Broadcom's AI Revenue Tripled and It Named a $230 Billion Year

Broadcom reported fiscal Q3 on September 2: record revenue of $29.6 billion, operating income of $20.1 billion, and an operating margin of 67.9%. AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially.

Then Hock Tan did the thing that actually moved the conversation. He guided AI revenue to approximately $58 billion for fiscal 2026, $115 billion for fiscal 2027, and $230 billion for fiscal 2028 — and said Broadcom has secured supply to support the 2027 number, with line of sight to the supply needed for 2028.

Read the guidance shape, not the headline number

Three-year revenue guidance from a semiconductor company is close to unheard of. Chip cycles are volatile, customer roadmaps change, and CEOs who name a number 24 months out are creating a rod for their own back. Tan named three.

The trajectory is roughly a doubling each year: $58B → $115B → $230B. To put $230 billion in perspective, that single line item would exceed the total 2025 revenue of nearly every semiconductor company on earth other than Nvidia. Broadcom's entire company did $29.6 billion this quarter.

What makes it more than a promise is the supply comment. Tan is not projecting demand — demand for AI silicon has not been the constraint for two years. He is claiming he has locked the manufacturing capacity, which in the current environment is the harder half. Advanced packaging at TSMC, HBM allocation from the memory makers, substrate supply: those are the bottlenecks, they are contracted years ahead, and a company that has secured them has effectively pre-booked its revenue.

That is why the number is credible in a way a pure demand forecast would not be. It is also why it is a claim worth checking against TSMC and memory disclosures over the next several quarters.

The customer concentration is the whole story

Broadcom pointed to expanding business with Anthropic and OpenAI. That is a short list, and it is the risk and the thesis simultaneously.

Broadcom does not sell a general-purpose GPU. It co-designs custom accelerators — XPUs — with a handful of very large customers who have concluded that buying Nvidia at Nvidia's margins is unsustainable at their scale, and that a chip tuned to their specific model architecture beats a general one. Google's TPU program is the archetype; OpenAI's Jalapeño effort with Broadcom is the newest large entrant.

Each of those relationships is enormous and each is singular. A custom accelerator program represents years of joint engineering and a multi-billion-dollar commitment, which makes it extremely sticky — and makes the loss of any one of them a material event. There is no long tail here to cushion a defection.

The structural bet Tan is making is that the largest AI buyers permanently move a meaningful share of their compute off merchant GPUs and onto silicon they specify. If that is right, Broadcom sits in the best position in the industry: it captures the shift away from Nvidia without having to beat Nvidia at general-purpose accelerators. If the labs decide custom silicon is more trouble than the margin savings — a real possibility as model architectures keep changing faster than chip design cycles — the whole curve flattens.

The networking half nobody talks about

AI revenue at Broadcom is not only accelerators. AI networking is the other engine, and it is arguably the more durable one.

Every large training cluster is a networking problem before it is a compute problem. Tens of thousands of accelerators are useless if the fabric between them cannot keep them fed, and Broadcom's Ethernet switching silicon is the default answer for hyperscalers who do not want to be locked into Nvidia's InfiniBand and NVLink stack.

That business is less concentrated than custom accelerators, benefits from every cluster built regardless of whose chips are inside it, and is the piece least exposed to a single customer changing its mind. When Broadcom talks about $230 billion, a substantial share is fabric — and fabric demand scales with total industry buildout rather than with any one lab's roadmap.

Why the stock barely moved

Broadcom beat expectations and issued a cautious near-term forecast citing rising competition, and the shares went approximately nowhere.

That is a reasonable market response to a genuinely mixed disclosure. The three-year guidance is spectacular and unverifiable. The near-term caution is specific and immediate. Investors have also spent the year being told extraordinary things about AI revenue by nearly every company in the supply chain, and the marginal informational value of another very large number has declined.

The competition Tan flagged is real. Nvidia is not passive about custom silicon eroding its share, and its recent moves — the $3.5 billion into MediaTek convertible bonds to put NVLink Fusion inside rival XPU designs — are precisely aimed at making custom accelerators plug into Nvidia's rack architecture rather than replace it. If that works, Broadcom's customers get their custom chips and Nvidia still collects on the system.

What to watch

Whether the fiscal 2027 supply claim survives contact with TSMC's capacity disclosures. Tan says it is secured. TSMC has warned its largest customers about a capacity squeeze. Both statements can be true, but only one of them is a commitment.

Whether a third or fourth mega-customer appears. Anthropic and OpenAI expanding is good. Concentration risk only falls when the list gets longer.

Whether Nvidia's NVLink Fusion strategy converts Broadcom's customers into Nvidia's ecosystem members. That is the competitive question that determines whether $230 billion in 2028 is a floor or a ceiling.

Tan has now put a number on the board that the entire custom-silicon thesis will be graded against. He does not get to revise it quietly.