Broadcom

IThe business

Broadcom sells chips and infrastructure software. It designs the chips and makes almost none of them: TSMC made about 95% of the wafers that Broadcom’s contract manufacturers produced in FY2025.² The software runs inside most of the Fortune 500.² Fiscal years end on the Sunday nearest 31 October: FY2025 ended 2 November 2025. TTM on this page means the four quarters to 2 August 2026, because the fourth quarter reports in December.²³

In the twelve months to 2 August 2026 revenue was $89.1B, up 49%. Semiconductor solutions brought in $59.4B at a 60.7% segment margin and infrastructure software $29.7B at 79.9% C (calculated). Segment margins leave out stock pay, the amortization of acquired intangibles, restructuring and acquisition costs, which together cost $17.0B.²³¹⁴

Revenue grew from $13.2B in FY2016 to $63.9B in FY2025, 19.1% a year. Five acquisitions, from Broadcom Corporation in 2016 to VMware in 2023, cost $158B C (calculated) and are in those numbers; the table under Management lists them.^1215161713

What it sells

• Custom AI accelerators, called XPUs: chips designed with one customer around its own models. Google’s TPU is the oldest program. Management names Anthropic, OpenAI and Meta as large XPU customers, and Apple signed for custom silicon through 2031. Broadcom, Google and Anthropic expanded their collaboration so that Anthropic, from 2027, accesses about 3.5 gigawatts of TPU-based capacity through Broadcom; the 8-K says Anthropic’s use of it is ‘dependent on Anthropic’s continued commercial success’.⁶²²⁵
• AI networking: Ethernet switches, network cards and optics, sold to OEMs and system integrators and inside Broadcom’s own racks.² Dell’Oro puts Ethernet at about two-thirds of AI back-end switch sales in the first quarter of 2026.²³
• Other semiconductors: wireless, broadband, storage and industrial chips. They brought in $4.1B in the quarter to August, a fifth of segment revenue C (calculated).³⁴
• Infrastructure software: VMware Cloud Foundation, mainframe software, cybersecurity and Fibre Channel storage networking. Revenue was $8.75B in the quarter, including $3.47B of upfront license revenue.³

The AI number

Broadcom’s AI revenue is not a line in any filing. The 10-K lists AI semiconductor solutions as custom accelerators, Ethernet switching and routing silicon, network cards, physical layer devices and optical components, plus racks and systems built on its accelerators, and the dollar figure appears only in the earnings releases.²⁴ On the call, management said accelerators were 73% of Q3 AI revenue, a claim from a third-party transcript; the filings do not split accelerators, networking and racks.⁵

AI revenue was 47.6% of the TTM, $42.4B of $89.1B, and 56.4% of Q3 revenue. It was 80% of Q3 semiconductor segment revenue C (calculated).⁴³

Who pays

One semiconductor customer, a distributor, was 50% of Q3 revenue and 32% of FY2025 revenue. Broadcom does not name it.³² By end customer, the better measure, Broadcom believes its top five were about 55% of Q3 revenue and 50% over nine months, against about 40% a year earlier.³ On the call Tan counted six XPU customers, four of them ‘just simply going to be huge’.⁵

Customers now want more than chips. The 10-Q says top customers have asked Broadcom to lease its accelerators, to purchase full racks and lease them on, and to accept financing such as the AI XPV platform or deferred payment.³ The 10-K says what that does to the margin:

will likely increase our operating margin but compress or lower future gross margin

Gross margin was 75% in Q3, 210 basis points below Q2, which the CFO attributed to AI’s larger share of revenue, and she guided about 73% for Q4.⁴⁵

Build-out or annuity

The case for an annuity is contracts. Remaining performance obligations were $179.2B at 2 August 2026, from $33.3B nine months earlier, and the 10-Q says they include a long-term custom AI accelerator contract signed in the quarter to 3 May.³² Apple’s agreements and Google’s supply assurance agreement run to 2031.⁶²² Custom chips are redesigned each generation, so revenue recurs for as long as customers keep building.

The case for a build-out is who pays. The 10-Q says the XPV platform bridges the gap between the leading frontier AI labs’ cash flows and their upfront investment, and does not name them.³ On the call, management said the platform would enable more than 20 gigawatts for OpenAI and Anthropic, and Tan said the other four of the six XPU customers can fund themselves.⁵ The 10-Q also says the AI upturn ‘may not be sustainable’ and names a perceived AI spending bubble among the causes.³

I lean toward annuity for the software, and for AI revenue through FY2028, where supply is bought and contracts are signed. After that I lean toward a build-out, which is where the flat-to-down year in my base case sits.

