income statement.

The software business has a second unit: one subscribed core-licence of VMware Cloud Foundation. On $27,029 million of FY2025 revenue it carried $1,902 million of segment cost of revenue — a 93.0% gross margin — and $20,765 million of segment operating income, a 76.8% operating margin (FY2025 10-K). In FY2025 the software segment generated 49% of the two segments’ combined operating income on 42% of revenue. Any statement about Broadcom’s margin is really a statement about the ratio between these two units.

The nature of the revenue

Roughly three-quarters of revenue is product sales and one quarter subscriptions and services (76% / 24% in Q2 FY2026). That understates how much of the software base is contractual, because $7,800 million of FY2025 revenue was upfront licence revenue recognised at signature on contracts the customer cannot terminate, and sits inside the products line after a FY2025 reclassification (FY2025 10-K). The consequence matters: software revenue in any given quarter depends on when large contracts are signed, not only on how much software is being consumed. The company says so directly — “the relative volume of customer contracts signed with the right to terminate causes variations in revenue recognized in each period” (FY2026 Q2 10-Q). About 64% of contract liabilities at May 2026 related to contracts terminable for convenience.

What it has stopped selling, and what it has started

Broadcom sells the parts of what it buys that do not clear its margin threshold. Since FY2018 it has divested Brocade’s IP networking business for $800 million, CA’s Veracode unit for $950 million, Symantec’s Cyber Security Services business, and VMware’s end-user computing division to KKR for $3.5 billion, each described in the filings as “not aligned with our strategic objectives”. The pattern is consistent: keep the entrenched, high-margin, low-capital core; sell the growth-capital-hungry adjacency.

The move now running the other way deserves attention because it cuts against that discipline. Broadcom has begun selling and leasing whole AI racks and systems built around its XPUs, not only the chips. Its own risk disclosure states the economics plainly: this “will likely increase our operating margin but compress or lower future gross margin” (FY2025 10-K). Racks carry memory, substrates and assembly that Broadcom passes through at a much lower mark-up than silicon. The first business in twenty years that Broadcom has entered because a customer asked rather than because the margin was attractive is worth watching for that reason alone.

3. Industry, Competitive Position and Moat

The AI data-centre industry sells one thing: usable compute delivered inside a power envelope. Everything else — design tools, silicon IP, wafers, memory, packaging, networking, sheet metal — is an input to that. The profit is distributed along the chain with unusual sharpness, and the most recent reported quarter from each layer shows the shape of it.

Each figure from that company’s own quarterly release. The distribution is bimodal: design, IP, foundry and memory earn 58–85% gross margins; physically assembling the result earns 6–12%.

Merchant silicon versus custom silicon

The two ways to sell AI compute have genuinely different economics, and the difference explains most of Broadcom’s position. NVIDIA sells a merchant part it designed, owns and prices: $89.0 billion 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: $16.7 billion of AI semiconductor revenue in Q3 FY2026, up 221% year on year and 54% sequentially. The customer funds the design, absorbs the memory and substrate bill, and takes the architectural risk. Broadcom takes less margin per unit than NVIDIA but also carries less inventory risk and no architectural bet — and it keeps the socket for the three to five years the platform lives.

That the model works commercially is now visible outside Broadcom’s own accounts. Amazon disclosed a custom-chip annual revenue run-rate “now over $20 billion” with “over $225 billion in revenue commitments for Trainium” (Q1 2026 earnings commentary), and Google has shipped seven generations of TPU. The competitive question is not whether custom silicon is real but who supplies it. Marvell is the only comparably scaled merchant alternative, at $2.171 billion of data-centre revenue and a 58.9% non-GAAP gross margin in its July 2026 quarter — roughly an eighth of Broadcom’s AI run-rate. Below that sit design-service houses such as Alchip, whose roughly 34.8% gross margin marks the commodity end of the same activity.

