Broadcom
Turns the AI chips that Google, Anthropic, OpenAI and Meta design for themselves into working silicon, and sells the network chips that wire thousands of them together.
The scores
Growth3/4 ↑
Revenue up 86% and organic, but it narrows to one driver and the top five customers' share rose from 40% to 55% in a year.
Business quality3/4 →
Agreements filed to 2031 and a nine-month tape-out a customer credits, but Google is dual-sourcing a generation and NVIDIA now funds the alternative implementer.
Financial position3/4 ↑
Net debt of $37B is covered many times by cash flow and spread over seven years, but an undrawn backstop and note facility sit behind.
Execution record4/4 →
Every print in the pack beat, the AI outlook was raised, VMware delivered; the one blemish is the June reversal from racks to chips only.
Per share capture3/4 →
Buybacks of $8.45B only offset $6.3B of stock compensation, so the count is flat rather than shrinking; nothing converts, deals are behind it.
Potential3/4 ↑
A doubling with supply bought, Google signed to 2031 and a tranche closed is committed; the 2028 doubling and the labs' gigawatts are expected.
Risk profile2/4 ↓
Two concentrations, not one: five customers are 55% of sales, and the largest planned customer's rent is guaranteed and may be lent by Broadcom.
Valuation3/4 ↑
$19 per dollar of next year's implied profit prices the 2027 doubling as described and nothing for 2028; 44 times reported earnings is not cheap.
Evidence quality3/4 →
Every central figure is filed or from the customer's own release; the 2027 and 2028 outlook and content per gigawatt exist only on the call.
If it works 1.6× · if it fails you keep about 60%.
The Verdict
We rate Broadcom attractive. It turns the AI chips that Google, Anthropic and others design into working silicon and sells the network chips that wire them together; that business more than tripled in a year, so the business is solid. The price is ordinary for next year's expected profit and pays nothing for the growth after it, so it is fair. As a bet it is fragile: about one and a half times the money if AI sales reach what management sees, a little over half kept if the build-out stalls. Next year's sales are more contracted than the price credits, though the largest planned customers lose money.
The long caseThe biggest AI builders want chips designed for their own work, and Broadcom builds them over years together, so each new data centre they power up buys more from it.
The evidence under this - 9 claims
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
- The adjusted basis excludes about $8B a year of stock compensation, roughly $1.65 a share, so on a basis that expenses it the fiscal 2027 multiple is nearer 22.
- The fiscal 2027 figure assumes the 66% operating margin survives another year of accelerator mix that management will not guide beyond one quarter.
- The panel's fiscal 2028 estimate of $30.58 matches management's, so the market already carries the number; the low multiple prices doubt about its quality and durability rather than ignorance of it.
- Anthropic states run-rate revenue above $30 billion from about $9 billion at the end of 2025 and raised $30 billion in February 2026, and management says it is well on its way to a public listing that would change its credit.
- The fair value of the backstop is stated as not material, no notes have been issued, and Apollo and Blackstone underwrote the $35 billion without Broadcom lending a dollar.
- Only two of six customers use the platform; the other four are, in management's words, financially secure enough to fund themselves.
- 2026-10-26editeda second financing with Broadcom residual value support was announced; the exposure to the same customer is larger or at least not yet bounded
- Management describes the guarantees as modest residual value guarantees with a low probability of being called, supported by the labs' profitability trajectory and the racks' sustaining value, and the remedies include reselling the racks to the seller at a fixed price.
- The financing itself is underwritten by Apollo, Blackstone and banks rather than by Broadcom, and Apollo's own release says the partners fund it.
- Anthropic's first gigawatt is funded and already deploying, so the stress removes a customer that has paid for its 2026 capacity through a third party rather than one that has not started.
- Management says demand exceeds the $115B and that supply, not customers, sets the figure, so a lab's shortfall could be backfilled by the four self-funded customers.
- Non-AI semiconductors grew 2% in fiscal 2025 and management called enterprise spending slow to recover; the 5% is off a flat base, and wireless declined in the quarter.
- Software growth of 29% is VMware's move from perpetual licences to subscriptions and the price rises that came with it, which the 10-K describes as a transition; the recurring-revenue figure of 15% is the underlying rate and exists only on the call.
- The 10-Q reports the software segment's revenue and operating income and carries no non-AI semiconductor line, so the $4.2B cannot be checked against a filing.
The company
The companies running the largest AI models no longer rely only on NVIDIA's general-purpose chips. Google, Anthropic, OpenAI and Meta each design a chip around their own model, and Broadcom turns that design into silicon that works, in volume. Management says such a chip runs the customer's model at under half the cost.
Broadcom also sells the network chips that link accelerators together, older chips for phones and broadband, and VMware software, 30% of sales.
The evidence under this - 2 claims
- Anthropic's own release says it runs on Trainium and NVIDIA GPUs as well as TPUs, and OpenAI's says its chip deploys with Microsoft, so neither lab has committed its compute to Broadcom's parts alone.
- The two largest planned customers for 2027 and 2028 are the two whose deployments require the financing platform; the four self-funded customers account for about 4 of the 10 fiscal 2027 gigawatts on Bernstein's arithmetic that management accepted.
How it makes money
It is paid per chip shipped, and each customer's chip runs over several generations, so one design win brings years of orders. The software is sold on multi-year subscriptions.
The evidence under this - 3 claims
- Anthropic's own release says it runs on Trainium and NVIDIA GPUs as well as TPUs, and OpenAI's says its chip deploys with Microsoft, so neither lab has committed its compute to Broadcom's parts alone.
- The two largest planned customers for 2027 and 2028 are the two whose deployments require the financing platform; the four self-funded customers account for about 4 of the 10 fiscal 2027 gigawatts on Bernstein's arithmetic that management accepted.