Where

By shipment or delivery location, the United States was $16.5B of FY2025 revenue, China including Hong Kong $11.2B (17%), Singapore $10.8B and Taiwan $6.5B. These are delivery points, not end customers.²

IIThe moat

I think Broadcom’s moat is switching costs, held up by scale in design and an installed software base. It is wide. Against rivals it is widening: the semiconductor segment margin rose from 48.5% to 61.3% as segment revenue tripled C (calculated).¹⁷³ Against customers it is narrowing, because five of them are 55% of revenue and each owns its chip’s architecture.³

An accelerator is designed with one customer around its models, its software and its racks. Moving to another supplier means redesigning the chip around the customer’s software, so a customer switches when it designs the next generation, not in the middle of one. The customer decides again at each generation. Counterpoint says Google dual-sources TPU v8 with MediaTek, Tan said Broadcom’s v8i is shipping ahead of MediaTek’s v8t, and the 10-Q lists customer-owned tooling as a scenario that could cut demand for Broadcom’s accelerators.²⁷⁵³

Customers run core systems on VMware and mainframe software, and Broadcom earns a 79.9% segment margin on it C (calculated).³¹⁴ I can test the software moat only at the edges. The one survey on churn was sponsored by a VMware support rival: of 269 users, nine in ten said licensing costs made them consider alternatives, and 48% had no plan to move assets to VMware Cloud Foundation.²⁸ It measures intent, not exits, and Broadcom does not disclose renewals.

The last soft year for chips was FY2019. Semiconductor revenue fell 8.5% and segment profit fell 7.7%, and the margin held at 49% C (calculated).¹⁷ FY2020 revenue fell 1%. Both years hit a semiconductor business less than a third the size of today’s C (calculated). Today’s margins show scale and mix as much as moat: gross margin fell from 77% to 75% in the quarter and is guided lower.⁴⁵

Evidence it is widening

Contracts. Obligations of $179.2B are 5.4 times the $33.3B of nine months earlier C (calculated), and Apple’s agreements and Google’s supply assurance agreement run to 2031.³²⁶²²

Operating leverage. The semiconductor segment margin rose from 48.5% in FY2018 to 61.3% in Q3 FY2026, with one dip, to 55.7% in FY2024 C (calculated). Segment revenue was 3.1 times larger over the same span C (calculated).¹⁷¹³²³

Supply. Management says it has secured the supply for its FY2027 and FY2028 outlook in leading-edge wafers, substrates and memory, is building a substrate fab in Singapore from FY2027, and is adding laser capacity. Purchase commitments are $126.8B, from $0.13B nine months ago.⁵³² In a market short of memory, lasers and substrates, contracted supply is itself a barrier.

Evidence it is narrowing

Concentration. The top five end customers went from about 40% to about 55% of revenue in a year. The 10-Q says the loss of any one of them could have a material adverse effect, and the five average about 11% of revenue each C (calculated).³

Second sources. Counterpoint expects Broadcom to hold about 60% of design-partner share of AI server compute ASICs in 2027 and Marvell 8%, a forecast from one research house.²⁷ NVIDIA bought $2.0B of Marvell preferred stock and the two agreed to connect Marvell’s XPUs to NVIDIA’s ecosystem.⁷

Credit. A supplier that backstops its customers’ leases competes on its balance sheet as well as on silicon. Broadcom has given a backstop of up to about $29B on one tranche, and on 27 October it filed a second, of up to about $43B.³⁴⁹

The moat is wide against competitors and narrow against customers, and five customers are 55% of revenue.

IIIThe competitive advantage

Broadcom’s advantage is scale and speed in custom AI silicon. On different definitions, in each company’s latest reported quarter, Broadcom’s AI revenue was 7.7 times Marvell’s whole data-center revenue, 2.5 times AMD’s Data Center segment and 19% of NVIDIA’s Data Center revenue C (calculated).⁴⁷⁹⁸

The margins show where the pricing power sits. Broadcom’s semiconductor segment earned 61.3% in the quarter, NVIDIA’s Compute & Networking segment 71.0% and AMD’s Data Center segment 31.3% C (calculated).³⁸⁹ Broadcom’s figure covers non-AI chips as well, AMD’s includes server CPUs, and all three leave out stock pay and the amortization of acquired intangibles.

The 10-K names no competitors. It says semiconductor rivals include integrated device makers, fabless companies and the in-house teams of large customers, and that some customers are entering Broadcom’s markets.² Marvell, MediaTek, NVIDIA and AMD are the ones with numbers.

Where Broadcom wins

Broadcom wins where a buyer has one large, stable workload and wants to own its cost. Management’s claim is that an XPU performs as well or better on a customer’s own workloads at less than half the cost of a GPU. The claim comes from the call, in a third-party transcript, and no outsider has tested it.⁵ NVIDIA’s segment earns 71.0%, a margin that buyers with a stable workload have a reason to design around.⁸

It also wins in networking, where Ethernet took about two-thirds of AI back-end switch sales in the first quarter of 2026, with InfiniBand rebounding, and Broadcom’s switch silicon sits inside other vendors’ systems.²³ And it wins on supply: it has contracted the wafers, substrates and memory for its own outlook, which a smaller design house cannot do.⁵³

Where it loses

It loses where the workload changes faster than a chip can be designed, and where a buyer wants one supplier for chips, network and software. Marvell’s agreement with NVIDIA points at that second case.⁷ It also loses a customer’s next generation when the customer builds the tooling itself, which the 10-Q names as a risk.³

The five customers that are 55% of revenue each have the scale to design the chip themselves, and Google already works with MediaTek on TPU v8.³²⁷⁵

IVManagement

Hock Tan has been chief executive since March 2006, first of Avago, which was renamed Broadcom after it bought Broadcom Corporation in 2016. He is 74. Before that he ran Integrated Circuit Systems, was vice president of finance at Commodore International, and worked at PepsiCo and General Motors. He joined Meta’s board in February 2024 and left it at Meta’s annual meeting on 27 May 2026, having told Meta in April that he would not stand again.⁶²⁶³ Management names Meta as an XPU customer. The proxy says Broadcom deals with entities where a director has relationships on arm’s-length terms; it does not link that statement to Meta and gives no amounts.²¹²¹⁰⁵