What would actually be hard to reproduce

Three things bind, and they are not the ones a filing usually emphasises. The first is the serialiser/deserialiser and optical IP that moves data between accelerators — the part of an AI system that has become harder than the compute itself, evidenced by the fact that every disclosed acquisition in the field over the past year has been in interconnect rather than compute: Qualcomm bought Alphawave for $2.4 billion, NVIDIA invested $4 billion into Coherent and Lumentum for photonics capacity rights, and Marvell, Credo, Semtech and Infineon all made photonics tuck-ins.

The second is access to advanced packaging, which is the industry’s binding physical constraint and 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” and that the shortage runs through 2029–2030; TSMC raised FY2026 capital expenditure to $60–64 billion in response. Broadcom sources approximately 95% of its wafers from TSMC and states that it does “not generally have long-term capacity commitments” with its contract manufacturers (FY2026 Q2 10-Q). Its advantage here is priority born of volume, not contract.

The third is the design win itself. Once a hyperscaler has committed a rack architecture, a compiler stack and a year of software work to a specific accelerator, switching suppliers means redoing all of it. This is why Broadcom can name six XPU customers — Google, Anthropic, OpenAI and Meta among them (Q3 FY2026 call) — and treat that number as the business rather than as a concentration problem. It is both.

Networking, where the position is more contested

Broadcom’s Ethernet switching silicon is the merchant standard, and the industry is standardising around Ethernet for AI fabrics through the Ultra Ethernet Consortium, whose specification reached v1.0.3 in July 2026 with AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft and Oracle among its steering members. But the technical lead is narrower than the AI numbers suggest. Cisco announced its Silicon One G300 at 102.4 Tb/s in February 2026 — the same throughput as Broadcom’s Tomahawk 6 — and NVIDIA reported $14.8 billion of data-centre networking revenue in its April 2026 quarter, up 199%, selling Spectrum and InfiniBand directly against Broadcom. No independent source publishes Ethernet switching-silicon market share, so the size of Broadcom’s lead here is unknown rather than established.

The software position, tested against outside evidence

Broadcom’s filings assert that infrastructure software is mission-critical and sticky. Independent evidence supports the mechanism and is more specific than the company is. When the European Commission cleared the VMware acquisition in July 2023 (Case M.10806) it required Broadcom to commit to open access to VMware’s APIs and certification programme and to grant an open-source driver — remedies aimed squarely at the I/O Vendor Partner certification programme, because hardware makers depend on it to have their adapters work. A competition authority mandating access is the strongest available proof that the certification matrix, not the hypervisor, is the lock.

Outside evidence also supports the durability claim so far and contradicts the loudest counter-narrative. Broadcom’s infrastructure software annual recurring revenue grew 15% year on year in Q3 FY2026 and revenue reached $8.8 billion, up 29%. Nutanix, the most direct beneficiary of any customer exodus, reported FY2026 revenue of $2.85 billion — about a twelfth of Broadcom’s software run-rate — growing 12% with net dollar retention of 106% and FY2027 guidance implying deceleration to 12–13%. 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 weaker is on price. CISPE, the European cloud-provider association, alleges cumulative cost increases “exceeding 1,000 percent” in its March 2026 Article 102 complaint and is separately seeking annulment of the merger clearance before the General Court (Case T-503/25). AT&T’s 2024 complaint documents the mechanics — 8,600 servers, over 75,000 virtual machines, 55,600 converted perpetual licences under one agreement — and states a transition off VMware “could take years”. But no primary source establishes a verified like-for-like price series, so the magnitude of the repricing is unknown. Both proceedings are pending.

What wins here

The industry rewards whoever owns a piece of the chain that is scarce, specified into a customer’s architecture, and cheap to reproduce once designed — and punishes whoever converts that position into volume that must be physically assembled. Broadcom is the first kind of company almost everywhere in its portfolio. The rack and leasing business is the one place it is deliberately becoming the second kind.

4. Growth Engine

Broadcom’s reported growth over the last three years has three quite different explanations, and conflating them produces the wrong business.

VMware contribution from the FY2024 10-K acquisition note; 9M FY2026 is Q1 and Q2 as filed plus the Q3 FY2026 release; the implied full 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.