Where the money comes from, stream by stream
| Stream | Share of sales (Q3 fiscal 2026) | How it earns it |
|---|---|---|
| AI semiconductors | 56% | Custom accelerators co-designed with a handful of AI labs and cloud companies, about three-quarters, and Ethernet switching and connectivity chips for AI clusters. |
| Non-AI semiconductors | 14% | Networking, broadband, storage, wireless and industrial chips sold to equipment makers, much of it through one distributor. |
| Infrastructure software | 30% | VMware's private-cloud software, mainframe and security software, sold as subscriptions and upfront licences to large companies. |
Who runs it
Hock Tan has run Broadcom since 2006 and built it by buying companies, most recently VMware; Amie Thuener, from Google's parent, became finance chief in June 2026. Insiders own about 2%, nearly all of it the chairman's. Tan owns little outright and is paid in shares he must still earn: first on the share price, then on AI sales in 2028 to 2030, at a target below what management already expects for 2027. Nothing pays on cash or on results per share.
The evidence under this - 5 claims
- The table excludes Tan's unvested 2023 and 2025 PSUs, so his economic stake in the share price is far larger than the 908,474 shares he holds outright, and he has agreed to hold the after-tax 2023 shares through fiscal 2030.
- Samueli sold about 1.4 million shares under a Rule 10b5-1 plan in March and June 2026 (C33), after the table's date.
- Management already guides fiscal 2027 AI revenue to about $115B (C04), above the $90B target and near the $120B maximum a year before the performance period opens, so the award pays mostly on a level already guided.
- Nothing in the plan pays on cash conversion, return on capital or per-share results, and AI revenue is management's own cut of the semiconductor segment rather than a filed line (C01).
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
Why it wins, and who could beat it
Why customers come back
Moving a chip to another implementer costs a customer a generation of time, because each design is built on Broadcom's own connection and packaging building blocks and would have to be redone. Google tested it: MediaTek started its version of a Google chip earlier and is shipping it later than Broadcom's.
From outside, the lock-in shows up as length. Google signed through up to 2031, and Apple through 2031. Neither agreement states a volume, and Google is keeping MediaTek anyway.
The evidence under this - 4 claims
- The switching cost did not stop Google hiring MediaTek for a generation, and management expects Google to keep more than one source.
- OpenAI's own page credits its team with the architecture and Broadcom with implementation and networking - the half of the work Marvell, now funded by NVIDIA, also sells.
- A 2031 agreement with undisclosed volume and no stated exclusivity is a duration of relationship, not of share.
- The Ultra Ethernet Consortium has published a 1.0 specification whose stated mission is an "open, interoperable" full-stack architecture for AI and HPC networking, "maintaining Ethernet interoperability". A scale-out layer that a standards body is deliberately making multi-vendor is the opposite of an incumbency, and this is the only witness here that is neither Broadcom nor one of its customers or rivals.
- 2026-09-23editedAdded the Ultra Ethernet Consortium as counter-evidence and as the moat's first non-filer source. Every other witness on this claim was Broadcom, a customer or a rival - all inside the filing system - and the one institution outside it argues against the networking half rather than for it.
- The same sentence concedes that Google has a second TPU supplier in MediaTek for a generation Google itself initiated first.
- Marvell's own filing describes custom accelerators for the same buyers and NVIDIA bought $2.0B of Marvell preferred stock in March 2026 to connect Marvell's custom XPUs to NVIDIA's ecosystem, so the alternative implementer is now funded by the merchant incumbent.
The threat
The obvious substitute is NVIDIA, whose data-centre sales reached $89.0B in its latest quarter, up 117%; Anthropic says it runs on NVIDIA chips too. The quieter threat is a second implementer: MediaTek on one Google generation, and Marvell, into which NVIDIA put $2.0B in March.
The evidence under this - 4 claims
- Anthropic is private, reports no audited figures, and files nothing; its run-rate revenue is its own statement.
- Anthropic says it runs on three vendors' chips and that Amazon is its primary training partner, so the TPU commitment is one of several and the 8-K itself conditions it on commercial success.
- The same sentence concedes that Google has a second TPU supplier in MediaTek for a generation Google itself initiated first.
- Marvell's own filing describes custom accelerators for the same buyers and NVIDIA bought $2.0B of Marvell preferred stock in March 2026 to connect Marvell's custom XPUs to NVIDIA's ecosystem, so the alternative implementer is now funded by the merchant incumbent.
The rivals: NVDA, MRVL, NTNX, QRVO, SWKS, and 1 more that is private, in-house or listed abroad.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| NVDA | NVIDIA Vera Rubin and Blackwell Ultra GPU systemsPriced against it: Different product | AI semiconductor solutions | On sale | Different product |
| MediaTek, building Google's TPU version 8tPriced against it: Listed abroad | AI semiconductor solutions | Announced | Listed abroad | |
| MRVL | Marvell Technology, with NVIDIA's $2.0B backingPriced against it: Same product | AI semiconductor solutions | On sale | Same product |
| NTNX | Nutanix Cloud Infrastructure, and the hyperscalers' own managed private cloudPriced against it: One segment only | Infrastructure software | On sale | One segment only |
| QRVO · SWKS | Qorvo and Skyworks Solutions in radio filters, and a handset maker bringing the part in-housePriced against it: One segment only | Non-AI semiconductor solutions | On sale | One segment only |
How big it could get
How big this could get
Measure it in power. The world's data centres can draw 161 gigawatts in 2026, up about 31%, a third of it for AI servers, and about 28% of this year's AI servers run on custom chips. Broadcom says each gigawatt carries $20B to $30B of its chips; nobody outside measures that.
The evidence under this - 3 claims
TrendForce itself says the headline gap overstates the shortage because behind-the-meter generation is not counted, and its 2030 figures are a forecast by a commercial research firm rather than a regulator's register.
A shipment share is not a revenue share: NVIDIA's data center revenue alone was $89.0B in its latest quarter against Broadcom's $16.7B of AI revenue, so in dollars the merchant GPU still carries most of the market.