He beneficially owns 908,474 shares, under 1%, worth about $296M at the Q4 2026 close C (calculated); a year earlier the figure was 1,844,746.¹⁰²⁹ The Form 4 data I retrieved for Tan runs from November 2023 to May 2026. It shows about 1.54M shares sold from 14 December 2023 to 6 January 2026, about $356M at each day’s close E (estimate), plus 855,500 given away; sales before November 2023 are not counted.¹¹ None of the senior team has a written employment contract.²

The chairman sold more than Tan did. The listing of insider transactions from 25 March to 23 September 2026 shows Henry Samueli selling about 1.54M shares, about $559M at each day’s close E (estimate), while 16.2M of his shares stand pledged for loans. Mark Brazeal sold about 83,000 shares, about $32M.³⁰¹⁰

Pay

Tan’s reported FY2025 pay was $205.3M, of which $202.4M was the target value of the September 2025 award; the award’s maximum grant-date value is $607M. He had no annual cash bonus for FY2025, because the 2023 award covers five years of pay. The CEO pay ratio was 543 to 1.¹⁰

Two awards drive his incentives, and both are set far below where Broadcom is now.

The 2023 award’s top price hurdle is $112.50, about a third of the current price. The 2026 proxy shows the award at its maximum of 10 million shares, worth $3.3B at today’s price if he stays to 31 October 2027.¹⁰³² The 2025 award pays nothing at $60B of AI revenue over four quarters, the target at $90B and the maximum at $120B. When the board set it, it expected FY2025 AI revenue of about $20B.¹⁰³¹ The Q4 guide already runs at $87B a year, within 4% of the target, and the FY2028 outlook of $230B is almost twice the maximum hurdle C (calculated).⁴⁵

These awards tie his pay to the stock and to AI revenue. The 2023 award is shown at its maximum, and the 2025 award’s target is about where the Q4 guide already runs. His disclosed holding is $296M, against awards worth up to $3.9B at today’s price, so the awards shape his incentives more than his shares do C (calculated).¹⁰

The pay rewards AI revenue and a higher share price. The 8-K sets the 2025 award’s goal on AI revenue alone, with no margin, cash or return test, so the backstops that help win revenue are not charged to him.³¹ His $296M of shares is under a tenth of awards worth up to $3.9B C (calculated).¹⁰ Over a full cycle the deals have paid, with FY2025 software profit at 17.0% of the $121.9B they cost,¹³² but the last buybacks paid $338 a share, above my base value.³

Shareholders have pushed back. Votes for the pay proposal, as a share of votes for and against, were 61.7% in 2024 and 66.4% in 2026, against the 92% the proxy reports for 2025, and the compensation committee chair, Harry You, received 74.0% on the same basis in 2026 against 79.6% in 2024 C (calculated).³⁴¹⁰³⁵ After the September 2025 award the company contacted holders of 45% of shares and met holders of 38%. The proxy says those holders acknowledged the importance of Tan’s continued leadership, and that top holders showed overwhelming support for keeping him beyond FY2027.¹⁰

The board says it is engaged in CEO succession planning and names no successor. Tan’s award runs to the end of FY2030, when he would be 78 C (calculated).¹⁰

The bench

The CFO changed on 12 June 2026. Kirsten Spears, CFO since December 2020, retired and consults to March 2027. Amie Thuener, Alphabet’s chief accounting officer from 2018 to 2026, joined in May.³³ Charlie Kawwas runs semiconductors, 67% of revenue C (calculated), and has been at Broadcom since 2015.²³

The chairman, Henry Samueli, co-founded Broadcom Corporation. He owns 85.6M shares, and 16.2M of them are pledged to secure loans for a development project; the proxy puts the pledge at about 0.3% of shares outstanding.¹⁰ The lead independent director of eight years, Eddy Hartenstein, retired at the April 2026 meeting at 75, and the board shrank to eight.³⁶¹⁰

Capital allocation

Tan buys mature software and cuts its costs. The cash goes to dividends and debt. The record is strong, and the newest decision, financing customers’ racks, has no record yet.

The software deals clear my 10.5% discount rate: $121.9B of consideration became $20.8B of segment profit in FY2025, 17.0% before stock pay and amortization and about 12% after 17% tax and a revenue-share of stock pay C (calculated).²¹³ From FY2016 to FY2025 Broadcom generated $130.9B of free cash flow, paid $57.1B in dividends and distributions, $56.0B of it on common shares, and $35.1B in program buybacks, and spent $68.1B in cash on acquisitions C (calculated); stock paid for the rest of VMware.²¹³¹⁵ The buybacks did not shrink the share count. Diluted shares were 4.31B in FY2018 and 4.85B in FY2025, and without the 544M issued for VMware the count is flat, so the buybacks offset stock pay C (calculated).¹⁵²¹³

The dividend grew from $0.41 a share in FY2017 to $2.36 in FY2025, 24.5% a year C (calculated).²¹⁶ The buybacks were made at low prices until this year: Broadcom paid an average of $64 a share in FY2023 and $107 in FY2024, and $338 in the nine months to August 2026, above my base value.²³¹³

The newest habit is financing customers. In June 2026 Broadcom launched the XPV platform, gave a backstop of up to about $29B on the first tranche’s leases, and accepted that a customer may issue it up to $42B of convertible notes.³ I found no earlier customer financing of this kind in the filings, so there is no record to judge it by.

