Broadcom Keeps Beating and Falling. The Headlines Blame the Wrong Things.

On June 3, Broadcom told Wall Street its custom AI chip business sold $10.8 billion in a single quarter, up 143% from a year earlier. Bookings for the quarter ran near $30 billion — roughly three times what it shipped. It beat every estimate. The stock fell more than 12% the next day and on the order of $300 billion of market value disappeared.

It repeated the move in September. Third-quarter revenue rose 86% to $29.59 billion. Non-GAAP earnings rose 96%. AI semiconductor revenue hit $16.7 billion, up 221%. Shares still slid in extended trading, because the forward-quarter revenue guide came in a few hundred million dollars light of consensus.

Broadcom has reached the point where only the gap between guidance and the market's imagination moves the price. A company growing 86% a year that beats, then raises its 2027 AI target to roughly $115 billion, and still gets sold, is being priced for absolute perfection. That reflex — not the antitrust case, not the insider selling — is the real signal. The headlines have been pointing at the wrong layer.

The two distractions

The news cycle around Broadcom runs on two stories. The European Union is circling the VMware acquisition, asking European cloud providers how dependent they are on VMware and whether they can easily switch; Broadcom already lost a court bid in August to shield U.S. documents from the probe. And the founder-CEO, Hock Tan, sold roughly 70,000 shares in early January at just under $347 a share.

Both are real. Neither is where the money is made or lost. VMware is the infrastructure-software half of the company, and its quarter tells the tale: that segment grew only about 9% in the spring while the rest of the business grew 48%. The AI-silicon engine prints the growth; VMware is the slow, regulatory-adjacent layer. The insider sales are a fraction of a stake a founder has spent a generation building — scheduled, disclosed, and small enough to be noise dressed as a signal. When investors need a story for a roughly $1.8 trillion stock, these grab the slot because they are legible.

The scarce layer under the glut

Every investor has heard the 2026 worry: AI compute is flooding in, GPUs everywhere, capex peaking, a glut coming. The numbers inside Broadcom contradict it. A quarter where bookings ($30 billion) ran roughly three times shipments ($10.8 billion) is not an industry with too much supply; it is an industry where buyers are fighting for allocation.

That is the scarcity story, and it has a sharp shape. As generic compute becomes abundant, what becomes scarce is custom silicon — accelerators designed to one customer's specifications, fused by years of co-design and advanced packaging. Broadcom makes almost nothing general-purpose anymore. It designs accelerators for a short, named list of AI giants: Google's TPUs under a supply agreement running to 2031, Meta's MTIA in a multi-generation deal, and a lengthening roster that runs to OpenAI, Anthropic, and ByteDance. The scarcity is not in the chips it sells; it is in the ability to design and package them, locked to a handful of buyers who can pay.

The asymmetry here is real, and conventional wisdom has it backwards. The risk is not that AI demand collapses into a GPU glut — against a booked backlog in the tens of billions, that is the wrong side of the trade. The risk is the reverse: a short list of customers, each a hyperscale giant with its own silicon ambitions.

What's actually priced in

A market cap around $1.8 trillion with a forward multiple in the low 30s is not the frothiest valuation in the trillion-dollar club, but multiples on hypergrowth are promises. The promise here is specific: keep the named customers writing eight-figure-per-quarter checks while adding more. When one wobbles — or concludes the in-house chip is better built elsewhere — the trajectory cracks in a way no antitrust fine or share sale ever will.

Which is why the June selloff was the informative one. Investors did not sell because $10.8 billion was weak. They sold because Hock Tan declined to raise an already extraordinary guide. At that bar, flawless execution was the base case, and only overdelivery counted. That is not the profile of a cheap, misunderstood stock. It is the profile of a great business whose reward for continuing to be great has already been banked.

Separate the layers and the decision stops being complicated. The EU and VMware are real but peripheral: a regulatory tax on a legacy business, not a threat to the engine. The insider sales are a rounding error against a founder's stake. The valuation is demanding but not absurd for the growth.

The only variable that matters is the short list of buyers. Watch whether Google, Meta, OpenAI, and the rest keep raising orders, and whether the pipeline keeps landing new names. On that, the evidence is unusually clean: bookings running far above shipments, a backlog in the tens of billions, and a CEO who keeps raising a number he previously defended as a ceiling. Read the headlines as weather. Price the customer list as the trade.