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
Who has signed
Google and Apple have signed through 2031, neither with a stated volume. Anthropic's first gigawatt is paid for by a $35B lease deal, and about 3.5 gigawatts from 2027 depend on its continued success. OpenAI and Meta are announced programmes with nothing filed.
The evidence under this - 5 claims
- 2026-10-26editedthe lenders announced a second, $63 billion Anthropic structure on October 26, so 'none has closed' no longer holds
- Anthropic is private, reports no audited figures, and files nothing; its run-rate revenue is its own statement.
- Anthropic says it runs on three vendors' chips and that Amazon is its primary training partner, so the TPU commitment is one of several and the 8-K itself conditions it on commercial success.
- Anthropic's own release says it runs on Trainium and NVIDIA GPUs as well as TPUs, and OpenAI's says its chip deploys with Microsoft, so neither lab has committed its compute to Broadcom's parts alone.
- The two largest planned customers for 2027 and 2028 are the two whose deployments require the financing platform; the four self-funded customers account for about 4 of the 10 fiscal 2027 gigawatts on Bernstein's arithmetic that management accepted.
Where the growth is coming from, business by business
Where the growth is coming from
| Business | Share of revenue | Growth: last quarter | Last half-year | Last year | Share in 3 years at its half-year rate |
|---|---|---|---|---|---|
| AI semiconductor solutions | 70% | +127.4% | +104.0% | +74.7% | 92% |
| Infrastructure software | 30% | +29.0% | +19.0% | +14.5% | 8% |
| Whole company | 100% | +85.5% | +67.3% | +48.7% | 100% |
On the filed split through 2026-08-02, if each business keeps its last half-year's growth, the company grows about 79% next year and 93% in year 3, against 48.7% over the last twelve months. AI semiconductor solutions carries the case: it grows 104% against 19% for the rest, and would be 92% of revenue by then. A scenario, not a forecast: no business holds its rate.
How we read it: The business carrying the case: AI is $16.7B of the $20.8B semiconductor segment and non-AI grows 5%, so the segment's 104% is AI's. Still 3: one driver, and five customers are 55% of sales.
The results
The numbers
Sales grew 49% over the twelve months to Aug 2026, faster than the 24% in the year to Nov 2025.
Operating margin is 48%, up from 45% in the year to Oct 2023.
Running costs take 21¢ of each sales dollar, down from 30¢ in the year to Oct 2021.
Each dollar put into the business earns 27% a year after tax, up from 13% in the year to Oct 2021.
It turned 44% of sales into free cash, $39.4B over the last twelve months.
It owes $35.4B more than the cash it holds, about 0.9 years of the cash its operations bring in.
The share count rose 15% in 3.8 years, and pay in shares costs 9.5% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-04. Fiscal years end in November.
| Oct 2021 | Oct 2022 | Oct 2023 | Nov 2024 | Nov 2025 | 12 mo to Aug 2026 | |
|---|---|---|---|---|---|---|
| Size and growth | ||||||
| Sales | $27.4B | $33.2B | $35.8B | $51.6B | $63.9B | $89.1B |
| Sales growth | +15% | +21% | +7.9% | +44% | +24% | +49% |
| Profit | ||||||
| Gross margin | 61% | 67% | 69% | 63% | 68% | 69% |
| Operating margin | 31% | 43% | 45% | 26% | 40% | 48% |
| Net margin | 25% | 35% | 39% | 11% | 36% | 43% |
| What growth costs, per sales dollar | ||||||
| Research and development | 18% | 15% | 15% | 18% | 17% | 13% |
| Selling, general and admin | 4.9% | 4.2% | 4.4% | 9.6% | 6.6% | 4.7% |
| Pay in shares | 6.2% | 4.6% | 6.1% | 11% | 12% | 9.5% |
| Capital spending | 1.6% | 1.3% | 1.3% | 1.1% | 1.0% | 1.4% |
| Cash | ||||||
| Cash from operations | $13.8B | $16.7B | $18.1B | $20.0B | $27.5B | $40.7B |
| Free cash flow | $13.3B | $16.3B | $17.6B | $19.4B | $26.9B | $39.4B |
| Balance sheet | ||||||
| Cash and short-term investments | $12.2B | $12.4B | $14.2B | $9.35B | $16.2B | $24.0B |
| Debt | $41.5B | $39.5B | $39.2B | $67.6B | $65.1B | $59.4B |
| Net cash | −$29.3B | −$27.1B | −$25.0B | −$58.2B | −$49.0B | −$35.4B |
| Returns and shares | ||||||
| Return on capital | 13% | 20% | 24% | 8.4% | 18% | 27% |
| Shares (diluted) | - | 4.23B | 4.27B | 4.78B | 4.85B | 4.89B |
What growth costs
Growing has made Broadcom more profitable: research, selling and stock pay each take a smaller slice of every sales dollar than when VMware arrived. The cost that grows with sales is the chips themselves, because custom AI chips earn a thinner margin than the rest, so profit per sale rises even as the margin on the chips slips.
The evidence under this - 2 claims
- Gross margin is falling as custom accelerators grow (C16): cost of revenue is the line that grows with sales, so the widening sits below gross profit.
- Stock pay per sales dollar is still higher than before VMware, whose staff came paid in stock.
- Management guides margin one quarter at a time and would not say where gross margin settles as memory content in each accelerator rises.
- Marvell fell after its own print on the same custom-silicon mix effect, so the dilution is the product's, not one vendor's execution.
The last report (2 Sep)
Quarter ended August 2026, filed figures.
The third quarter of fiscal 2026 was large on every line: sales up 86%, AI chips up 221% to $16.7B, and $13.7B of free cash, 46 cents of every sales dollar.
The one figure that fell is the one the market watches. Custom chips keep less profit after production costs than Broadcom's older products, so that share slipped to 75 cents in the dollar and is guided to 73. Operating margin still rose to 67.9%, because costs grew 22% while chip sales grew 127%.
