Broadcom Posts Record AI Chip Revenue, Then Watches Its Stock Drop 12%

Broadcom’s fiscal second quarter of 2026 tells two stories simultaneously. The first is a triumph: $10.8 billion in artificial intelligence semiconductor revenue, up 143% year-over-year, setting a new record and confirming the company’s position as one of the primary beneficiaries of the AI compute buildout. The second is a market disappointment: Q3 AI chip guidance of $16 billion fell short of analyst expectations of $17.2 billion, sending the stock down 12% in after-hours trading despite the record results.

The contrast captures something essential about the current moment in AI infrastructure — expectations have run so far ahead of reality that even extraordinary growth can be perceived as a miss.

Record Results, Complex Reception

Broadcom reported Q2 consolidated revenue of $22.19 billion, up 48% year-over-year and slightly ahead of Wall Street estimates. Adjusted earnings per share came in at $2.44. In virtually any other sector of the economy, these would be unambiguously excellent results.

But in AI infrastructure, where Broadcom competes alongside Nvidia and a growing cohort of custom chip designers, the bar has shifted. CEO Hock Tan stated that “Q2 semiconductor revenue from AI of $10.8 billion grew 143% year-over-year, above our forecast” — factually true. Yet what investors asked was whether next quarter would continue accelerating. The Q3 guidance implied over 200% growth year-over-year, but still came in $1.2 billion below what analysts had modeled.

The Custom Chip Business Model

Broadcom’s AI chip business is structurally different from Nvidia’s. While Nvidia sells GPU clusters to hyperscalers at high margins, Broadcom designs custom AI accelerators — known as ASICs — tailored to the specific needs of individual customers. The company has identified six core AI chip customers, including Google, Meta, Anthropic, and OpenAI, with two additional hyperscalers disclosed over the past year.

These custom chips offer better performance-per-watt for specific inference or training tasks compared to general-purpose GPUs — a significant advantage at the scale these companies operate. Google’s TPU line, for instance, is built on Broadcom silicon and represents a key competitive differentiator for Google Cloud.

For Broadcom, the trade-off is that revenue is more concentrated and custom design cycles are longer. Each new chip generation requires months of engineering collaboration with the customer before a single die is manufactured. This creates high switching costs — but also means that revenue cadence is lumpier than a product company selling off-the-shelf hardware.

Margin Pressure Emerges

One detail attracting significant attention beyond the top-line numbers was Hock Tan’s candid acknowledgment of margin pressure. “It’s the mix,” Tan told analysts on the earnings call, “between software and non-AI to the very rapidly growing AI semiconductor that is just diluting gross margin.”

This is the tension embedded in Broadcom’s growth story: the AI chips growing fastest are accretive to revenue but dilutive to margins compared to the company’s high-margin software and networking businesses, including its large VMware enterprise software portfolio. As AI semiconductor revenue becomes a larger share of the total, overall gross margins face structural pressure.

This dynamic is not unique to Broadcom. Across the semiconductor industry, the custom ASIC business — which requires extensive non-recurring engineering investment for each new design — carries inherently different economics than a mass-market product. Broadcom has managed this tension skillfully, but the margin question will intensify as AI chip revenue scales toward its stated $100 billion target for fiscal 2027.

The $100 Billion Bet

That target — over $100 billion in AI semiconductor revenue for fiscal year 2027 — remains Broadcom’s north star. The company did not raise it, which some investors interpreted as caution. In context, though, maintaining a commitment to more than doubling AI revenue again from an already elevated base represents substantial ambition.

The six custom chip customers driving this growth are spending at extraordinary rates. Collectively, they represent an increasingly essential slice of the AI inference and training stack — not merely as chip buyers but as companies designing and deploying AI systems at a scale that demands custom silicon. As these customers expand their AI fleets, Broadcom’s ASIC revenue grows almost mechanically with them.

Tan specifically called out Anthropic and OpenAI — both of which recently went public or are preparing IPOs — as among the customers with expanding AI accelerator programs. The implication is that even new entrants into the public markets are committed to long-cycle custom silicon programs, suggesting the AI infrastructure buildout has durability beyond near-term hype cycles.

The Sell-the-News Reaction

The 12% post-earnings decline is best understood as a sell-the-news phenomenon in a stock that had priced in perfection. Broadcom shares had more than doubled over the prior 12 months, anticipating the very growth that Q2 results confirmed. When even 143% AI revenue growth doesn’t produce a guidance raise, momentum investors reappraise their positions.

This is increasingly a pattern across AI infrastructure stocks. Nvidia faced similar dynamics earlier in the cycle. Super Micro Computer and Marvell Technology have also seen strong results punished by guidance that merely met rather than exceeded stretched expectations. The lesson seems to be that once a company becomes widely recognized as an AI beneficiary, its stock price begins to reflect increasingly optimistic future scenarios — making any single quarter’s results a coin flip.

What It Signals for the Chip Market

For the broader semiconductor industry, Broadcom’s results carry important signal: the custom ASIC model is working. The six hyperscaler customers are spending more than ever on specialized AI silicon, suggesting that the AI infrastructure buildout is not decelerating. If anything, the gap between general-purpose GPU demand and custom chip demand may continue to grow as AI systems mature and per-task optimization becomes more valuable.

The guidance miss, while notable, is perhaps better understood as the difficulty of predicting quarterly lumpiness in large custom chip programs than as evidence of structural slowdown. When revenue depends on bespoke multi-billion dollar chip programs for a handful of customers, the timing of tape-outs and production ramps creates inherent volatility.

Broadcom’s challenge heading into fiscal 2027 will be demonstrating that the path to $100 billion in AI chip revenue is credible — while managing the margin evolution that comes with it. If it can do both, the stock’s post-earnings reaction may ultimately look like noise rather than signal.