FY2024’s 44% is almost entirely purchased: strip VMware’s $12,384 million and the underlying business grew 9.4% in a year that also had an extra week. FY2025 and FY2026 are a different matter. Broadcom has made no acquisition since a $600 million 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.

What is actually driving it, ranked

• Custom AI accelerator volume — structural, and overwhelmingly the largest driver. AI semiconductor revenue reached $16.7 billion in Q3 FY2026, up 221% year on year, with Q4 guided to $21.7 billion. Six customers deploying committed gigawatts of their own architectures is a demand source that does not depend on a broad market recovering. Management has stated FY2026 AI revenue of roughly $58 billion, and targets of approximately $115 billion in FY2027 and $230 billion in FY2028 (Q3 FY2026 call) — these are the company’s own forward statements, not established results.
• Attached networking and optical content — structural. Every accelerator deployment pulls Ethernet switching, network-interface controllers and optics. Broadcom does not disclose the split between XPU and networking within its AI revenue, so the relative contribution is unknown; NVIDIA’s $14.8 billion networking quarter indicates the attach economics are large.
• Repricing the VMware installed base onto subscription — management-driven, and finite. Software revenue grew 29% to $8.8 billion in Q3 FY2026 with ARR up 15%. This is a conversion, not a market: it ends when the base is converted, and management has not disclosed how much remains. The gap between 29% revenue growth and 15% ARR growth is the signature of upfront licence recognition rather than underlying consumption.
• Mix shift toward AI within the semiconductor segment — structural, but hollowing the base. Non-AI semiconductor revenue was $4.2 billion in Q3 FY2026, up 5% year on year and flat sequentially, with broadband and server storage up and wireless down (Q3 FY2026 call). At that run-rate the entire non-AI franchise is smaller than Broadcom’s whole semiconductor segment was in FY2023, when it earned $28,182 million. Growth is not broadening; it is concentrating.
• The extra week and the acquisition base effect — temporary, and now behind. FY2024’s 53rd week and the VMware lap flattered FY2024 and depressed FY2025 optics. Neither affects FY2026.

5. Margin, Cash and Capital Allocation

Broadcom converts revenue to cash better than almost any company of its size, and the mechanism is simple: it owns designs rather than plant. Capital expenditure over the eight years FY2018–FY2025 totalled $4,020 million against $126,722 million of operating cash flow — 3.2%. It fabricates almost nothing itself; the capital intensity of the industry sits on TSMC’s balance sheet, which is why TSMC is spending $60–64 billion this year while Broadcom spends around $2 billion. 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; and share counts before July 2024 are restated for a ten-for-one split. 9M FY2026 operating cash flow and share count are derived from reported quarters plus the Q3 release; total debt at Q3 FY2026 is carrying value, not principal.

What actually moves the margin

Gross margin is a mix question, not a cost question. In FY2025 the semiconductor segment carried a 68.1% gross margin and the software segment 93.0%, so every point of revenue mix that shifts toward semiconductors mechanically lowers the blended figure. That is exactly 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” (FY2026 Q2 10-Q). 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 rack business will push the same way harder, since the company expects it to “compress or lower future gross margin” while raising operating margin.

The largest non-cash cost is stock compensation, at $7,568 million in FY2025 — 11.8% of revenue, up from 6.1% in FY2023. It rose because Broadcom granted two-year equity awards in Q2 FY2025 in place of the usual annual grant, which front-loads expense; $20,106 million remained unrecognised at May 2026 over roughly three years. This is a real cost of the business that non-GAAP margin excludes and share count eventually reflects.