What it did to the thesisIt strongly supported the case: AI chip sales and cash surged, and the yearly AI outlook rose. The one weakness was the shrinking share of each sale kept after production costs, though operating margin still rose. Whether growth lasts beyond next year, which the price does not credit, stays open.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $3.32 against $3.22 expected; $2.68 as filed, the gap being what the company adds back.
The evidence under this - 4 claims
- Management guides margin one quarter at a time and would not say where gross margin settles as memory content in each accelerator rises.
- Marvell fell after its own print on the same custom-silicon mix effect, so the dilution is the product's, not one vendor's execution.
- The press-reported consensus for Q4 was about $35.03B, so the guide was read as a miss on the day and the shares fell 2.7%.
- Gross margin has been guided down every quarter since the accelerator ramp began - 79% a year ago to 73% - and management now declines to guide it beyond one quarter.
What management did with the money
Over nine months it paid $9.28B of dividends, bought back $8.45B of shares and repaid $6.05B of debt net of new borrowing, and still added $7.8B of cash. The buybacks did not shrink the share count, which rose 0.7%, because it paid staff $6.29B in stock.
Insiders in 2026Q2: 11 purchases and 62 sales, from their Form 4 filings.
The evidence under this - 2 claims
The price, and where we disagree
In short. Fair. An ordinary price for next year's expected doubling of profit, and nothing for the year after. The bigger risk is whether two loss-making AI labs can pay.
The business has run far ahead of the price. Over the last year sales per share rose 47.9% while the shares rose 3.3%, and they sit 27.8% below their high. Each of the last three falls followed results that beat forecasts, on a worry about margin, financing or a customer's pace rather than a missed number.
The evidence under this - 3 claims
- A flat multiple on a doubling earnings base is also what a market does when it believes the earnings are peaking: 44 times trailing is not a low number in absolute terms, and the six-year median it matches was set in years the company grew 8% to 24%.
- The panel raised its price targets 6.7% and its positive share rose after the September print, which is not a market abandoning the name.
How we read the multiple
On reported profit buyers pay about 44 times, near Broadcom's normal. On its adjusted profit, which leaves out about $2B a quarter of stock paid to staff and other costs, about 31 times this year's.
On the profit next year's outlook implies, about 19. The price needs next year's doubling and pays nothing for 2028's.
If neither Anthropic nor OpenAI takes its chips next year, profit would be about $11 to $13 a share, and the price 28 to 33 times that: dear, not ruinous.
Against its own history. On reported profit it costs about what it usually has while the profit underneath doubles; on sales and cash it looks dearer because the growth arrived first.
The evidence under this - 4 claims
- The adjusted basis excludes about $8B a year of stock compensation, roughly $1.65 a share, so on a basis that expenses it the fiscal 2027 multiple is nearer 22.
- The fiscal 2027 figure assumes the 66% operating margin survives another year of accelerator mix that management will not guide beyond one quarter.
- The panel's fiscal 2028 estimate of $30.58 matches management's, so the market already carries the number; the low multiple prices doubt about its quality and durability rather than ignorance of it.
- Anthropic's first gigawatt is funded and already deploying, so the stress removes a customer that has paid for its 2026 capacity through a third party rather than one that has not started.
- Management says demand exceeds the $115B and that supply, not customers, sets the figure, so a lab's shortfall could be backfilled by the four self-funded customers.
- The pack's own cross-check flags its trailing sales multiple, but the flag is caused by a vendor error in the pack's earnings history - the Q2 fiscal 2026 revenue row reads $1.14B against a filed $22,187M - and the 19.1 multiple is struck on the filed $89,104M and is right.
- Reported earnings include $8B a year of acquisition amortisation that ends within about three years; on the adjusted basis the trailing multiple is about 30.
Priced against its rivals
Only Marvell Technology sells the same thing to the same buyers as a large part of what it does. A price comparison needs at least two, so the price is judged against its own history instead.
Private or foreign rivals that cannot be priced: MediaTek Inc., Alchip Technologies and Global Unichip Corp.
The bet: what it pays if it works
This bet is on one business only: the AI chips Broadcom designs for Google, Anthropic, OpenAI and others. Software and the older chips are counted separately, at a price that assumes they never grow.
If AI sales reach what management says it can see for fiscal 2028, the stock is worth only about half as much again as today, so this is an expensive bet. If the build-out stalls, software and the older chips cover about a third of today's price: the price already pays for most of the ramp.
| AI semiconductor solutions revenue in two to three years | Gigawatts of custom accelerator compute a year needed | At the stated multiple |
|---|---|---|
| $115.00B | about 10 at 100% kept | 1.0× |
| $230.00B (the big win) | about 20 at 100% kept | 1.6× |
| $241.00B | about 21 at 100% kept | 1.6× |
To pay 2× (a Fair bet) it would need about $305.53B of AI semiconductor solutions revenue (about 27 gigawatts of custom accelerator compute a year); to pay 5× (a Cheap bet), about $850.50B of AI semiconductor solutions revenue (about 74 gigawatts of custom accelerator compute a year), at the same multiple.
Each row is AI semiconductor solutions alone, plus the rest of the company (Infrastructure software and non-AI semiconductor solutions) counted at about $520.00B.
The steps it has to take
- Done Google signed for custom TPUs and rack components to 2031, and Apple for custom silicon to 2031
- Done The first $35B financing tranche closed, funding over a gigawatt for Anthropic from mid-2026
- Done Supply bought for the ramp: $52.7B of fiscal 2027 and $73.0B of fiscal 2028 binding purchase commitments
- In progress Q4 fiscal 2026 AI revenue of about $21.7B, completing a $58B year due 2026-12-09
- Not started Anthropic's 5 gigawatts of TPU version 8i and OpenAI's first 1.3 gigawatts deployed during 2027 due 2027-12
- Not started Fiscal 2027 AI revenue near the $115B outlook, about ten gigawatts shipped due 2027-12
- Not started Fiscal 2028 AI revenue near $230B and earnings above $30 a share due 2028-12
The detail behind the label
The evidence under this - 5 claims
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
- Anthropic's first gigawatt is funded and already deploying, so the stress removes a customer that has paid for its 2026 capacity through a third party rather than one that has not started.
