Broadcom builds the custom AI chips that big labs design
Custom AI chips
built for Google, Anthropic, OpenAI and Meta
Network chips
link AI chips together inside a data centre
VMware software
sold to companies on multi-year subscriptions
The companies with the largest AI models no longer rely only on NVIDIA's general-purpose chips. Each designs a chip around its own model, and Broadcom turns that design into working silicon in volume.
The company · 2 · What exactly does it sell?
AI chips are now more than half of sales
Each $1 of sales last quarter, by product
56%30%14%
AI chipsSoftware, mostly VMwareOlder chips for phones and broadband
It is paid per chip shipped. Each customer's chip runs over several generations, so one design brings years of orders. The software is sold on multi-year subscriptions.
The company · 3 · How does it make money?
Of each sales dollar, 46 cents arrives as spare cash
Of every $1 customers pay
73 centskept after the cost of making the chips
46 centsarrives as spare cash
Up 22%running costs, while AI chip sales grew 127%
Custom AI chips keep less after production costs than older products, so that first figure has slipped from 75 cents. But research and selling costs grow far slower than sales, so profit still rises.
Is it a good business? · 4 · What gives it an edge?
Switching chip builders costs a customer a whole generation
Through 2031
Google and Apple have signed on
MediaTek
started a Google chip first, yet ships it after Broadcom's
Each design rests on Broadcom's own connection and packaging parts, so moving it means redoing it, which costs a year the customer cannot spare. Google is keeping MediaTek anyway, so the hold is strong but not total.
Is it a good business? · 5 · What is the catch?
A few customers, and the newest ones lose money
55%of last quarter's sales came from five customers
$89.0BNVIDIA's data-centre sales in one quarter, the main rival
$2.0BNVIDIA put into Marvell, the other custom-chip builder
Its largest new customers, the AI labs, lose money. Anthropic also runs on NVIDIA chips, and a customer that keeps a second builder can push Broadcom's prices down.
Is it a good business? · 6 · What does its record show?
Sales more than tripled in five years, and costs lagged behind
Sales, $B
27.4
33.2
35.8
51.6
63.9
89.1
Oct 2021Oct 2022Oct 2023Nov 2024Nov 2025last 12 mo
Sales, latest quarter+86%against a year earlier
AI chip sales+221%to $16.7B
Spare cash, one quarter$13.7B46 cents of every sales dollar
This is fast growth with rising profit. Broadcom also grew by buying companies, most recently VMware; the recent surge is its own AI chips.
Is it a good business? · 7 · How much room is there to grow?
Custom chips are a growing slice of a fast-growing market
About 28%of this year's AI servers run on custom chips
$20B to $30Bof Broadcom chips per gigawatt, management says
About $11.5Bper gigawatt in one analyst's arithmetic
The industry sizes data centres by the power they draw. Nobody outside the company measures how many dollars of Broadcom chips each gigawatt carries, so the two estimates are far apart.
Is it a good business? · 8 · Who is in charge?
A deal-maker in charge since 2006, paid on AI sales
Since 2006Hock Tan has been chief executive
About 2%owned by insiders, nearly all the chairman's
2028 to 2030the AI sales years his share award depends on
Tan built Broadcom by buying companies, most recently VMware. His AI sales target sits below what management already expects for 2027, so it is not a hard one. Nothing in his pay rewards cash.
The story right now · 9 · Why is everyone buying AI chips?
AI companies are racing to build giant data centres
161 gigawattsof data-centre power worldwide in 2026, up about 31%
1/3of that power is for AI servers
Under 50%of a standard AI chip's running cost for a custom one, management says
Training and running AI models takes huge numbers of chips, so data centres are now sized in gigawatts of power. Labs want cheaper chips than NVIDIA's, which is Broadcom's opening. Whether they can keep paying is the argument.
The story right now · 10 · Who pays for all these chips?
Broadcom is helping its biggest new customer pay the bills
Anthropic rents data centresFilled with Broadcom's chips.
Anthropic loses moneySo its rent needs someone's backing.
Broadcom signs a guaranteeUp to $29B of rent if Anthropic cannot pay.
Sales look strongBut part of the risk has come back to Broadcom.
When a seller backs its buyer's bills, sales tell you less about real demand. Broadcom may also lend Anthropic $42B. We would worry if a customer ever borrows from Broadcom to pay its rent.
The story right now · 11 · What changed lately?
AI chip sales more than tripled in a year
+221%AI chip sales, to $16.7B
+86%total sales in the latest quarter
75 centskept per $1 after making chips, and slipping
$63Ba second Anthropic deal, again backed by Broadcom
Strong quarter
The shares fell after each of the last three reports, though each beat forecasts. The worry each time was profit per sale, financing or one customer's pace, never a missed number.
The story right now · 12 · What is everyone arguing about?
Is this growth real demand, or borrowed money?
