The Biggest Threat to Broadcom's AI Bet Is Its Own Coming Top Customer

Broadcom shares fell about 4.8% on Monday, close to $344.72, after Anthropic CEO Dario Amodei published a weekend essay calling for a global slowdown in the pace of AI development. Then, in after-hours trading, Broadcom CEO Hock Tan answered the worry directly on CNBC, and the stock nudged back up. The exchange is worth reading closely, because it is a collision between the two halves of Broadcom's entire investment story — and the man urging caution is about to become the chipmaker's largest customer.

What Broadcom actually sells in AI

Broadcom's AI business is not like Nvidia's. Instead of selling one standardized accelerator to everyone, Broadcom co-designs custom chips — it calls them XPUs — for a handful of hyperscalers and model labs, then sells them in enormous volume. Its customers read like a who's who of the frontier-model race: Google (its custom TPUs), OpenAI (the new "Jalapeño" accelerator), Meta, and Anthropic.

This is why the numbers are so large. In the fiscal third quarter reported on September 2, Broadcom's AI semiconductor revenue hit $16.7 billion, up 221% from a year earlier and now 56% of total revenue. Management guides to about $21.7 billion of AI revenue next quarter, and lays out a roadmap of $58 billion this fiscal year, roughly $115 billion in fiscal 2027, and roughly $230 billion in fiscal 2028 — a doubling each year, built on six custom-chip customers it says wants even more than the roadmap shows.

That roadmap is the whole bull case. Broadcom is priced not as a diversified chip and software company but as a toll road on frontier-AI infrastructure, with a market cap near $1.6 trillion and a growth multiple that leaves little room for the traffic to slow.

Why one essay spooked the market

The market's fear was never hard to name. Broadcom's revenue permission comes from one thing: that the frontier-model arms race keeps buying compute. If the labs that pay Broadcom slow down, the toll road loses traffic.

Amodei's essay lands precisely on that assumption. He is not a bystander lecturing from outside; Anthropic is projected to surpass Google as Broadcom's largest custom-chip customer in 2027 and hold that through 2028. The person calling for the industry to hit the brakes is, per Broadcom's own roadmap, the person expected to feed it the most.

The market read it as a threat to the permission itself, and it sold the whole chain: Broadcom dropped about 4.8% and a broad semiconductor ETF fell about 5.6%. Even though Hock Tan had just delivered a quarter that beat estimates and more than tripled AI revenue, the stock had already sold off at the September report on guidance that came in below aggressive expectations. The essay gave worried holders a second reason to exit.

Tan's rebuttal, and what it does and doesn't prove

Tan's answer in the after-hours interview was threefold, and each part separates what he actually controls from what he is betting on.

First, he affirmed the targets. Asked whether the slowdown debate had changed the fiscal 2027 or 2028 forecasts, he said, "No, not in the least."

Second, he reframed the debate by splitting training from inference. Tan said he cannot speak to training, but on inference — the day-to-day running of trained models — he called demand "very, very strong." That matters: a model lab can pause building a bigger next model and still need enormous computing capacity to serve the models it already has to paying users.

Third, he argued the constraint is supply, not demand. Broadcom has said actual customer demand for fiscal 2027 already exceeds the $115 billion target, and that growth is capped by fabrication and packaging capacity — TSMC's CoWoS advanced packaging is sold out deep into 2027, and AI chips absorbed about 90% of global high-bandwidth memory supply in 2025. If that is true, orders are not the risk; physical capacity is.

Take the three together and the rebuttal is coherent. Take them exactly as they are written and the after-hours bounce of about 0.4% stays sentiment, not proof. Supply being the constraint tells you Broadcom's near-term numbers are safe only if the contracts hold. And the newest, fastest-growing legs of the roadmap — Anthropic and OpenAI — are exactly the customers asking for, or vulnerable to being pulled into, a slower race.

The honest reading

The scorecard here is a good quarter against a higher hurdle. The reported AI numbers beat; the forward guide came in under the most optimistic estimates; and the headline about slowing AI rattled a stock that had already given back ground from its highs. Monday's drop plus the modest after-hours recovery is a market arguing with itself about pace, not a fundamental fact.

What would break the thesis is not another essay. It is a customer, especially Anthropic, cutting XPU orders, or a model lab deciding not to build the next generation of capacity. Until then, watch whether fiscal 2027 revenue tracks toward the $115 billion mark, and whether Q4's guided $21.7 billion in AI revenue actually converts. The essay is a risk to the most aggressive years of the plan, not to the quarter already in the bank. That is the difference between a company you can own through the noise and one you should reconsider if the next order book wobbles.