A $230 Billion Forecast That Rests on Six Phone Calls

Broadcom's AI business has six customers. Analysts estimate the top three account for more than 60% of a roughly $16 billion quarterly run rate.

That single fact reframes everything else, so start there.

Nvidia sells a merchant product to thousands of buyers. If one hyperscaler pauses, others absorb the capacity. Broadcom's XPU business is the opposite: it co-designs a bespoke accelerator for a specific customer, and that revenue exists exactly as long as that customer's programme does.

Nvidia carries market risk. Broadcom carries counterparty risk. On a growth chart those look identical. They are not remotely the same business.

And the guidance: ~$58B AI revenue in FY2026, ~$115B in FY2027, ~$230B in FY2028, with a target of exceeding $30 EPS in fiscal 2028.

Those are not normal numbers. Neither is the doubling, then doubling again.

Broadcom is the arms dealer of the anti-Nvidia trade

Here's the part that explains why this is happening, and it's the piece most coverage leaves out.

Broadcom's six XPU customers reportedly include Google, Meta, OpenAI and Anthropic. It delivered Ironwood TPUs to Google and Anthropic, is ramping next-generation TPU v8i shipments to Google with Anthropic deploying v8i in 2027, and is building OpenAI's first-generation accelerator, reportedly codenamed Jalapeno. Anthropic is described as on track to become Broadcom's largest XPU customer in 2027 and into 2028.

Look at that list again. Every one of those companies is trying to reduce its dependence on Nvidia — and Broadcom is who they call to do it.

Google's TPUs, Meta's MTIA, OpenAI's and Anthropic's custom silicon: these are the programmes that show up in every "hyperscaler builds own chip" headline. Broadcom is the common supplier underneath all of them.

So the honest way to describe this company isn't "the overlooked AI chip stock." It's the pure-play on everyone else's desire to stop paying Nvidia. Every custom-silicon press release is, somewhere in the footnotes, a Broadcom purchase order.

"Secured supply" doesn't mean what it sounds like

Broadcom says it has secured supply for the 2027 number and has line of sight on 2028. That's genuine and it matters — wafer allocation and advanced packaging capacity are the hard constraints in this industry, and locking them years out is a real competitive act.

But read the sentence precisely.

Securing supply proves Broadcom can build it. It does not prove anyone is obliged to take it.

Supply commitments run from Broadcom to its foundry partners. Customer commitments run the other way, and they're the ones that determine whether $230 billion happens. The coverage keeps treating those as the same guarantee. They aren't.

Which brings us back to the concentration. The FY2028 figure isn't a market forecast. It's the sum of six roadmaps — and roughly three of them do most of the work.

The stock fell anyway

The detail I find most telling: this guidance disappointed.

Reporting suggests some investors were positioned for FY2027 AI revenue nearer $130–150 billion. Against that, $115 billion read as a miss — and Broadcom sold off despite reporting one of the strongest quarters any semiconductor company has ever posted.

Sit with that. A company guided to double its AI revenue, then double it again, and the market's reaction was disappointment. That tells you exactly how much of this is already in the price, and how little room there is for an ordinary quarter.

What analysts say is still unresolved

Worth listing plainly, because it's the checklist for the next two prints:

→ Customer concentration — the oligopsony problem above → Google multisourcing — its largest historical XPU customer has options → Bookings conversion — commitments becoming revenue on schedule → AI networking mix— margin differs sharply from XPUs → First-generation XPU yields — new customers, new silicon, unproven ramps → The FY2027 bridge — how $58B becomes $115B, quarter by quarter

There's also a subtler concern being raised: the financing capacity of some of the newer AI customers. Several are private labs whose spending depends on continuing to raise capital. Broadcom's 2028 number quietly assumes the AI funding environment of 2026 persists through 2028. That's an assumption about capital markets, not about semiconductors.

My read

The instinct is right — Broadcom deserves far more attention than it gets, and this is arguably the most important AI earnings report nobody outside the sector read.

But I'd frame the reason differently. Broadcom isn't the quiet second Nvidia. It's the leveraged bet on hyperscalers succeeding at getting away from Nvidia — with all the upside and all the fragility that implies.

Fourteen consecutive quarters of AI-led growth have not yet tested what happens when one of three large customers pauses a programme. That test is coming, because it always does. The question is whether it arrives before or after the 2028 number is in the base.

Question for you: is a business with six customers and 200%+ growth a better risk or a worse one than a business with thousands of customers and 60% growth? I think most people are answering that with the growth rate, and the growth rate is the part that's easiest to see and least informative.

Address

Boston, Massachusets

Contacts

news@techproduit.com

Get your weekly newspaper