- Management says demand exceeds the $115B and that supply, not customers, sets the figure, so a lab's shortfall could be backfilled by the four self-funded customers.
- Non-AI semiconductors grew 2% in fiscal 2025 and management called enterprise spending slow to recover; the 5% is off a flat base, and wireless declined in the quarter.
- Software growth of 29% is VMware's move from perpetual licences to subscriptions and the price rises that came with it, which the 10-K describes as a transition; the recurring-revenue figure of 15% is the underlying rate and exists only on the call.
- The 10-Q reports the software segment's revenue and operating income and carries no non-AI semiconductor line, so the $4.2B cannot be checked against a filing.
- The adjusted basis excludes about $8B a year of stock compensation, roughly $1.65 a share, so on a basis that expenses it the fiscal 2027 multiple is nearer 22.
- The fiscal 2027 figure assumes the 66% operating margin survives another year of accelerator mix that management will not guide beyond one quarter.
- The panel's fiscal 2028 estimate of $30.58 matches management's, so the market already carries the number; the low multiple prices doubt about its quality and durability rather than ignorance of it.
Where we disagree with the market
The second-best way to own AI: thinner margins than NVIDIA, a few customers, and growth paid for with those customers' borrowing.
Next year's sales are more contracted than the price credits: Broadcom has bound itself to $52.7B of 2027 parts, and Google to 2031.
That evidence sits in the notes to the quarterly filing, in the purchase commitments and the terms of the lease guarantee, while coverage reads gross margin and the labs' essays. A company does not commit $52.7B to parts for a year it doubts.
We are wrong if a customer borrows from Broadcom to pay its rent, the second round of lease financing fails to close, or 2027 AI sales are guided below $100B.
What analysts say, and the evidence
The evidence under this - 7 claims
- Management describes the guarantees as modest residual value guarantees with a low probability of being called, supported by the labs' profitability trajectory and the racks' sustaining value, and the remedies include reselling the racks to the seller at a fixed price.
- The financing itself is underwritten by Apollo, Blackstone and banks rather than by Broadcom, and Apollo's own release says the partners fund it.
- Anthropic states run-rate revenue above $30 billion from about $9 billion at the end of 2025 and raised $30 billion in February 2026, and management says it is well on its way to a public listing that would change its credit.
- The fair value of the backstop is stated as not material, no notes have been issued, and Apollo and Blackstone underwrote the $35 billion without Broadcom lending a dollar.
- Only two of six customers use the platform; the other four are, in management's words, financially secure enough to fund themselves.
- 2026-10-26editeda second financing with Broadcom residual value support was announced; the exposure to the same customer is larger or at least not yet bounded
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
What could go wrong, and how we would know
The risks
- Financing its own customer. Broadcom guarantees up to $29B of Anthropic's data-centre rent and may lend $42B. On 26 October a second $63B Anthropic deal was announced, again backed by Broadcom.
- Customers building a second source. Google hired MediaTek for one chip generation, and NVIDIA put $2.0B into Marvell, the other custom-chip implementer.
- Fewer, bigger customers. Five customers were 55% of last quarter's sales, up from about 40% a year earlier, and one distributor handled half.
- Less per gigawatt. An analyst's arithmetic on management's own gigawatts gives about $11.5B of chips per gigawatt, not the $20B to $30B management states.
The evidence under this - 8 claims
- Management describes the guarantees as modest residual value guarantees with a low probability of being called, supported by the labs' profitability trajectory and the racks' sustaining value, and the remedies include reselling the racks to the seller at a fixed price.
- The financing itself is underwritten by Apollo, Blackstone and banks rather than by Broadcom, and Apollo's own release says the partners fund it.
- Anthropic states run-rate revenue above $30 billion from about $9 billion at the end of 2025 and raised $30 billion in February 2026, and management says it is well on its way to a public listing that would change its credit.
- The fair value of the backstop is stated as not material, no notes have been issued, and Apollo and Blackstone underwrote the $35 billion without Broadcom lending a dollar.
- Only two of six customers use the platform; the other four are, in management's words, financially secure enough to fund themselves.
- 2026-10-26editeda second financing with Broadcom residual value support was announced; the exposure to the same customer is larger or at least not yet bounded
- The same sentence concedes that Google has a second TPU supplier in MediaTek for a generation Google itself initiated first.
- Marvell's own filing describes custom accelerators for the same buyers and NVIDIA bought $2.0B of Marvell preferred stock in March 2026 to connect Marvell's custom XPUs to NVIDIA's ecosystem, so the alternative implementer is now funded by the merchant incumbent.
- Anthropic's first gigawatt is funded and already deploying, so the stress removes a customer that has paid for its 2026 capacity through a third party rather than one that has not started.
- Management says demand exceeds the $115B and that supply, not customers, sets the figure, so a lab's shortfall could be backfilled by the four self-funded customers.
- Bernstein's Stacy Rasgon computed $11B to $12B of revenue per gigawatt from the same outlook and gigawatt count, against management's $20B to $30B; management's answer was that not every gigawatt outlined will deploy inside the fiscal year.
- Roughly six of the ten fiscal 2027 gigawatts belong to Anthropic and OpenAI, whose deployments Broadcom is financing through a platform whose first tranche covers about 1 gigawatt; the second has not closed.
- In June the same management held the fiscal 2026 AI figure at $56B while analysts modelled above $60B, and said it would ship chips rather than the racks it had previously described.
What we still do not know
- Who the fifth and sixth custom-chip customers are; management names four of six.
