Broadcom's 25% Selloff Decomposes to One Google Question That September 2 Answers

Broadcom has lost roughly a quarter of its value since its June peak near $495, and the final leg of that slide came the way panic usually does: with a headline. On August 19, Marvell disclosed an expanded custom-chip agreement with Google that included a warrant tied to up to $120 billion of future purchases, and Broadcom — which designs the TPU chips at the heart of Google's AI push — fell about 5% that day. A week later, with the stock near $370 and down mid-teens in two weeks, Mizuho's Jordan Klein called it capitulation and argued for getting in before the September 2 earnings report, noting the company has now gone nearly 30 straight quarterly reports without two disappointing quarters in a row.

The question behind that call — is this a real breakdown in Broadcom's most important customer relationship, or fear priced past the facts — is answerable. The report is scheduled to settle it.

Start by breaking the company apart, because one customer now matters so much. In the quarter ended in early May, Broadcom reported $22.2 billion of revenue, up 48%, and that headline splits into two very different businesses. Custom AI accelerators and AI networking chips — the pieces built for hyperscalers — brought in $10.8 billion, up 143% from a year earlier. The other half is mature semiconductors plus infrastructure software led by VMware, which produced $7.18 billion of that revenue, grew only 9%, and came in below what analysts expected. The read: Broadcom is today roughly half an AI-compute company growing at triple-digit rates, partly funded by a software annuity with roughly 93% gross margins whose growth is decelerating.

That mix matters when you evaluate the Google fear. The bear case, most concretely stated by Macquarie when it cut Broadcom to Neutral, is that Google is adding silicon suppliers and moving work in-house — that Broadcom's share of Google's custom-chip revenue erodes from roughly 95% today to 65% by 2028, with MediaTek taking a next-generation TPU variant. Read that number one way and it is a nightmare. Read it the way a share calculation should be read, and the story is different: a falling share of a much larger pool is not a falling business.

Two facts anchor that reading. First, Google has not replaced Broadcom — in April it signed a long-term agreement to develop future TPU generations and supply AI rack components through 2031. Second, the $120 billion warrant that spooked investors is itself evidence of how enormous Google's custom-chip purchasing is going to be. Hyperscalers add second vendors because they are scaling compute faster than any single design house can supply — dual-sourcing is a demand signal, not an exit. The closest precedent cuts the same way: when Apple moved its Wi-Fi and Bluetooth chip in-house, the report surfaced in January 2023 and Apple's own chip arrived in the iPhone 17 in September 2025 — an Apple that had been roughly 20% of Broadcom's revenue. Design-aways take years, and the bear view Mizuho describes — that nothing management says on the call can refute the share-loss thesis — is a sentiment position, not a math position.

The non-Google backlog is the rest of the argument. Broadcom is developing custom accelerators for six hyperscale customers with supply agreements stretching into 2028 — Meta's initial gigawatt-order for its MTIA accelerators, Anthropic scaling past a gigawatt this year toward several gigawatts in 2027, OpenAI's first custom XPUs reaching volume production in 2027. And the bookings line is where the company shows its hand: in the second quarter it booked more than $30 billion of AI orders while shipping $10.8 billion, and it extended revenue visibility from 2027 into 2028.

Now the part that should make a calm buyer hesitate, because the selloff was not irrational. In June, Broadcom delivered the strongest AI numbers in its history — and the stock dropped more than 12% the next session. The cause was guidance, not results: management reaffirmed its $100 billion fiscal 2027 AI revenue target without raising it, and a market already paying roughly 31 times forward earnings at a record price had been waiting for an upgrade. That is the pattern that governs this stock now. Broadcom no longer trades on what it ships; it trades on what its guidance promises. At the current multiple — near the low end of its 2026 range, but still a rich price — the same reaffirmation again is what produced June's drop, and two structural currents pull against the optimism: the shift toward faster-growing AI chips pushes consolidated gross margin down on mix alone (guided to about 74% in the current quarter), while the high-margin software engine that funds the roadmap keeps decelerating. Recent order-flow data adds a sour note typical of a crowded unwind: net selling even in retail-sized orders.

So the September 2 report has a pre-committed bar. The quarter itself is largely known — management already guided to about $29.4 billion of revenue with AI semiconductor revenue around $16 billion, up more than 200%. The number that decides the stock is the fiscal 2027 AI revenue figure. If Hock Tan raises it above $100 billion, the capitulation pricing of the past two weeks gets repriced — the "confront the shorts" move Mizuho is explicitly banking on. If it is again a reaffirmation, investors have an experiment from June telling them what that does to the share price. Mizuho's own framing is the honest calibrator: this is not a call for a sudden 25% rally next week, but a six-months-plus view that the market has priced the worst Google outcome while the company keeps booking three months of shipments in a single quarter. The entire bet is that asymmetry — $30 billion booked against a guidance number the market wants raised — and it is checkable in a single call on Wednesday.