Where the cash went

Over FY2018–FY2025 Broadcom generated $126,722 million of operating cash flow and raised $124,054 million of gross debt against $88,932 million repaid. It deployed the money in a consistent order:

• Acquisitions — $57,990 million of cash net of cash acquired, against roughly $122 billion of gross announced consideration across Brocade, CA, Symantec’s enterprise business, VMware and a Seagate asset purchase, of which $53.4 billion was Broadcom stock and about $10.5 billion was debt assumed.
• Dividends — $54,612 million, with the per-share rate rising from $0.700 to $2.360 on a split-adjusted basis between FY2018 and FY2025, a 19% compound rate, and a further 10% increase to $0.65 a quarter in FY2026.
• Share repurchases — $50,627 million, of which $35,143 million under buyback programmes and $15,484 million as shares withheld to settle employee tax on vesting. Despite this, split-adjusted diluted shares rose from about 4,310 million in FY2018 to 4,853 million in FY2025, because 544 million shares were issued for VMware and equity awards keep vesting.
• Capital expenditure — $4,020 million, smaller than a single quarter of dividends. Management has said capital expenditure will be higher in FY2026, and Q4 is guided to $1.4 billion against $532 million in Q3 — a real step-up tied to the rack business.

The ranking says what the filings do not: Broadcom is an acquirer that returns cash between deals, financed substantially by debt. Debt principal went from $17,609 million at FY2018 to $67,120 million at FY2025 and $59.4 billion at Q3 FY2026, and 64.5% of the FY2025 principal matures after FY2030. Notably, no 10-K in the archive states a leverage target, a covenant ratio or a deleveraging commitment — the language is permissive throughout, and repayment has followed cash availability rather than a stated plan.

What post-dates the reported figures

Two arrangements struck after the last audited year change the risk profile and are not in any annual report. On 8 June 2026 Broadcom arranged for an investor partner to take on agreements to purchase AI racks and the related customer leases, entering a backstop of the customer’s lease obligations over five-year terms with a maximum exposure of $29 billion (FY2026 Q2 10-Q subsequent events). Management described the same structure on the Q3 call as the “XPV platform”, with a first tranche of $35 billion closed in June for one gigawatt for Anthropic. The economic substance is that Broadcom is standing behind a customer’s multi-year payment obligation in order to book the revenue now. There is no announced but unclosed acquisition.

6. Cyclicality, 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, $29.6 billion, exceeded the whole of FY2019. A record margin at a cyclical peak and a record margin at a trough are opposite facts, and this is the first kind.

The industry it sits in is genuinely cyclical, and the historical 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 (SIA). Every one of those resolved within 12 to 24 months, but every one also happened. Against that, the current expansion is larger than any prior contraction: sales reached $403.3 billion in Q2 2026, up 35.1% sequentially, and WSTS forecasts 2026 industry revenue of $1.51 trillion, roughly double 2025. A market that doubles in a year is not one where the base rate of past downturns can simply be assumed to apply, in either direction.

The demand behind it, and what it depends on

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

What is committed and what is not

The commitment asymmetry is the sharpest thing in the recent filings. Purchase commitments — binding obligations to buy goods, primarily inventory — went from $132 million at November 2025 and $54 million at February 2026 to $128,110 million at May 2026, with $55,214 million falling in FY2027 and $72,870 million in FY2028. Against that, Broadcom states it does “not generally have long-term capacity commitments” with its contract manufacturers and buys substantially all manufacturing on purchase orders. It has bound itself to buy; it has not bound anyone to supply. Inventory rose from $2,270 million at FY2025 year-end to $4,328 million at May 2026, with finished goods up 172%, “primarily to support higher expected shipments for custom AI accelerators”.

The other side is the order book. Committed remaining performance obligations went $27.5 billion (August 2025), $33.3 billion (November 2025), $45.0 billion (February 2026), then $164.6 billion at May 2026 — a $120 billion jump in a single quarter driven by one long-term custom accelerator contract, with about 30% expected in revenue within twelve months. The company cautions that these “are not indicative of revenue for future periods”. Concentration of that magnitude in a single agreement is the central fact about Broadcom’s forward visibility, and it is both the strongest and the most fragile thing on the page.

