Broadcom's AI chip revenue tripled in a year. Investors sold the stock anyway

Broadcom Q3 2026 AI chip revenue up 221 percent stock falls on guidance

Broadcom's AI chip revenue tripled in a year. Investors sold the stock anyway.

Q3 FY2026 Results

Broadcom's Earnings Call

$16.7B

AI chip revenue, up 221% year over year

$29.6B

Total quarterly revenue, up 86% year over year

$21.7B

Q4 AI revenue guidance, below the $22B-$23B analysts modeled

Confirmed custom chip hyperscale customers

Published: September 5, 2026 | Updated: September 5, 2026 | Category: Business | By Mahesh

Broadcom reported third-quarter fiscal 2026 revenue of $29.6 billion on September 2, an 86% increase from the same period a year earlier, with AI-specific semiconductor revenue hitting $16.7 billion, up 221% year over year, driven by custom accelerator chips built for Google, Meta and OpenAI, according to Quartz's reporting on the results.^[1] By almost any conventional measure, this was a blowout quarter. The stock fell more than 3% in extended trading anyway.

The disconnect traces to a single number: guidance. CEO Hock Tan told analysts on the earnings call that the company expects fourth-quarter AI semiconductor revenue of approximately $21.7 billion, according to Tech Times' detailed account of the call.^[2] Analysts had modeled Q4 AI revenue in the range of $22 billion to $23 billion, based on Broadcom's own prior disclosure of $30 billion in Q2 AI bookings. That gap between guidance and expectation, even though the $21.7 billion figure itself represents a 236% year-over-year acceleration, was enough to trigger the sell-off.

A Company Whose Own Success Has Become the Standard It's Measured Against

Tech Times' reporting flagged a pattern worth naming explicitly, because it explains why a genuinely record quarter produced a negative market reaction. The exact same dynamic played out after Broadcom's second-quarter fiscal 2026 results in June, when the stock fell approximately 13% despite AI chip revenue climbing 143% year over year, because CEO Hock Tan did not raise the company's full-year AI semiconductor revenue target beyond $100 billion at that time, according to CNBC's coverage of that earlier report.^[3] Two consecutive quarters now show the identical shape: extraordinary year-over-year growth, paired with forward guidance that falls just short of what Wall Street had already priced in. Tech Times' own framing of this pattern was direct: Broadcom's "own extraordinary numbers have become the floor beneath investor expectations, not the ceiling."

That dynamic is a specific and increasingly common feature of how the market is treating AI infrastructure stocks broadly in 2026, not a Broadcom-specific issue. When a company's growth rate is already triple digits, the market's baseline expectation shifts to assume that pace continues or accelerates, and even modestly slower sequential growth relative to the prior quarter's trajectory gets priced as disappointment rather than as the continuation of exceptional performance it objectively still represents.

What Tan Actually Said About Broadcom's Customer Base

The earnings call itself offered a level of specificity about Broadcom's custom chip business that is worth reading directly rather than through headline summary. According to Tech Times' reporting, Tan confirmed Broadcom now co-designs custom AI accelerators for six confirmed hyperscale customers: Google, on its seventh generation of Tensor Processing Units with the current Ironwood chip; Meta Platforms, through its Meta Training and Inference Accelerator program; OpenAI, whose custom Jalapeño inference chip was unveiled in June 2026 and is targeted for deployment by the end of this calendar year; Anthropic; ByteDance; and Fujitsu. Combined, Tan said Broadcom now holds roughly 70% of the custom AI accelerator market, alongside Marvell Technology as its sole meaningful rival in that specific category, a claim Depth Grid can contextualize against Marvell's own disclosed customer economics.

That competitive framing connects directly to reporting Depth Grid covered on Google's warrant arrangement with Marvell last month, where Marvell disclosed a custom silicon agreement structured around $500 million purchase tranches through fiscal 2033. Quartz's own reporting on this week's Broadcom results noted the connection directly: "Marvell last month closed a custom chip partnership with Google under which the search giant could contribute $120 billion in revenue through fiscal 2033." Reading both companies' disclosures side by side confirms that Google is deliberately splitting its custom silicon business between two suppliers rather than concentrating it with either one exclusively, a detail Tan himself acknowledged separately in an earlier quarter, according to Yahoo Finance's coverage of the June results, when he told analysts Google would likely continue drawing on multiple chip suppliers.

"We are empowering two of our most strategic customers, the leading AI labs, to bridge the gap [between training and inference infrastructure needs]." — Broadcom CEO Hock Tan, on the Q3 FY2026 earnings call, September 2, 2026^[4]

The Margin Story Underneath the Revenue Story

Beyond the headline revenue figures, Tan's commentary on the call touched a structural tension that has followed Broadcom's AI business for several quarters now. Yahoo Finance's earlier coverage of Broadcom's June results captured Tan explaining the mechanism directly: "Semiconductor margins remain very stable and very solid. It's the mix, particularly the mix between software and non-AI to the very, very rapidly growing AI semiconductor that is just diluting gross margin." In plain terms, Broadcom's AI chip business, while growing explosively, carries structurally different margin characteristics than the company's legacy software and non-AI semiconductor lines, meaning the faster AI revenue grows as a share of the total, the more it pulls down Broadcom's blended gross margin, even though every individual business line remains healthy on its own terms.

Broadcom's board declared a quarterly dividend of $0.65 per share alongside the results, according to Quartz's reporting, and the company set fourth-quarter non-GAAP operating income guidance at approximately 66% of projected revenue. Those figures suggest the underlying business remains highly profitable even as the specific gross margin mix shifts, a distinction that matters for separating a genuine profitability concern from what is, more precisely, a change in the composition of an already strong and growing profit base.

Reading Broadcom's Guidance Against the Rest of the AI Infrastructure Sector

Broadcom's earnings this week landed inside a broader wave of AI infrastructure disclosures Depth Grid has tracked closely over recent weeks. Broadcom's own full-year AI semiconductor revenue guidance remains "in excess of $100 billion," a figure the company has held steady across multiple quarters even as individual quarterly results have consistently beaten expectations. That steadiness itself is a signal worth noting: rather than repeatedly raising a headline annual target only to fall short later, Broadcom appears to be managing investor expectations conservatively at the full-year level while its actual quarterly execution continues to run ahead of what that conservative full-year framing would imply, a pattern that stands in some contrast to OpenAI's own recent advertising revenue disclosure, where the company's actual run rate fell short of a specific dollar target it had previously stated publicly.

The stock market's reaction to Broadcom's results this week, alongside the identical pattern in June, suggests investors are increasingly parsing AI infrastructure earnings on a much finer margin than the simple question of whether growth is happening. The question being priced now is whether growth is accelerating or merely continuing at an already extraordinary pace, a distinction that barely registered a year ago when any meaningful AI-related revenue growth was treated as unambiguously positive news, but has become the central axis on which these stocks now trade.