Broadcom returned $17.7B to shareholders in nine months and in the same period added a backstop of up to $29B for customers’ leases.³

VFinancials

Profit is growing faster than cash. Four things need watching: upfront software license revenue, contract assets, the tax rate and stock pay.

Gross margin was 68.9% in FY2023, 63.0% in FY2024, the first year with VMware, 67.8% in FY2025 and 68.8% in the TTM. Operating margin went from 45.2% to 26.1%, after $9.3B of acquisition amortization, then to 39.9% and 48.0%.¹³²³

Gross margin by line is not in the filings, and the CFO gave it on the call: about 67% for semiconductors and 94% for software in Q3, which reconcile to the reported 75% for the company. The split of the semiconductor margin between XPUs and networking is not disclosed, and the 10-K says rack sales will likely compress it. Each point of semiconductor gross margin in FY2028 is $1.9B of gross profit in my base.⁵⁴²

Earnings quality

Upfront license revenue. Broadcom recognized $7.8B of upfront license revenue in FY2025 and $7.2B in the first nine months of FY2026, $3.5B of it in the latest quarter, on contracts where customers cannot terminate.²³ It arrives when a contract is signed, not evenly, so software revenue growth of 29% in the quarter overstates the recurring rate. Management cites annual recurring revenue growth of 15%, a claim from the call.⁵ Contract assets, revenue recognized ahead of invoices, rose to $11.9B from $8.9B in nine months.³

Tax. The effective tax rate was −1.7% in FY2025 and 5.4% in the TTM C (calculated), helped by Singapore and Malaysia incentives worth about $2.7B in FY2025. The 10-K expects the global minimum tax to ‘materially increase’ the rate; the Malaysia tax holiday expires in FY2028 and the Singapore incentives through November 2030, and management guides a non-GAAP rate of about 16% for the fourth quarter. I keep the rate at 17% for ten years, and each point of tax rate costs about $3 a share of base value E (estimate).²³⁵

Cash. Operating cash flow was $40.7B in the TTM against net income of $38.3B. An $11.0B rise in working capital held it down, and stock pay of $8.5B is a cost that I charge.²³

Stock pay. The expense understates the cost. RSUs that vested in FY2025 were worth $21.7B at vesting prices against $7.6B of expense, and $19.9B in the nine months to August 2026 against $6.3B.²³ There are 165M unvested RSUs, worth $54B at today’s price against $18.2B of expense still to be recognized C (calculated).³ RSUs granted in the nine months were valued at $333.44 each at grant, against $142.47 on average for those outstanding. My cases charge $11.3B in FY2027 and $13.2B in FY2028, against $8.5B in the TTM; keeping stock pay at the TTM’s 9.5% of revenue would cost $23 a share E (estimate).³

The balance sheet

Debt principal was $61.1B at 2 August 2026 against cash of $24.0B, so net debt was $37.1B: 0.62 times non-GAAP operating income, and GAAP operating income covered interest 13.7 times C (calculated).³⁴ Principal peaked at $69.8B at the end of FY2024 after VMware.² In the nine months to August Broadcom paid $10.5B on debt obligations and borrowed $4.5B.³ Coupons run from 1.400% to 5.700%, and management puts the weighted average at 4% with 7.4 years to maturity. The maturity table shows $2.3B due in the rest of FY2026, nothing in FY2027 and $5.1B in FY2028.³⁵ The $7.5B revolver and the $4.0B commercial paper program are undrawn.³ Goodwill of $97.8B is 52% of assets.³

Off the balance sheet, Broadcom has stepped into its customers’ financing:

VIValuation

First, how this goes badly wrong

Before the numbers, I asked what would most likely ruin this, and worked back. Three things can. The labs may not be able to pay: two of the six custom-chip customers rely on Broadcom’s financing, and it has backstopped up to about $29B and filed a second of up to about $43B.³⁴⁹ Broadcom may have bought too much supply: it has $126.8B of purchase commitments, and my bear writes off $26.5B.³ A few buyers may pause: the top five are 55% of revenue, and the four largest cloud buyers spend 91% of their operating cash flow.³⁴⁵ The first two set my bear case.

An owner takes out less than the profit. Free cash flow was $39.4B, and after $8.5B of stock pay it is $30.9B, just over half of the $59.8B operating income the company reports C (calculated).⁴³ Standing still takes shares, not plants: RSUs that vested in nine months were worth $19.9B, against $6.3B of expense.³ The $30.9B is 1.9% of the $1.59T market value, or 2.3% on $37.3B with working capital steady and 17% tax C (calculated).³⁵⁰

Some of it I cannot know. OpenAI is private and no filing I used shows its cash flow, and the FY2028 outlook is management’s claim, not updated each quarter.³⁵ So I show a range, $95 to $431, not one number.