Many investors think
The second-best way to own AI: few customers, and growth paid by their borrowing
We think
Next year's sales are more firmly signed than the price gives credit for
Broadcom has committed $52.7B to buy parts for 2027, and Google has signed to 2031. A company does not commit that much to a year it doubts. We are wrong if a customer borrows from Broadcom.
The price and our view · 13 · What is a share worth?
The price is about half our value
$656what we think a share is worth · the last close is of that
$656our value
Attractive below $525
Neutral
Unattractive above $788
$132gloomy: two loss-making labs cannot pay
$656our estimate: management's AI targets, then slowing
$1,176sunny: the 2028 plans all arrive
Value per share in three futures. None is a prediction of the share price.
We estimate what a share is worth from the cash the business should earn over many years. Our rule: buy only when the price is at least a fifth below our estimate, as a cushion for being wrong.
The price and our view · 14 · What if it all works?
The AI chips are already priced for success
Priced for successthe kind of bet this is
1.6×if it works
60%kept if it fails
Multiples of the company's value when the bet was last struck, 1 Oct 2026.
If the AI build-out stalls, what is left is software and older chips: about $52 billion of sales a year, very profitable, but worth far less than the whole company is valued at today.
The price and our view · 15 · How could we be wrong?
The biggest risk is customers who cannot pay
Financing its own customerBiggest
It guarantees up to $29B of Anthropic's rent.
Fewer, bigger customersShowing now
Five customers were 55% of last quarter's sales.
A second builderWatch
Google uses MediaTek; NVIDIA backs Marvell.
Who the others areUnknown
Management names four of six custom-chip customers.
Custom chips also keep less of each sales dollar than older products, so profit per sale slips as they grow.
The price and our view · 16 · How will we know?
December's report tests AI sales and the customer financing
AI chip sales reach $21.7B in the quarterGood sign
Management repeats its $115B outlookGood sign
A customer borrows from Broadcom to pay its rentWe turn neutral
DEC 20269
Results for its fourth quarter and full year. Companies report every three months; each report is a test of the story.
Our view also changes on price alone: if the shares rise to within a fifth of our estimate, the cushion is gone and we move to Neutral.
The price and our view · 17 · So, what do we think?
AI's chip builder, priced as if next year were the peak
$656our value
Attractive below $525
Neutral
Unattractive above $788
AttractiveMedium conviction
Broadcom's AI sales more than tripled and next year's are largely signed. The price sits well below our estimate. Conviction is Medium because its largest new customers lose money and lean on Broadcom's backing.
↑ ↓ to move · C for contents
What it does
More about the company
Who it is
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.
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.
Growth and profit
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.
Its record
The numbers
Is it growing?Fast
Sales grew 49% over the twelve months to Aug 2026, faster than the 24% in the year to Nov 2025.
Is it profitable?Strong
Operating margin is 48%, up from 45% in the year to Oct 2023.
Does growth pay off?Pays off
Running costs take 21¢ of each sales dollar, down from 30¢ in the year to Oct 2021.
Does it earn good returns?Strong
Each dollar put into the business earns 27% a year after tax, up from 13% in the year to Oct 2021.
Is it making cash?Strong
It turned 44% of sales into free cash, $39.4B over the last twelve months.
Is the balance sheet safe?Light debt
It owes $35.4B more than the cash it holds, about 0.9 years of the cash its operations bring in.
Are shares being diluted?Diluting
The share count rose 15% in 3.8 years, and pay in shares costs 9.5% of sales.
Why it wins
The evidence
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.
Who could beat it
The rivals in full
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.
Each rival
NVIDIA Vera Rubin and Blackwell Ultra GPU systemsNVDA
The accelerator purchase itself: the buyer runs its training and inference on merchant GPUs it can also rent from any cloud, instead of on a chip built for its model. NVIDIA's data center revenue was $89.0B in its latest quarter against Broadcom's $16.7B of AI revenue, and Anthropic and OpenAI both state they run on NVIDIA parts as well.
MediaTek, building Google's TPU version 8t2454.TW
Broadcom's implementation role on a Google accelerator generation: Google keeps the architecture and hands the physical design, packaging and production to a second vendor. Management says the 8t was initiated before Broadcom's 8i and is not yet shipping.
Marvell Technology, with NVIDIA's $2.0B backingMRVL
Broadcom as the custom-accelerator implementer and attach-silicon supplier at a hyperscaler: Marvell's own filing describes custom solutions to individual customer specifications for AI and data center buyers, its data center revenue grew 46%, and NVIDIA bought $2.0B of its preferred stock to connect Marvell's XPUs into NVIDIA's ecosystem.
Nutanix Cloud Infrastructure, and the hyperscalers' own managed private cloudNTNX
The VMware renewal itself. An enterprise facing a multi-year VMware subscription moves its virtual machines to another hypervisor or to a managed cloud instead of renewing. Broadcom's own 10-K names no rival at all here - it describes competitors only as classes, which is what makes the substitution unquantified rather than absent.