- What Anthropic's gigawatts are worth in dollars, and whether it can walk away. It files no accounts.
- Whether the $29B guarantee is really worth as little as the filing says.
The evidence under this - 3 claims
- Anthropic's own release says it runs on Trainium and NVIDIA GPUs as well as TPUs, and OpenAI's says its chip deploys with Microsoft, so neither lab has committed its compute to Broadcom's parts alone.
- The two largest planned customers for 2027 and 2028 are the two whose deployments require the financing platform; the four self-funded customers account for about 4 of the 10 fiscal 2027 gigawatts on Bernstein's arithmetic that management accepted.
- Anthropic is private, reports no audited figures, and files nothing; its run-rate revenue is its own statement.
- Anthropic says it runs on three vendors' chips and that Amazon is its primary training partner, so the TPU commitment is one of several and the 8-K itself conditions it on commercial success.
- Management describes the guarantees as modest residual value guarantees with a low probability of being called, supported by the labs' profitability trajectory and the racks' sustaining value, and the remedies include reselling the racks to the seller at a fixed price.
- The financing itself is underwritten by Apollo, Blackstone and banks rather than by Broadcom, and Apollo's own release says the partners fund it.
What would change our view
- Attractive→NeutralThe 2027 doubling gets priced as certain rather than as due
It needs all of:
- ○The share price rises until this year's adjusted profit costs what reported profit already does, and next year's discount is gone.about $474.52 today, 38% above the last close
- ○2027 AI sales land near the $115B outlook, so the rising price is the market agreeing rather than the business surprising.checked at the next results, due by 9 Dec
- Attractive→NeutralThe customer backstop stops being a financing question
It needs any of:
- ○A customer borrows from Broadcom to pay its rent, or Broadcom pays anything under its lease guarantee.watched as it happens
- ○Five customers pass 65% of sales with no new custom-chip customer named.checked at the next results, due by 9 Dec
We would also be wrong if
- 2027 AI sales come in below $90B, against the $115B outlook.
- Google's next chip ships from another supplier while Broadcom's share of it falls.
- Adjusted operating margin stays below 60% for two quarters running as custom chips grow.
What is coming
- December 2026 Fiscal 2026 Form 10-K: Whether any convertible notes were issued or backstop paid under the AI XPV platform, full-year customer concentration, and the purchase commitments by year.
Every dated event ahead
- December 2026 Fiscal 2026 Form 10-K: Whether any convertible notes were issued or backstop paid under the AI XPV platform, full-year customer concentration, and the purchase commitments by year.
- 9 December 2026 Q4 and full-year fiscal 2026 results: AI revenue against the $21.7B guide, operating margin against 66% as gross margin dilutes, and whether the $115B fiscal 2027 AI outlook is kept.
- March 2027 Q1 fiscal 2027 results: The first quarter of the year management says brings about $115B of AI revenue, Anthropic's version 8i ramp, and the Singapore substrate fab supplying.
- March 2027 2027 proxy statement: Whether the fiscal 2026 bonus again paid the maximum, whether AI-revenue goals are reset above the guided level, and what insiders own after this year's sales.
Since we wrote this
- 26 OctNewsSecond $63B Anthropic financing announced, Broadcom backing itBanks and Blackstone will fund chips leased to Anthropic, with Broadcom supporting the first lien. The shape repeats the June deal; the size of Broadcom's exposure is not yet stated.
- 9 OctNewsOpenAI's $50 billion revenue pace and early financing talks weigh on Broadcom. View heldOpenAI's reported $50 billion revenue run-rate lands as Broadcom's OpenAI chip financing is reported in early talks.
- 8 OctNewsBroadcom reportedly seeks over $50 billion of lender funding for OpenAI chips. View heldWSJ reports Broadcom is arranging more than $50 billion of financing for OpenAI chips.
- 5 OctThesis reviewThe bet now names what it ridesThe owner asked that every bet name the macro trend behind it, after BWXT's bet was missing its chief executive's case for nuclear over gas.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the proxy and filings.
- 3 OctNewsBlackstone and banks line up $60 billion for Broadcom's AI chip financing. View heldBloomberg reported Blackstone and banks gathered $60 billion for a Broadcom AI chip financing.
- 2 OctNewsAnthropic's prospectus restates the $42 billion Broadcom facility the 10-Q already carried. View heldAnthropic's IPO prospectus describes the up-to-$42 billion Broadcom financing.
- 19 SepThesis writtenAttractive, medium convictionWritten two weeks after third-quarter results, with the shares 27.8% below their high while sales grew 86%.
Also checked 3 times with nothing new to act on.
Sources 33
Sources - 33
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-09-10the balance sheet and cash flow statement, the $126,821M purchase-commitments table by fiscal year, the AI XPV platform note with the $29B maximum backstop liability and the $42B convertible-note facility, customer concentration of 50% to one distributor and 55% to the top five end customers, the debt note, the repurchase note, the cover-page share count of 4,773,629,865 at 28 August 2026, and the risk factorsbacksC01C05C06C07C08C17C18C19C20C22C49C50
- S02primarycompany filing2026-09-02Q3 fiscal 2026 statements, the segment split, the Q4 guide of $34.8B revenue and 66% non-GAAP operating margin, the full GAAP-to-non-GAAP reconciliation including $2,006M of acquisition amortisation and $2,019M of stock compensation in the quarter, the cash flow statement with buybacks, dividends and debt repayment, and the quarterly dividendbacksC01C02C03C16C17C19C20C21C22C41
- S03primarytranscript2026-09-02Q3 fiscal 2026 earnings call transcriptmanagement's AI revenue of $16.7B, the fiscal 2026 AI figure of $58B, the fiscal 2027 outlook of about $115B and fiscal 2028 line of sight of about $230B, the per-customer gigawatt schedule, $20B to $30B of content per gigawatt, the XPV platform with Apollo and Blackstone, gross margin by segment, the $59.6B fixed-rate debt at a 4% coupon, and eleven analyst questions on supply, financing, margin and content per gigawatt
In 2027, we have secured the supply to again double AI revenue to approximately $115 billion. Our demand actually exceeds this outlook, and we will work to improve supply. In 2028, we expect the trajectory of growth to continue. We have line of sight for fiscal 2028 AI semiconductor revenue growth to again double to $230 billion.