What the price implies

At $326.08 my base case returns 7.5% a year, below the 10.5% I require. The weighted value of my cases is $262 a share, and the skew at the price is −71% to my bear value and +32% to my bull value E (estimate). I compare each return with my own 10.5% first and the 5.00% ten-year Treasury yield second.⁵⁰³⁹ Broadcom fell 5.5% on 30 October in a sector move: the Nasdaq Composite fell 2.9% and the PHLX Semiconductor index 7.2%.⁵⁰⁵²

Start with earnings. At $326.08, Broadcom trades at 18.7 times Nasdaq’s mean for FY2027 and about 10.9 times the FY2028 earnings that management says will exceed $30. Neither figure states its basis, and management does not say whether the $30 is before stock pay.⁴⁰⁵ My base earnings are $15.08 a share for FY2027 and $20.32 for FY2028 after stock pay, and $22.56 for FY2028 before it, on 4.887B shares E (estimate). The outlook on time, at my margins, gives $28.82 before stock pay, so the claim needs more than my margins: on the company’s 4.94B shares, $30 needs an operating margin before stock pay of about 64% on $282B of revenue, below the 66% guided for Q4 E (estimate).⁵⁴

Broadcom’s 18.7 times FY2027 earnings sits between NVIDIA’s 13.0 times and the 33 to 36 times of AMD and Marvell, on one source and one day.⁴⁰⁴¹⁵¹ The multiple does not separate my value from the price. The cash-flow path does.

The reverse DCF starts from $37.3B: the TTM’s owner free cash flow of $30.9B, which is free cash flow less stock pay, plus the $11.0B rise in working capital, less $4.7B of tax to move the TTM’s 5.4% rate to the 17% my cases use. I start before working capital, which the cases charge year by year, and after tax at 17%, the rate the cases use. That start is a yield of 2.3% on the diluted market value, against 1.9% on the $30.9B. I solve for one growth rate for ten years. The price needs 18.8% a year, to $208B in year ten C (calculated). At my base-case margin of 41.2% that is about $505B of revenue, 5.7 times the TTM E (estimate). At the reported 5.4% tax rate the start is $41.9B and the price needs 17.2% a year. Owner free cash flow grew 18.2% a year from FY2019 to FY2025, and my base case grows 13.8% a year from the $37.3B start.²³¹⁶

If the outlook arrives on time, the price needs owner free cash flow to grow only 7.5% a year after FY2028, so the debate is what comes after it.

The discount rate

I use 10.5%. The capital asset pricing model uses the 5.00% ten-year Treasury yield in Q4 2026 and Damodaran’s implied equity risk premium of 4.20% on 1 October 2026. It gives 10.1% on Broadcom’s ten-year beta of 1.21, 11.2% on the five-year beta of 1.48 and 12.0% on the three-year beta of 1.66 C (calculated). The ten-year window includes years before VMware and before AI was close to half of revenue, so it understates the risk of the business I am valuing. The five- and three-year windows are mostly the AI run-up in one stock. I take 10.5%, a beta of 1.31, between the first two, and the base is $237 at 11.2% and $214 at 12.0% E (estimate). The betas use month-end price returns without dividends.^3943144 One point on the rate moves the base by 13% to 17%.

Three cases

I build each case from the segments down: AI, other semiconductor and software revenue, a segment margin for each, stock pay, a 17% tax rate on the profit after stock pay, capital spending of 2% to 3% of revenue against depreciation of 0.8%, and working capital at 20% of each added revenue dollar C (calculated).²³ The 17% is the rate Singapore applies without its incentives. Stock pay is a cost, as under GAAP. The amortization of acquired intangibles is non-cash and left out. The working capital build was 21% to 40% of added revenue in the filings, depending on the measure: $10.0B on the $46B rise in annualized quarterly revenue since November 2025, $11.0B on the $29.2B rise in TTM revenue, or $10.0B on the $25.2B rise in nine-month revenue. I use 20%, the low end, charge it only when revenue grows, and never release it when revenue falls C (calculated).²³

The base case takes most of the outlook and then assumes a normal cycle.

• AI revenue grows 10.5% a quarter for eight quarters from the Q4 guide of $21.7B, to $112B in FY2027, 97% of the outlook, and $167B in FY2028, 73% of it; FY2029 is the Q4 FY2028 run rate of $193B E (estimate). Revenue is $162B and $219B, against $165B and $282B on the outlook. The base’s FY2028 depends on the ramp of 2026 continuing at the same rate through the second half of 2027.
• A flat-to-down year in FY2030, AI revenue down 5%, then growth of 7% falling to 5% a year E (estimate).
• Other semiconductors grow 4% a year, and software 3% in FY2027, because FY2026 carries unusually large upfront license revenue, then 5% falling to 3.5% E (estimate).³
• The semiconductor margin falls from 60% to 54% by FY2033 and stays there, because racks and memory pass through at lower margin. The CFO attributed Q3’s 210 basis point drop in gross margin to AI’s larger share of revenue and guided about two points more for Q4, and the 10-K says rack sales will likely compress it. The FY2018 to FY2025 average of 53.9% is a check, not the reason E (estimate). The software margin is 78%, between FY2025’s 76.8% and the TTM’s 79.9%.⁵²³¹⁷
• Stock pay is 7.0% of revenue in FY2027 and 5.0% from FY2032; there are no losses on supply or the backstop.

