Cerebras Says Data-Center Capacity Is Its Main AI Growth Constraint

Cerebras’ answer to its post-earnings selloff is that the margin pressure investors focused on comes from a deliberate capacity decision: the company plans to rent back equipment it had already sold to customers so it can meet demand for fast inference. Ed Ludlow opened with the mismatch. Cerebras’ sales outlook beat Wall Street estimates, but the stock was down about 17.5%, which he described as the biggest drop in the company’s short history as a public company. Bloomberg’s intraday chart showed Cerebras shares at 187.18, down 39.54, or 17.44%.

Andrew Feldman argued that the company is ahead of the plan it had shown investors at the start of 2026 and during the public offering process. He cited record revenue of $191 million, up 92% year over year, and said the cloud business was up 167% year over year. Bloomberg’s on-screen results graphic used a slightly different revenue figure, listing first-quarter revenue at $193.4 million versus an estimated $181.4 million, along with a 22-cent loss per share, $110.6 million in hardware revenue, and $82.8 million in cloud revenue.

The margin issue, in Feldman’s account, is concentrated in the second and third quarters. Cerebras told investors it would go back to some customers and rent back gear it had sold them in order to keep up with demand. That would have a gross-margin impact “on the order of 10 or 15 points,” he said. The reason was to keep customers close and satisfy what he called “extraordinary demand” for fast inference.

That was Feldman’s central rebuttal to the market reaction. He said Cerebras beat margin consensus “substantially” and guided full-year gross margins 10% better than planned. Later, he described the full-year guidance as 10 gross margin points above consensus. His position was that the company is ahead of plan on the disclosed metrics, even though the market is penalizing the temporary margin cost of securing capacity.

For Andrew Feldman, the largest constraint is not demand and not, primarily, the chips themselves. It is data-center capacity. He said IPO proceeds had helped Cerebras significantly expand its data-center pipeline and pointed to a partnership with Bell Canada for 120 megawatts, to be delivered in 2027.

Cerebras is pursuing data centers across the US, Canada, Europe, and the Middle East. Feldman argued that the “vast resources” now available to the company give it an advantage in chasing the limiting factor: data centers.

Ed Ludlow pressed on the physical side of that claim. The issue is not just compute supply, he said, but buildings, planning approval, concrete, labor, and construction. Feldman called that “the irony of this market”: AI is moving at “blistering speed,” while the industry is constrained by data centers that “move with the speed of real estate.” He described the problem as category-wide, affecting hyperscalers, neoclouds, and newer cloud providers alike.

Cerebras is therefore not presenting the bottleneck as a lack of demand for its systems. Nor is the company’s near-term explanation mainly a semiconductor shortage. The constraint, as Feldman described it, is the pace at which powered, permitted, usable data-center capacity can be brought online.

Cerebras’ technical claim is that its wafer-scale architecture insulates it from several supply constraints affecting other AI compute providers. Andrew Feldman said Cerebras does not use high-bandwidth memory, or HBM. He described HBM as a type of DRAM made by three companies — Micron, Hynix, and Samsung — and said it is globally short, expensive, and subject to long lead times. Because Cerebras does not use it, he argued, the company has “a tremendous advantage.”

He listed two additional constraints Cerebras says it avoids. The first is CoWoS, a TSMC process used by many advanced AI chips. Cerebras does not use it, Feldman said. The second is capacity at TSMC’s three-nanometer node. Cerebras, he said, is on the five-nanometer node instead.

Feldman tied those avoided bottlenecks to performance as well as availability. Cerebras’ architecture, he said, lets the company deliver “the fastest inference in the world by an order of magnitude” while avoiding the main supply-chain constraints faced by others in the field.

Ludlow pressed on whether that advantage had translated into actual customer delivery, not just sales claims. Feldman pointed to OpenAI: Cerebras signed a contract on December 24 and had OpenAI in full production on February 1. He called that “unheard of speed.” Ludlow’s response was simpler: “That’s quick.”

The claim is not that Cerebras faces no bottlenecks. Feldman’s argument is narrower: where other providers may be constrained by HBM, advanced packaging, or leading-edge node capacity, Cerebras’ architecture shifts the pressure toward deployment capacity — buildings, power, and the real-estate pace of data centers.

OpenAI’s Jalapeno intelligence processor fits the market structure Andrew Feldman said Cerebras has expected all along. The AI compute market, in his view, is too large to consolidate around GPUs. It will be served by a heterogeneous mix of hardware architectures.

That means ASICs from hyperscalers, ASICs from labs, and companies such as Cerebras with “pioneering architectures.” Feldman did not argue that Cerebras would own the market. His expectation is that multiple architectures will each take “big bites” out of a market large enough to support them.

The reason is the breadth of demand AI creates. Feldman said AI makes “tractable for compute” much of the world around us in a way that was not true before AI. In that context, custom silicon from large labs is not a sign that the market is closing; it is part of the heterogeneity Cerebras expects.

Asked whether Cerebras might return to capital markets to move faster, Andrew Feldman said the company has more than $9 billion on the balance sheet and is “really pleased” with that position. He did not rule out further financing, saying Cerebras is always scanning equity and debt markets for ways to accelerate growth.

The milestone he offered investors was execution rather than a single product target. When a company lays out an aggressive plan and “crushes it,” he said, that demonstrates both its ability to execute and its ability to predict its own execution. Cerebras, he said, needs to keep setting high bars and beating them.