backsC01C03C04C07C08C11C15C16C18C19C29C41C43 - S04primarycompany filing2025-12-18the business description by end market and software portfolio, the Competition section that names classes and no company, distributors at 48% and the top five end customers at about 40% of fiscal 2025 revenue, the manufacturing note naming TSMC and the internal FBAR and laser fabs, 33,000 employees with 57% in R&D, and the fiscal 2025 statementsbacksC06C27C40C44C47C48C49
- S05primarycompany filing2025-12-11fiscal 2025 revenue of $63,887M, Q4 fiscal 2025 revenue of $18,015M, and the Q1 fiscal 2026 guide of $19.1B revenue and $8.2B AI revenuebacksC26
- S06primarytranscript2025-12-11Q4 fiscal 2025 earnings call transcriptfiscal 2025 AI revenue of $20B up 65%, the AI order backlog above $73B to ship over eighteen months, the consolidated backlog of $162B, the AI switch backlog above $10B, and the Q1 fiscal 2026 gross margin guide down about 100 basis points on AI mix
All these components combined with our XPUs bring our total order on hand in excess of $73 billion today, which is almost half Broadcom's consolidated backlog of $162 billion. We expect this $73 billion in AI backlog to be delivered over the next eighteen months.
backsC26 - S07primarytranscript2026-06-03Q2 fiscal 2026 earnings call transcriptthe fiscal 2026 AI outlook held at $56B, the chips-only statement, Google's expected diversity of sources, $30B of AI bookings in the quarter, AI networking at about 40% of AI revenue, and twelve analysts' questions on gross margin, financing and supply
we fully expect that there will be some diversity of sources for them. But our commitment from them is a very substantial dollar amount.
backsC04C11C29C38 - S08primarycompany filing2026-04-06the Long Term Agreement to develop and supply Google's future TPU generations and the Supply Assurance Agreement through up to 2031, and Anthropic's approximately 3.5 gigawatts of TPU capacity through Broadcom from 2027, stated as dependent on Anthropic's continued commercial successbacksC09C10
- S09primarycompany filing2026-07-06new multi-year long-term agreements with Apple through 2031 for custom ASIC silicon across multiple generations of Apple productsbacksC28
- S10primarycompany filing2026-04-02Kirsten Spears' retirement effective June 12, 2026 and the appointment of Amie Thuener, Alphabet's chief accounting officer since 2018backsC39
- S11primaryvendor data2026-09-19Research pack, frozen 2026-09-19the market tier: last close $357.61, the 252-day range, the price-move decomposition, the analyst panel and its history, price-target drift, forward estimates by panel size, the vendor valuation ratios against six-year medians, FINRA short interest, insider statistics, the FMP daily price history behind the repricing episodes, and headcount rows citing the 10-KbacksC20C21C22C25C33C34C40
- S12primarycompany filing2026-08-26data center revenue of $89.0B up 117% on the Blackwell Ultra ramp, and the risk factor naming customers who develop their own internal solutionbacksC12
- S13primarycompany filing2026-08-28quarterly revenue of $2.7B up 37% with the data center end market up 46%, the NVIDIA strategic partnership connecting Marvell's custom XPUs to NVIDIA's ecosystem, NVIDIA's $2.0B purchase of Marvell convertible preferred stock, and a risk factor naming Broadcom's consolidationbacksC13
- S14primarycompany filing2026-03-11Marvell's own description of developing custom semiconductor solutions tailored to individual customer specifications for AI and data center applications - the same product to the same buyerbacksC13
- S15primarycompany filing2026-07-23capital expenditures of $44.9B in the quarter and purchases of property and equipment of $80,598M in the half against $39,643M, $49.6B of equity and mandatory convertible preferred raised for AI infrastructure, and product sales described as primarily the sale of TPU systemsbacksC32
- S16primarycompany filing2026-07-30anticipated 2026 capital expenditures of approximately $130 billion to $145 billion to support AI effortsbacksC32
- S17primarycompany release2026-06-24the customer's own account of who designed the chip and what Broadcom did, the nine-month design-to-tape-out, the gigawatt-scale deployment with Microsoft beginning in 2026, and the early performance-per-watt claim
OpenAI designed the chip from scratch around its deep understanding of LLM fundamentals, informed by its roadmap of models, kernels, serving systems, and product needs, with partners Broadcom and Celestica, helping industrialize the platform through chip implementation, board, rack system integration, high-performance networking, and scalable production systems. Jalapeño is the first step in a multi-generation compute platform designed for initial deployment by the end of 2026 and expanding in the years ahead, combining OpenAI-designed accelerators with Broadcom silicon implementation, networking, and connectivity technologies; and Celestica's board, rack, and system expertise.
backsC14 - S18primarycompany release2026-04-06the customer's own statement of the agreement, its run-rate revenue above $30 billion from about $9 billion at the end of 2025, and that it trains and runs Claude on AWS Trainium, Google TPUs and NVIDIA GPUs with Amazon as its primary cloud provider
We have signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity that we expect to come online starting in 2027.
backsC09 - S19primarycompany release2026-04-06Google's own statement that TPUs reach Anthropic through Broadcom as well as through Google Cloud
This capacity expansion will be delivered through Google Cloud services, as well as access to Google-built TPUs supplied through Broadcom.