Durable versus borrowed

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

What to monitor, and where it is published

• Hyperscaler capital expenditure — quarterly releases and 10-Qs of Alphabet, Amazon, Microsoft and Meta. This is the demand curve; a guided reduction would show here first.
• TSMC monthly revenue and quarterly capital expenditure and packaging commentary — tsmc.com investor relations, monthly. The binding physical constraint.
• Broadcom’s remaining performance obligations and purchase commitments — each 10-Q. Whether the $164.6 billion converts, and whether the $128.1 billion of purchase commitments moves with it.
• Non-AI semiconductor revenue — disclosed only on the quarterly call. A recovering base would broaden the story; continued flatness confirms the concentration.
• Infrastructure software ARR against reported revenue — quarterly call. The gap between them measures how much software growth is signature timing rather than consumption.
• Free cash flow as a percentage of revenue — quarterly release. The rack and leasing model is the thing most likely to break the relationship between reported profit and cash.
• Global semiconductor sales — SIA monthly and WSTS forecasts, for the industry-wide turn.

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality is covered in Section 6. What follows is what cyclicality does not capture, ordered by how much each compounds the others.

• Revenue and receivables now 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. Broadcom names six XPU customers. Two of the four identified — OpenAI and Anthropic — fund purchases from capital raised rather than from 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. The $29 billion maximum backstop of a customer’s five-year lease obligations means a customer failure would not simply cost a future order; it would create a direct payment obligation against revenue already recognised. This exposure grows as racks deploy.
• Broadcom has committed $128.1 billion to purchase inventory without securing the capacity to build it or a matching cancellation right. If deployment slows, the obligation does not. The filings do not disclose whether these commitments are cancellable or on what terms, which is the single most consequential undisclosed fact in the current accounts.
• 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. It is being adopted because customers are asking, which is a different reason than Broadcom has historically used to enter a business.
• A single foundry relationship carries the entire semiconductor franchise. Approximately 95% of wafers come from TSMC, with no long-term capacity commitment and an acknowledgement that TSMC “could choose or be required to materially prioritize capacity for other customers”. Concentration of manufacturing in Taiwan is a second-order version of the same exposure.
• The balance sheet has no tangible cushion. Goodwill and intangibles of $130,074 million are 76.0% of total assets, against $81,292 million of equity — tangible book value of roughly negative $48.8 billion. No goodwill has ever been impaired. An impairment would not affect cash but would test the debt structure’s optics at exactly the moment cash flow was under pressure.
• Management’s pay is levered to the one number an investor most needs to evaluate independently. The September 2025 CEO award pays nothing below $60 billion of AI revenue and 300% above $120 billion, 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 FY2028 target of approximately $230 billion 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 of these 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. It determines whether the growth is a low-margin pass-through or high-margin silicon.
• Revenue by end market. 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. The last quantification was over a year ago; no renewal-pricing or churn data is published.
• Whether the $128.1 billion of purchase commitments is cancellable, and on what terms.
• The commercial terms of the custom accelerator contracts — who owns the IP, who bears a cancellation, and whether pricing is fixed across the platform life. Not disclosed by any party in the industry.
• The identity of the distributor representing 42% of revenue, and the concentration of end customers behind it.
• A verified like-for-like VMware price series. Alleged increases run from tenfold to 1,200% in filed complaints; no primary source establishes the actual magnitude.

8. Investor Takeaways

• This is a capital allocator that happens to own semiconductors and software, not a chip company — it has spent $58 billion of cash on acquisitions and $4 billion on plant since FY2018, and it sells whatever it buys that does not clear its margin threshold.
• The engine is the sole-source socket: a design the customer paid to develop, produced with almost no Broadcom capital, at 68% segment gross margin in silicon and 93% in software, held for the three to five years the customer’s platform lives.
• The growth lever is the number of gigawatts of custom AI compute six customers deploy, and for the first time in a decade the growth is organic rather than acquired.
• What could break the story is not a demand collapse but the shape of the commitment: $128 billion of purchase obligations and a $29 billion lease backstop against six customers, two of which are funded by capital markets rather than profits.
• Watch free cash flow as a share of revenue and the non-AI base. The first tells you whether the rack model is eroding the economics; the second tells you whether anything is growing that is not four capital budgets.