Qorvo and Skyworks Solutions in radio filters, and a handset maker bringing the part in-houseQRVOSWKS
The filter socket in a phone. The same 10-K names "the internal resources of large integrated OEMs" among its competitor classes - the customer becoming the supplier - and this is the business management itself guides to about $4.3B a quarter and flat.
How big it could get
The market and who has signed
The market
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.
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.
Who runs it
Leaders, pay and the cash
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.
What they 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.
The last report
The latest report
Broadcom's sales rose 86% as AI chip sales rose 221% to $16.7B
What it did to the thesis Strong quarter
It 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 report
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's happening now
The background
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.
Where we disagree
Both sides in full
The market
The second-best way to own AI: thinner margins than NVIDIA, a few customers, and growth paid for with those customers' borrowing.
We believe
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.
The argument
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 it's worth
How we value it
A business whose AI sales are guided by management with supply secured is valued on that guide for three years, slowing toward 4% by year ten, each year closing 30% of the gap, so the view reads what the price pays against what the outlook delivers if it lands.
What the price assumes
Sales growth in the first year, then slowing as ours does, at the same profit per sale.
The price assumed
42.2%
Our estimate
66%
Measured at the close of
6 Nov 2026
The three cases
Bear case$132
The two loss-making labs behind the outlook cannot pay (C09, C11): fiscal 2027 grows 20% and fiscal 2028 5%, the margin falls to 42% as accelerators dilute it (C16), and new capital earns only 9%.
Our value$656
Fiscal 2026 grows 66% to about $106B on the Q4 guide of $34.8B (C03); fiscal 2027 55%, about the $115B AI outlook with supply secured plus flat other sales (C04); fiscal 2028 45%, below the $230B line of sight; the official margin settles at 55% as amortisation rolls off and stock pay stays a cost.
Bull case$1,176
Fiscal 2027 and 2028 grow 64% and 62%, close to the $230B of fiscal 2028 AI revenue management sees (C04), backed by $126.8B of purchase commitments (C05), and the margin reaches 60%.
The bet
Seen as a bet
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.
What could go wrong
Every risk
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.
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.
What changes our view
Every route to a new view
Every route to a new view
→ Neutral
The share price rises to about $525, 80% of what we think the business is worth, so it is no longer cheap enough. 2027 AI sales land near the $115B outlook, so the rising price is the market agreeing rather than the business surprising.
→ Neutral
A customer borrows from Broadcom to pay its rent, or Broadcom pays anything under its lease guarantee. Five customers pass 65% of sales with no new custom-chip customer named.
The next test
It supports the case if
AI revenue at or above $21.7B, adjusted operating margin at or above 66%, the 10-K still reporting no notes issued and nothing paid under the backstop, the second financing structure announced on October 26 ($63 billion, third-party lenders) closed with Broadcom's support disclosed and bounded, and the $115B fiscal 2027 outlook reaffirmed or raised. One open-market purchase by a named officer would also count.
It pressures the case if
AI revenue below $20B or the fiscal 2027 outlook cut, adjusted operating margin below 63%, convertible notes issued to Broadcom or a backstop payment, or Anthropic's 2027 gigawatts reduced. Also pressures: Broadcom's support for the second structure proves to be an uncapped or larger guarantee than the $29 billion first backstop.
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.
The verdict
Scores and how we weighed it
The scores
Growth
3 of 4
Business quality
3 of 4
Financial position
3 of 4
Execution record
4 of 4
Per share capture
3 of 4
Potential
3 of 4
Risk profile
2 of 4
Valuation
3 of 4
Evidence quality
3 of 4
How we weighed it
ValuePositive
The price is less than half our value on the analysts' path, and asks far less growth than the signed outlook shows.
BusinessPositive
It designs the custom AI chips several of the largest buyers rely on and the network chips that connect them, and its record of delivering is the best on the roster.
GrowthPositive
Next year's sales are more contracted than the price credits, and the price pays nothing for the years after.
RisksNegative
The largest planned customers lose money, Broadcom guarantees and may lend to one of them, buyers are fewer and bigger, and some are building second sources.
RivalsMixed
Rivals are winning chip generations at some customers; it still holds most custom AI chip work.
The groupMixed
AI chip makers move together on spending news, so part of its fall and recovery is the group's.
Latest newsPositive
The latest report showed AI sales and cash surging and raised the yearly AI outlook; only the gross margin slipped.
In plain words
Gigawatt
A unit of electric power. AI data centres use so much electricity that the industry now measures their size by the power they draw, not by the number of chips inside.
In plain words
Our estimate of value
We forecast how much cash the business can earn over the next ten years and beyond, then ask what that stream is worth today to an investor who wants an 8% yearly return. We do it three times: a gloomy future, an expected one and a sunny one. We call a share Attractive only when its price is at least 20% below the expected estimate, and Unattractive when it is 20% above.