The base grows at the same rate in both years. The outlook, read as steady growth within each year, needs 11.6% a quarter in FY2027 and 22.7% in FY2028. The doubling in FY2028 is the ramp I do not take.⁴⁵

Can the outlook be delivered

Every figure in the outlook comes from the call, in one third-party transcript, and not from a filing: the $115B, the $230B, the earnings above $30, the gigawatt schedule, the 73% accelerator share and the gross margins by segment. This section and the answer treat them as company claims.⁵ Four tests follow. Gigawatts and supply say the outlook is physically possible. The customers’ budgets say the self-funding customers can carry a quarter of it. The funding of the labs is the risk, and it sets my number.

Gigawatts. The outlook rests on a schedule. An analyst’s tally on the call puts about 10 gigawatts in 2027 and 20 in 2028, with about 15 of the 20 in 2028 at Anthropic and OpenAI. If revenue follows gigawatts, the other four customers carry a quarter of the 2028 outlook, $57.5B. Tan said the gigawatts can be added up and that not all of them will reach production in the two years.⁵ The tally is for calendar years and the outlook is for fiscal years to 31 October, and chips ship before the gigawatts are energized, so I read the test as a ceiling on what the schedule can carry.

At the $20B to $30B of content a gigawatt that Tan described, the $230B outlook needs 7.7 to 11.5 gigawatts in FY2028, 38% to 58% of the 20; my $167B needs 5.6 to 8.4, 28% to 42%. At the $11.5B an analyst got by dividing the outlook by the tally, the outlook needs all 20 and my $167B needs 14.5, 73% of them C (calculated).⁵ On either reading the schedule can carry my base.

Supply. Broadcom has bought supply for the outlook in FY2027 and for about my base case in FY2028: commitments of $52.7B for FY2027 against $49.4B of cost on the outlook, and $73.0B for FY2028 against $72.4B on my base E (estimate). The test depends on gross margin. I assume a company gross margin of 70% in FY2027 and 67% in FY2028: the semiconductor margin falls from about 67% in Q3 to 64% and 62% as racks and memory grow, with software at 94%. The Q4 guide for the company is 73%. At 73% the commitments cover the outlook’s FY2028 cost to within $3.2B; at 67% they cover four fifths of it.³⁵ So supply cannot separate my base from the outlook, and I take no write-down in the base because the supply is mostly wafers and memory that carry forward.

The customers’ budgets. Alphabet and Meta, the two public XPU customers Broadcom names, spent $129.7B on capital in the first half of 2026, $259B at an annual rate, against operating cash flow of $148.9B. The self-funding customers’ quarter of the FY2028 outlook, $57.5B, is 22% of that rate. The four largest cloud buyers together spent 91% of their operating cash flow, and the $230B outlook equals 39% of their spending at the same rate; my $167B equals 28% C (calculated). These are total capital spending, not AI-only, and Microsoft and Amazon are not named customers. Broadcom does not disclose revenue by customer, including Google’s share.⁴⁵⁵

Funding. On the call, management said the XPV platform would enable more than 20 gigawatts of compute for OpenAI and Anthropic by the end of 2028, and Tan said the other four of the six XPU customers can fund themselves.⁵ The 10-Q says the platform bridges the gap between the leading frontier AI labs’ cash flows and their upfront investment, and the 8-K says Anthropic’s use of its expanded capacity is dependent on its continued commercial success.³⁶ Three quarters of the 2028 gigawatts rest on those two customers, whose payment depends on lenders and on their own revenue. OpenAI is private, and no filing I used shows its revenue or cash flow. Anthropic is private too, but its S-1 of 26 October now shows both: $16.3B of revenue and $3.9B of operating cash flow in the first half of 2026. It also shows about $161.2B of rent under leases with Broadcom that neither side can cancel except on a default. Section VII comes back to it.⁵³

My $167B is a judgment, not a derivation. It is the four self-funding customers’ quarter of the outlook at full value, $57.5B, plus 64% of the $172.5B that the two labs carry in the other three quarters E (estimate). I would not give the labs’ plan the same weight as a hyperscaler’s own spending. The price already pays for $227B of FY2028 AI revenue, which on this split is 98% of the labs’ share.⁵⁵⁰

Bear, bull and stress

The bear case has the funding fail. AI revenue is $88B in FY2027 and $92B in FY2028, then falls 24% to $70B in FY2029, and total revenue falls 15%; the semiconductor margin drops to 46%. Broadcom writes off $26.5B of supply, the excess of FY2028 commitments over the cost of goods on $141B of revenue at a 67% gross margin, and pays $12B under the backstop, 35% of the $35B first tranche, a default with the racks recovered at half the lease amount E (estimate).³

The bull case is the outlook delivered. AI revenue is $115B and $230B, then $270B, $300B and $321B, reaching $393B in FY2036, with semiconductor margins of 60% to 62%. Its FY2028 earnings before stock pay are $30.60, which matches the ‘exceed $30’ E (estimate).⁵

The bear still has AI revenue at 1.5 times the TTM’s $42.4B in its worst year. The stress case does not. AI revenue falls to the TTM level, $42B, in FY2029, the semiconductor margin is 46%, the supply write-off is $26.5B and the whole $29B backstop is paid. It is worth $77 a share, −76% against the price E (estimate). It has no probability, so it sits outside the weighted value, and that is why the margin-of-safety price comes from the weighted bear.³⁴

Weighted 25%, 50% and 25%, the cases give $262 a share E (estimate), against a price of $326.08.⁵⁰ The weights are judgments. I give the bear one chance in four because two of the six XPU customers depend on financing, and the bull one in four because the outlook is management’s own claim and its FY2028 needs the growth rate to double. At 20%, 50% and 30% the weighted value is $279, and at 30%, 50% and 20% it is $245 E (estimate). The value after FY2036 is 53% of my base, 48% of the bear and 56% of the bull C (calculated).

