Broadcom Inc. (NASDAQ:AVGO) is selling AI chips faster than ever. But the company just revealed an unexpected cost of that success: memory is squeezing its margins.
The world’s second biggest semiconductor company told investors on Wednesday that AI semiconductor revenue surged 221% year over year to $16.7 billion in the third quarter. Total revenue jumped 86% to a record $29.6 billion.
Broadcom has line of sight to roughly $350 billion of artificial intelligence semiconductor revenue across fiscal 2027 and fiscal 2028.
The stock fell anyway.
The reason sits one line below the revenue guidance.
For the fourth quarter ending Nov. 1, Broadcom guided consolidated gross margin to approximately 73%, down from 78% in the same quarter a year ago.
The reason is not weaker demand. It is the growing amount of memory required inside Broadcom’s custom AI accelerators. That creates a new problem for shareholders.
The faster AI chips become, the more memory they need. And the more memory they need, the harder it becomes for Broadcom to preserve its exceptionally high gross margins.
Broadcom’s Quarter Was a Record, the Margin Line Was Not
- Revenue of $29.6 billion, up 86% year over year, against consensus of $29.36 billion, according to LSEG
- Adjusted earnings of $3.32 per share, against $3.24 expected
- AI semiconductor revenue of $16.7 billion, up 221% year over year and up 54% sequentially
- Operating income of $20.1 billion, a record, at 67.9% of revenue
- Free cash flow of $13.7 billion, or 46% of revenue, the same conversion rate as the prior quarter
- Gross margin of 75%, down 210 basis points sequentially and 10 basis points better than the company’s own 74% guide
Trace that last number backward and the direction is the story.
Gross margin ran at 77.1% in the second quarter, 75% in the third and lands near 73% in the fourth on the company’s own forecast.
That is 410 basis points of compression across two quarters, during the fastest revenue growth Broadcom has ever posted.
The Memory Bill Is Coming Due for Broadcom
increasingly powerful chips require more memory and bandwidth.
That changes the economics.
Chief Financial Officer Amie O’Toole made the connection directly during the earnings call. "This reflects the increasing mix of XPUs with their increasing memory content, which is diluting our consolidated gross margin," she said.
The timing matters.
Conventional DRAM contract prices rose 90% to 95% quarter over quarter in the first quarter of 2026, the steepest single-quarter move the research firm TrendForce has recorded, with NAND up 55% to 60% over the same stretch.
TrendForce has projected DRAM pricing to climb more than 70% across the full calendar year.
Three companies – SK Hynix Inc. (NASDAQ:SKHY), Micron Technology Inc. (NASDAQ:MU) and Samsung Electronics Co. Ltd. (OTC:SSNLF) – control over 95% of DRAM output and all three have tilted wafer capacity toward high-bandwidth memory, which consumes several times the capacity per bit of ordinary DRAM.
Broadcom is on the paying side of that trade.
The Real Constraint Is Moving Down the Supply Chain
There is a bigger implication hidden inside Broadcom’s comments. The AI industry is no longer simply asking whether companies want to build more data centers. It is asking whether the physical supply chain can support them.
CEO Hock Tan said Broadcom considers the availability of leading-edge wafers, substrates and HBM memory when building its AI revenue forecasts.
The company is even building additional substrate capacity in Singapore to address part of the bottleneck. That tells investors something important. AI infrastructure is becoming a system-level investment story.
Chips are only one piece. Memory, substrates, networking, optical components and power infrastructure all need to scale together.
Broadcom is already seeing shortages elsewhere in that chain. Memory is therefore not an isolated issue.
It is one symptom of a much broader problem: AI infrastructure is becoming increasingly expensive and increasingly difficult to build.
Photo: Shutterstock
Login to comment