Wall Street Sold AI Chips on Cost Fears. This Company Gets Paid When Big Tech Cuts Costs.

Research Desk · July 25, 2026

Marvell’s headquarters campus in Santa Clara, California. Photo via Wikimedia Commons. [Confirm author and license on the file page, then credit here.]

The first two weeks of July erased more than $1.3 trillion of semiconductor market value. Micron fell 13% in a single session, roughly $138 billion gone in a few hours. Intel dropped 9%, Advanced Micro Devices 7%, and the VanEck Semiconductor ETF shed 5% after posting a record 71% gain in the second quarter. The triggers arrived together: SK Hynix slowed its next-generation memory expansion, new Federal Reserve Chairman Kevin Warsh signaled that nine of eighteen policymakers now want higher rates this year, and Meta suggested it had surplus AI capacity worth renting out. Underneath all of it sat one question. Will the enormous sums spent on artificial-intelligence infrastructure ever earn a return?

Marvell Technology (NASDAQ: MRVL) fell with the group, down roughly a third in a month and about 36% from its June high. But it is worth reading the fear precisely, because the market’s specific worry is that the technology giants will get disciplined about what AI costs them. And the single most concrete thing a hyperscaler does when it decides to cut AI cost is stop buying general-purpose accelerators at Nvidia’s margins and design a chip of its own instead. That decision is Marvell’s entire business.

The Rail Line, Not the Taxi

A graphics processor is a taxi. It will take you anywhere, it is available on demand, and you pay a premium for that flexibility every single trip. That is a fine arrangement when your journeys are varied and unpredictable. It becomes an expensive habit when you make the exact same trip several million times a day. At that point the sensible move is to lay a rail line between the two points you actually travel between. The rail line costs a fortune up front, it is useless for any other journey, and per rider it is dramatically cheaper.

Marvell is the firm that engineers the rail line. Amazon’s Trainium processors, the silicon behind AWS training workloads, are custom chips Marvell helped design. Microsoft’s Maia accelerators for Azure run on the same model. Meta has custom AI silicon built with Marvell’s involvement, and Google is a partner on its Axion processor work. These are application-specific integrated circuits: less flexible than a GPU, far more efficient at the one calculation the customer runs constantly. When a company is spending billions on compute, shaving twenty percent off the power bill is not a rounding error. The second half of the business is the signaling system that lets the trains run seconds apart, the optical interconnect and switching silicon that allows thousands of chips to behave as a single machine. Every accelerator in a modern data center, custom or otherwise, needs that plumbing to be useful.

What the Order Book Actually Says

The most recent quarter, reported in late May, was solid rather than spectacular, and that is the point. Revenue came in at $2.418 billion, up 27.6% year over year, with non-GAAP earnings of $0.80 per share in line with estimates. Free cash flow jumped 127% to $483 million, and operating cash flow rose 92% to $639 million. Data center is now roughly 76% of the company, up from a business that used to be dominated by storage controllers and carrier equipment.

The guide carried the real information. Management pointed to current-quarter revenue near $2.70 billion, roughly 35% growth and above the $2.62 billion the Street expected, with earnings of about $0.93 against a $0.90 consensus. Full-year guidance was raised toward $11 billion, more than 30% above fiscal 2026’s record $8.195 billion. Chief executive Matt Murphy told investors he expects revenue growth to keep accelerating each quarter through the year. Custom silicon alone generated $1.5 billion last fiscal year and is guided to grow more than 20%. Behind that, Nvidia completed a $2 billion investment agreement in March, and acquisitions of XConn Technologies, Celestial AI and Polariton have filled out the interconnect portfolio alongside a new 102.4-terabit switch chip. None of those numbers changed in July. The share price did.

The Margin Problem Nobody Puts on the Slide

Honesty requires naming the blemish, and here it is a real one. Gross margin now runs around 51%, down from the roughly 60% Marvell used to earn, because custom silicon is a thinner-margin business than the legacy products it is replacing. Operating margin in the first quarter was 14%, flat against the same quarter a year earlier, even with revenue up 28%. Read that twice. The top line is compounding at nearly thirty percent while the profit on each incremental dollar is not improving at all. Growth of this kind has to come through sheer volume, and volume depends on a short list of customers continuing to place very large orders. The company’s newly appointed chief financial officer also made his first open-market stock sale since joining, which several observers noticed.

Cheaper, Not Cheap

At roughly $211 the stock carries a market capitalization near $207 billion and trades around 55 times the consensus earnings estimate for this fiscal year, close to $3.80 per share. A 36% drawdown has made Marvell considerably cheaper than it was five weeks ago without making it cheap. Coverage is wide and warm: about 82% of the 49 analysts following the company rate it a buy, with targets clustering in the $230s, a high near $400 and a low at $180.

The bear case starts with the same concentration that makes the bull case work. Data center is 76% of revenue, custom silicon programs lean on a handful of buyers, and a hyperscaler can re-bid a socket or pull the design in house when a contract comes up. If growth decelerates and the multiple compresses toward 30 times earnings, a level that would still be generous for a semiconductor company, the math lands near $115. That is roughly where these shares traded in the spring, and it is the bear case with a number attached rather than a vague warning about valuation.

The Bottom Line

The fear that knocked a third off this stock in July was about memory pricing, interest rates, and whether AI spending pays for itself. Marvell’s exposure to the first two is indirect. Its exposure to the third cuts in an unusual direction, because if the giants do get religion about what AI costs them, custom silicon is the instrument they reach for, and Marvell is who they call. What sold off was the multiple, not the order book. The question for investors is whether 55 times forward earnings is a fair price for growth that is genuine, margin-dilutive, and dependent on a few very large customers staying committed. The next piece of evidence arrives with second-quarter results on August 26.