Broadcom's Blowout AI Quarter Just Got Sold. The Inference Shift Is Why 2027 Looks Cheap

Broadcom just reported one of the best quarters of its life, and the stock fell anyway. Revenue rose 86% to $29.59 billion, and AI chip revenue hit $16.7 billion, up 221% from a year earlier. Adjusted earnings of $3.32 a share topped estimates for a ninth consecutive quarter. By Friday the shares had dropped another 4%, to about $353, down roughly 16% over the past month. If you are new to this, that looks backwards — a company growing that fast, sold off on its best quarter. The confusion is the point worth resolving.

The selloff isn't about the numbers Broadcom delivered. It's about the number it didn't. The company guided next quarter's revenue to about $34.8 billion — a hair below the roughly $35 billion Wall Street had quietly hoped for. A stock priced for perfection doesn't just need to beat expectations; it needs to beat the raised expectations everyone is already banking on. Miss the whisper number by a rounding error, and the whole trade unwinds for a day. That is a statement about expectations, not about execution.

Here is the part that makes this worth separating, rather than dismissing as overreaction: underneath that optics, management kept raising its own bar. Full-year AI revenue guidance went from $56 billion to $58 billion, and Hock Tan — who likes to say "we try to be conservative" — laid out an AI chip business that nearly doubles again to about $115 billion in 2027 and $230 billion in 2028. Demand, he said, currently exceeds even the raised outlook.

The inference bet is the whole question

To a beginner, "Broadcom" sounds like a chip company competing with Nvidia head-on. That's the wrong frame, and it matters. Nvidia sells a general-purpose chip that works for every AI job and wins on training, where its CUDA software moat is deepest. Broadcom is the designer behind custom chips that hyperscalers commission for themselves — Google's TPUs, Meta's MTIA accelerators, and a new inference chip Broadcom built with OpenAI called Jalapeño, which went from design to production in nine months and, per the two companies, delivers materially better performance per watt than current state of the art.

That distinction is where the AI cycle is turning. Training is the phase where you build a model once, burning as much compute as you can afford. Inference is the phase that runs after — every time a user queries a model — and it is where cost, latency, and power per task dominate. Training runs on Nvidia's moat. Inference rewards the chip that does a given job most cheaply at scale, which is precisely the case custom silicon makes. Hyperscalers do not build their own chips out of enthusiasm; they build them to stop paying Nvidia's prices on the workloads they run billions of times a day. When CEOs talk about a "training-to-inference" shift, Broadcom is the architecture that stands to gain as demand scales.

This is why the demand signals — the forward orders, not the price target — are the load-bearing evidence. Broadcom's custom silicon pipeline reads like a series of building-sized commitments: Anthropic moving from 1 gigawatt deployed this year toward 5 gigawatts of TPUs in 2027 and another 10 in 2028; OpenAI scaling Jalapeño to 1.3 gigawatts in 2027 and over 5 gigawatts in 2028; Meta with a line of sight to 3 gigawatts through 2028. Put simply, a gigawatt is a rough physical reading of how many chips a customer is buying and the power they will draw — the unit hyperscalers use when they commit compute years in advance.

Why "19x 2027" isn't just a bull's math

Now the cheaper-than-it-looks case, and it rests on which earnings you measure. On trailing results, Broadcom is expensive — somewhere north of 40 times last year's earnings. But trailing earnings are about the past, and this business is compounding fast. Simple arithmetic: fiscal 2026 adjusted earnings are on track to land near roughly $12 a share, and if AI revenue doubles next year as guided while margins hold, 2027 earnings could push toward the high teens. At a $352 share price, that is a stock trading near 20 times 2027 earnings — not cheap, but a far cry from the multiple the trailing number suggests, and markedly cheaper than peers like Marvell or Advanced Micro Devices on the same basis.

That arithmetic is only as good as the doubling it assumes. This is where Broadcom's gigawatt commitments carry a genuinely different weight than most analyst forecasts, because they are contractual supply commitments from the biggest AI spenders on earth, not screen-driven guesses. A customer that has committed multi-gigawatts of purchasing has already signed up capital and power.

What would make this a mirage instead

But a commitment is not a check. Broadcom's own CEO named the binding constraint, and it was not demand — it was "land, power, and shell," the physical capacity to stand up data centers, plus leading-edge wafers, substrates, and the high-bandwidth memory AI chips need. Hock Tan said outright that demand would let the company ship "significantly more" chips if the infrastructure to deploy them existed. That is a supply-and-delivery problem, and it is exactly the kind that decides whether a rosy 2028 number becomes revenue or stays a PowerPoint.

There is also a competitive fence around the whole thesis. Nvidia has noticed that inference rewards efficiency and is pushing into custom and inference silicon of its own, and Marvell is chasing the same hyperscaler contracts. Broadcom argues its designs win on cost and performance per watt — Jalapeño, Tan claims, matches Nvidia's newest silicon on inference at roughly half the cost — but that is a claim the market is still testing, not a delivered fact.

So the selloff is not a verdict against Broadcom. The pattern has repeated twice now — a June drop of about 12% after management held long-term guidance static, and this one on a whisper-number miss — and both times the underlying revenue kept compounding. In that sense a selloff on a strong report is closer to the market's expectations catching their breath than to the company failing. The genuine tension runs the other way: Broadcom is only near-20-times-cheap if the inference turn actually lands on custom silicon and if the gigawatt commitments ship, and neither is guaranteed. Demand is not the issue. The issue is whether the supply commitments and the shape of the next return curve still justify the price today. That is the question worth holding, whatever the next quarter does.