Let’s do the math quickly.
Broadcom (NASDAQ:AVGO) just reported $10.8 billion in AI chip revenue for a single quarter. That’s up 143% from a year ago. The company’s free cash flow was $10.3 billion — nearly half of its total revenue. It booked $30 billion in AI orders in the quarter, against $10.8 billion it actually shipped.
Three times more demand than supply. Record revenue. Record free cash flow. Record operating income.
The stock fell 13% in after-hours.
Something doesn’t add up. Or does it?
What Actually Happened
Let’s start with what the market saw on the surface.
Broadcom’s total revenue of $22.187 billion came in $80 million below the analyst consensus of $22.27 billion. That’s a 0.4% miss. On $22 billion. The software segment delivered $7.18 billion against a $7.32 billion estimate — another miss. And most importantly: CEO Hock Tan did not raise the company’s full-year AI chip revenue target beyond the $100 billion figure he had already telegraphed.
The market had been pricing in a guidance raise. It didn’t get one. And for a stock that had surged 17% in the two weeks heading into earnings, that was enough to trigger a wave of selling.
But here’s the thing. The headline numbers are almost entirely the wrong thing to focus on.
The “Chips Only” Reversal
The most important strategic development from last night’s call got almost no coverage.
When asked about the “rack versus chip” model — whether Broadcom would supply complete integrated AI systems to its customers — Hock Tan answered simply: “No racks. It’s all chip. We are only chips.”
That’s a reversal from earlier messaging. For months, investors had been pricing Broadcom as a company that might provide end-to-end AI infrastructure — not just the semiconductors inside it. The “rack” narrative would have made Broadcom a direct competitor to Dell (DELL) and HPE (HPE) in the AI factory buildout.
Tan just killed that thesis in three words.
What it means in practice: Broadcom’s TAM is enormous, but it’s the silicon layer only. The companies that build the systems around those chips — Dell, HPE, Supermicro — are not being displaced. And Marvell (MRVL), which we covered yesterday, is actually a beneficiary. Broadcom’s networking chips and Marvell’s networking chips both fit into a world where nobody is building the full rack.
The Six Customers — & The New One Nobody Saw Coming
Broadcom has six core AI chip customers. Hock Tan named four of them last night with extraordinary specificity:
Google: A long-term multi-generation agreement for TPUs and AI networking announced in April. Hock called it “very substantial in dollars” but acknowledged Google will have some diversity of sources. Translation: Broadcom isn’t Google’s only supplier, but it has a very large committed contract.
Continue Reading With Our 7-Day Free Trial
Anthropic: This is the new one. Broadcom is providing Anthropic with 1 gigawatt of TPU-based compute in 2026. In April, they signed a deal for another 5 gigawatts starting 2027. This is the company behind Claude — and it’s now one of Broadcom’s largest compute customers.
OpenAI: Silicon delivered, production on track for late 2026. 1.3 gigawatts committed for 2027, part of a 10-gigawatt deal running through 2029.
Meta: Partnership to deliver multiple generations of MTIA chips. 3 gigawatts through 2028, with 1 gigawatt starting delivery in H2 2027.
The other two customers were unnamed. Shipments begin late 2026. Purchase orders of $6 billion already received.
Total gigawatts planned for 2027: 10 gigawatts. Unchanged from prior guidance.
The $30 Billion Problem
Here is the number that should have made every investor’s jaw drop — and that the market almost entirely ignored.
In Q2, Broadcom shipped $10.8 billion in AI chips. But it received $30 billion in AI orders. In a single quarter.
Three times more demand than what it actually shipped. That’s not a supply chain problem — Hock Tan specifically said supply is fine and improving. That’s a demand picture so enormous that even Broadcom, with its extraordinary manufacturing relationships and cash position, is running well behind what its customers want.