What closes the gap

Only a higher AI path closes the gap between the price and my base: FY2028 AI revenue of $227B does it alone, and the outlook on time, at $330, sits just above the price. No down year in FY2030, or a semiconductor margin held at 58% or 60%, each close between a fifth and two fifths of the gap, and the discount rate alone moves the base by more: at 9.5% it is $305, still short of the price E (estimate).⁵⁰⁵

Cash flows are discounted at year end; a mid-year convention adds $14 a share to the base. Two notes on what the model holds. It treats the backstop as free in the base: a 10% chance of paying the full $29B would cost $0.59 a share E (estimate).³ It also counts the fourth quarter’s owner free cash flow of $13.7B, less the $3.1B dividend paid on 30 September, in net debt of $26.5B; without it the base would be about $2 lower. The $13.7B is the guided revenue of $34.8B at Q3’s owner free cash flow margin of 39.4%, which is Q3’s free cash flow of $13.7B less $2.0B of stock pay, over $29.6B of revenue E (estimate).⁴⁴⁶

The second backstop. On 27 October Broadcom filed an 8-K for a backstop of up to about $43B, on top of the $29B in the 10-Q if the two are separate; the filing does not say. A 10% chance of paying all $43B would cost $0.88 a share. My signposts list a new backstop as lowering my value, so this one counts against my value. I have not moved the case values: the 8-K gives no payment schedule and no default case, and the base takes no loss on backstops as I stated it.⁴⁹

The margin of safety

• Price$326.08Close, Q4 2026 [50]
• Base value$261My estimate E (estimate)
• Margin of safety below$126set from my bear case C (calculated)

I set the margin of safety from the weighted bear case, because one cycle and a handful of customers drive this business. Below $126, 52% under my base value, my bear value of $95 is at least three quarters of the price paid, and my base case would return about 22% a year E (estimate). That price is 6.2 times my FY2028 earnings after stock pay. It is also 57% below $293.41, the lowest close of the past year, and 63% below the $338 Broadcom paid on average for its buybacks in the nine months to August, so it describes a crash, not an entry. At $326.08 the stock is priced 25% above my base value and above the weighted value, and there is no margin of safety in the current range. The bear case is built on losses the filings make possible: $126.8B of purchase commitments and a backstop of up to $29B.³¹⁵⁰ I do not cut the value again for concentration or the cycle, because both are in the cash flows. The cushion between the bear value and the margin-of-safety price is for what the model does not hold, such as a Taiwan supply shock or the loss of Hock Tan, and for my own error. The price sits 159% above the margin-of-safety price.

The next best use of the same dollar is NVIDIA, the closest comparable here. On its page my base returns 9.7% a year, against 7.5% here, and its price is 5% above my base value, against 25% here.⁶⁴ A business that turns 44% of revenue into free cash flow is one question. Paying $326.08, 25% above my base, is another. The price needs four things at once: the labs pay, 60%; the self-funding four keep spending, 80%; no customer moves its next chip in-house, 80%; the margin holds near 66%, 80%. That multiplies to 31% E (estimate). They are guesses and overlap, so I read it as near the 25% I give the bull.

Where this could be in 5 and 10 years

These are the value of my model at later dates. Value in a year is the present value, at 10.5% and in that year’s dollars, of the owner cash that comes after it, per today’s share count, less today’s net debt of $26.5B. Owner cash is free cash flow after stock pay, all counted as paid out. Years are fiscal, to the Sunday nearest 31 October.

For the base to hold at 5 years, AI revenue has to reach about $183B in FY2030 after one down year, with free cash flow near $103B, and the business has to be growing 5% to 7% from there. That is worth $320 then. At 10 years it is $377 if free cash flow reaches $136B, a 41% margin, with semiconductor margins at 54% and 3% growth after. Value rises less than 10.5% a year because the cash paid out on the way is counted separately. The base returns 7.5% a year from today’s price, below the 10.5% I require E (estimate).

The bear and the bull differ on four numbers: AI revenue, the semiconductor margin, the capital the build-out ties up, and the discount rate. In the bear, the build-out stalls in FY2028, I write off $26.5B of supply, margins drop to 46% and free cash flow reaches $50B. In the bull, the outlook arrives on time at $115B and $230B, margins hold near 60%, and free cash flow reaches $220B. One point on the discount rate moves the base by 13% to 17%.

The most likely way the base is wrong is FY2028. It needs $167B of AI revenue, and the ramp of 2026 has to keep its pace through 2027. At $140B, the base is worth $232 today.