backsC09 - S20primarycompany release2026-06-09Apollo Global Management press release - $35 billion capital solution for Broadcom AI XPV Platform Apollothe financier's own statement of the $35 billion initial tranche, the Blackstone and bank partners, the 20 gigawatt platform through 2028, and that the first transaction funds Anthropic's expansion of more than 1 gigawatt starting mid-2026
The Platform is designed to enable over 20GW in compute capacity for leading frontier AI labs through 2028.
backsC08 - S21primaryindustry data2026-01-20the ASIC share of AI server shipments, GPUs at 69.7%, AI server shipment growth above 28%, the top-five North American cloud providers' capital spending up 40% in 2026, and TPUs being sold to external clients such as Anthropic
The share of ASIC-based AI servers is expected to reach 27.8% by 2026, the highest since 2023, as North American firms like Google and Meta expand their own ASIC efforts. GPUs will remain the leading category, accounting for 69.7% of shipments. TrendForce data show that the combined capital expenditures of the top five North American CSPs-Google, AWS, Meta, Microsoft, and Oracle-are expected to increase by 40% YoY in 2026.
backsC30 - S22primaryindustry data2026-09-16the gigawatt base the prize and the driver are struck on, AI servers' share of it at about 25% in 2025 and 33.4% in 2026, the 490.7 GW 2030 projection against 222.6 GW of grid capacity, and the expected grid delays after 2028
Global data center power demand capacity reached 122.9 GW in 2025 and is estimated to rise to 161 GW in 2026. In 2027, continued CapEx expansion by CSPs and sustained overall server shipments are expected to keep annual growth in data center power demand capacity above 30%. AI servers accounted for nearly 25% of total data center power demand capacity in 2025, with their share rising to approximately 33.4% in 2026. This share could potentially exceed 40% in 2027. By 2030, global data center power demand capacity is projected to reach 490.7 GW. Assuming all other electricity demand is fully met, grid capacity available to data centers is estimated at only 222.6 GW. This leaves a supply-demand gap of 268 GW. However, after 2028, delays in grid power delivery are expected to become more frequent, substantially contributing to the growing gap between data center power supply and demand capacity.
backsC31 - S23secondarynews2025-12-12the stated objection on the day: the Q1 fiscal 2026 gross margin guide of about 76.9% against 79% a year earlier, and analysts questioning the $73B backlogbacksC25
- S24secondarynews2026-06-04the stated objection on the day: the fiscal 2026 AI outlook not raised and the chips-only statement, with the close at $418.91 down 12.59% on 79.9 million sharesbacksC25
- S25secondarynews2026-09-03the stated objection on the day: a Q4 guide of about $34.8B against a consensus of about $35.03B, a beat judged too thin for a stock this levered to AIbacksC25
- S26secondarynews2026-09-14the stated objection on the day: Anthropic's chief executive's essay calling for a slower pace of capabilities development, Broadcom shares down 4.8%, and Hock Tan's statement that the fiscal 2027 and 2028 forecasts are intact. The page returns 403 to every non-browser client; read through the search index and cited for the objection onlybacksC25
- S27primarycompany filing2026-09-17the Rule 10b5-1 flag on every open-market row: Henry Samueli's sales all under a plan dated December 16, 2025; Mark Brazeal's, Charlie Kawwas', Ram Velaga's and two directors' sales discretionary; Harry You's two 1,000-share purchasesbacksC33
- S28primarycompany filing2026-06-03Q2 fiscal 2026 revenue of $22,187M, non-GAAP earnings of $2.44 a diluted share, and the Q3 guide the September print was measured against; read for the prior-quarter comparable and for the repricing episode of June 4, 2026backsC25C38
- S29primarycompany filing2026-06-18the tender offers for six series of 2030-to-2038 notes, upsized from $2.5 billion to $3.0 billion of consideration, with about $5.5 billion tendered and about $2.9 billion accepted - the June leg of the $5,641M of notes repurchased in the quarterbacksC18
- S30primarycompany filing2026-06-18the consideration cap raised from $2.5 billion to $3.0 billion, about $5.5 billion of notes tendered and about $2.9 billion accepted, settling June 18 and 23, 2026backsC18
- S31primaryother2026-09-23Ultra Ethernet Consortium - mission statement and Specification 1.0.3 (ultraethernet.org, home page and specification download) Ultra Ethernet ConsortiumA standards body, not a filer and not a vendor, whose stated purpose is to make the AI scale-out layer open and interoperable across suppliers. That bears directly on the networking half of the moat and it points the WRONG way for the thesis: an interoperable standard is what dissolves an incumbency in scale-out silicon. Specification 1.0 is published and at revision 1.0.3.
Deliver an Ethernet based open, interoperable, high performance, full-communications stack architecture to meet the growing network demands of AI & HPC at scale ... UEC specifically focuses on functionality, performance, TCO, and developer and end-user friendliness, while minimizing changes to only those required and maintaining Ethernet interoperability. ... Backwards compatibility to widely-deployed APIs and definition of new APIs that are better optimized to future workloads and compute architectures.
backsC35 - S32primarycompany filing2026-03-02beneficial ownership as of February 24, 2026, director and officer biographies, the fiscal 2025 annual bonus goals and payout, the 2023 and 2025 Tan PSU awards and their metrics, the other officers' relative-TSR PSUs, and the pledge of 16,175,000 of the chairman's sharesbacksC44C45C46
- S33primarycompany release2026-10-26Bank of America, Blackstone, Citi and Morgan Stanley press release - $63 billion financing structure for AI compute capacity in partnership with Broadcom (PR Newswire) Bank of America, Blackstone, Citi and Morgan Stanleythe lenders' own statement of a second, $63 billion special-purpose-vehicle financing to buy Broadcom accelerators and networking for lease to Anthropic, deliveries from 2027, first lien by Bank of America, Citi and Morgan Stanley, second lien committed by Blackstone alone, and Broadcom's residual value support equal to the first lien amount. The release gives no first lien figure and no filing by Broadcom was found.
Broadcom is providing residual value support to the SPV for the amount equivalent to the first lien portion of the financing.
backsC51