And now Broadcom’s visibility into forward demand runs all the way to 2028 — up from 2027 just last quarter. The reason: customers are placing orders years in advance because building the infrastructure to actually receive and run these chips — the power grid connections, the data centers — takes that long to organize.
This is why Broadcom and Apollo and Blackstone are creating the XPV platform — a $35 billion fund (first tranche) specifically to help Anthropic and OpenAI access the compute capacity their contracts require. These frontier AI labs have enormous ambition but smaller balance sheets than Google or Meta. The XPV platform is the financing mechanism that lets them get there.
The $100 Billion Question
Here’s why the stock actually sold off.
Going into last night, many investors assumed that with $30 billion in quarterly bookings and 6 hyperscaler customers placing multi-year gigawatt orders, Hock Tan would raise his FY2027 AI revenue guidance above $100 billion.
He didn’t.
But read what he actually said more carefully: “If you drive on the same basis of what we’re seeing here, almost 2x what 2026 will be. In 2027, you will very easily see that it exceeds $100 billion.”
He’s not capping at $100 billion. He’s saying the math gets you there comfortably. FY2026 is tracking to $56 billion in AI chips. Doubling that is $112 billion. He reiterated “in excess of $100 billion” — not “approximately $100 billion.” And he said 2028 will be “substantial growth” from 2027.
The market wanted a number like $130 billion. It got a floor instead of a ceiling. That’s a communication choice, not a business deterioration.
What To Watch
Three things matter over the next 90 days.
First: whether gross margin stabilizes. In Q3, Broadcom guided gross margin down to 74% from 77%, purely because AI chips are a larger share of revenue than software. This is not a structural problem — Hock Tan said that explicitly. But it will be a recurring question every quarter as AI revenue grows.
Second: the XPV platform with Apollo. A $35 billion first tranche is being launched now. This is how Anthropic and OpenAI will fund their compute access. If this works, it becomes the template for how frontier AI labs — who can’t write $10 billion checks the way Google or Meta can — access custom silicon at scale. That’s a structural change in the financing of AI infrastructure.
Third: the unnamed two customers. Broadcom has six core customers. We know Google, Anthropic, OpenAI, and Meta. The other two begin shipping late 2026. Their combined purchase orders already stand at $6 billion. When their identity becomes public, it will move the stock.
Fourth — and this one is specific to Baptista Research readers: Next week, we are launching the Baptista Research Conviction Portfolio — our first-ever model portfolio of high-conviction, research-backed positions tracked in real time against the S&P 500. Broadcom is currently under active evaluation for a BASE tier position, entered at post-selloff price levels. The thesis is exactly what you’ve read above. The entry point, allocation, target price, and thesis break condition will be published exclusively to subscribers the moment the portfolio goes live.
The Bottom Line
Broadcom’s quarter was genuinely historic. $10.8 billion in AI chip revenue. $10.3 billion in free cash flow. $30 billion in bookings. Visibility into 2028.
The 13% after-hours drop reflects a market that had priced in a performance beyond historic. When you rally 17% in two weeks heading into an earnings report, you don’t just need to be good. You need to be extraordinary.
The business is extraordinary. The stock had gotten ahead of it. The selloff has now created the entry window the thesis required.
Here is what we are doing: Broadcom is entering the Baptista Research Conviction Portfolio as a BASE position next week at post-selloff levels — a 4% allocation, thesis-based stop, and a 12-month base case target of $235. We will publish the full position card — entry price, target, thesis break condition, and catalyst to monitor — exclusively on our subscriber page. The Conviction Portfolio tracker goes live alongside our next article.
If you are reading this as a free subscriber, this is the article that explains why the portfolio exists. The portfolio itself — every position, every update, every exit — is where your subscription pays for itself.
Disclaimer: The Baptista Research Conviction Portfolio is a hypothetical analytical exercise for informational and educational purposes only. It does not constitute investment advice. Baptista Research is not a registered investment advisor. All investment decisions should be made in consultation with a qualified financial professional.