A reader’s guide: these are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Anthropic filed to go public on 26 October, and its S-1 gives my funding test its first lab accounts. Anthropic has five-year equipment leases with Broadcom, with rent of about $161.2B before interest. Neither side can cancel except on a default, and Anthropic itself guarantees them. It may also issue Broadcom up to $42B of notes to pay part of that rent, which matches the unnamed customer in Broadcom’s 10-Q. It had $131.7B of cash and short-term investments at 30 June, and it earned $1.0B of operating income on $11.6B of revenue in the second quarter.⁵³

From Q3 last year to Q3 this year, semiconductor revenue rose $11.7B and the segment’s research spending rose $156M. Segment profit rose $7.6B, so each added dollar earned 65 centsC (calculated), above the 61.3% average.³ Racks are supposed to dilute that margin. So far it has risen, though the segment’s gross margin slipped from 67.4% to 66.9%. My base fades it to 54%.

Alphabet’s inventory went from $2.4B to $13.3B in nine months, mostly TPU systems and devices. Thomas Kurian, who runs Google Cloud, says all the TPU inventory sits against signed contracts.⁵⁴⁵⁵

The bull case

Hock Tan makes it best. On 2 September he said Broadcom has secured the supply for about $115B of AI revenue in FY2027 and $230B in FY2028, and that FY2027 demand is higher. He said a custom chip costs less than half a GPU for the same work, and that each gigawatt could earn a lab $30B a year, so helping them pay makes sense.⁵ These are company claims.

JPMorgan’s Harlan Sur had a $580 target in June. Press reports have JPMorgan saying the Google deal “locks in Broadcom’s TPU design win roadmap” through v11, and that FY2028 AI revenue could approach $300B. That was before the September outlook.⁵⁶⁵⁸ Investing.com reports Bernstein raised its target to $575 on 3 September.⁵⁷ On 27 October DA Davidson’s Gil Luria raised his rating and took his target to $400. CNBC quotes his note: he expects Anthropic to raise its capital and OpenAI to follow, and says “AVGO is trading as if the chances of that happening are near zero.”⁵⁹ ARK’s 13F listed 407,202 shares at 30 June, about 1% of the portfolio, from none at the end of 2025. Three of its ETFs held 353,706 at 31 July, 1.2% to 2.3% of each fund.⁶⁰⁶¹

Against reality

If Broadcom delivers $115B and $230B at my base margins, and the labs pay, the price is possible. That path is worth $330 a share, and $227B in FY2028 alone gets to $326.08E (estimate).⁵⁵⁰

Today is a long way off. The Q4 guide runs at $87B a year, so $227B is 2.6 times that paceC (calculated).⁴ FY2027 needs 11.6% growth a quarter, which looks easy after 29% to 55% this fiscal year, counting the Q4 guideC (calculated).⁴²⁵²⁶ FY2028 needs 22.7% a quarter on far larger numbersE (estimate).

Funding is the other gap. The price pays for 98% of the $172.5B I give the two labs in FY2028, where my base takes 64%E (estimate). Anthropic’s rent averages about $32B a year, about 70% of its Q2 revenue annualizedC (calculated), and the S-1 leaves its total compute commitments blank.⁵³ Rent is also not Broadcom’s revenue: Broadcom has assigned part of the leases and intends to assign more. I found no public accounts for OpenAI. The self-funding side looks better: Alphabet and Meta spent $85.3B on capital in Q3, $341B a year against the $259B in my test, though both spent more than their operating cash flowC (calculated).⁵⁴⁴⁵

Where I come down

Anthropic’s filing moved me: part of my doubt about the labs is now a signed lease and real accounts. It did not move my $95, $261 and $431, because a lease pays only if the tenant can, OpenAI is still a blank, and the 27 October backstop puts more of Broadcom’s balance sheet behind its customers. I would raise my values if Q4 meets the guide, the $115B stands on 9 December, and more of the labs’ share shows up in filed contracts; at $200B of FY2028 AI revenue my base would be $297E (estimate).

VIIIRisks, and what would change my view

1. The labs’ funding. Two of the six XPU customers rely on the XPV platform, and Broadcom’s backstop of up to about $29B is my bear case’s trigger; on 27 October it filed a second, of up to about $43B (see Valuation).³⁴⁹⁵
2. Concentration. One distributor was 50% of Q3 revenue and the top five end customers about 55%.³
3. Customers’ own silicon. The 10-Q names customer-owned tooling, and Google works with MediaTek on TPU v8 (see The moat).³²⁷⁵
4. Supply. TSMC made about 95% of the wafers, and Broadcom has $126.8B of purchase commitments.²³
5. Margin. Gross margin is guided down as racks and memory grow; operating margin guided at about 66% is the line I watch.⁴⁵
6. Tax and stock pay. The 5.4% tax rate is expected to rise, and vested stock was worth three times its expense (see Financials).²³
7. The CEO. Tan is 74, has no employment contract and no named successor, and the CFO changed in June.²¹⁰³³
8. Software. Regulatory inquiries in Korea, Japan and the European Union, and churn evidence that comes only from a rival-sponsored survey.²³²⁸
9. Geography. China including Hong Kong was 17% of FY2025 revenue by delivery location, and TSMC is in Taiwan.²

What would raise my base value: the report Broadcom plans to give after the close on 9 December holding the $115B FY2027 outlook with the Q4 AI revenue at or above $21.7B, remaining performance obligations rising with more due inside twelve months, and the top five customers falling below half of revenue. What would lower it: any new backstop, which the 8-K of 27 October already is, or note issuance, inventory above $8B, an operating margin below 62%, or a Q4 AI figure under the guide.⁴³⁵