Cerebras, in its first earnings report since going public, said it will rent its own AI hardware back from a customer to bring more compute online, and guided full year gross margin to 38–41% from about 47% last quarter.
Cerebras, the AI chipmaker, told Wall Street on Wednesday that it will start renting its own AI hardware back from a large customer so the company can bring more compute capacity online faster than its own data-center buildout would allow. Investors read that move in the margin line, sent the stock down roughly 14% to 20% intraday, and pushed it close to its IPO price. The selloff was not about demand. It was about a sale-leaseback dressed up as a margin guide.
The mechanics matter. Cerebras sells its wafer-scale AI systems to a customer, then leases the same hardware back so Cerebras can route that compute to other buyers while its own facilities catch up. On the company's first quarterly print since going public, revenue came in at $193 million, up 94% year over year, and the net loss narrowed to $14 million from $23.9 million a year earlier. None of those numbers justified a crash. What did was the guide: full-year gross margin of 38% to 41%, against roughly 47% in Q1, because the rent-back transfers sale economics into lease expense.
The framing from Cerebras is that this is a temporary, capacity-driven choice. Investors read it as a margin story they were not willing to forgive in a debut quarter. Both can be true. The move turns what looks like a one-quarter optics problem into a structural question: when AI compute is scarce enough that buyers become quasi-financiers, the accounting starts to do work the underlying demand never asked it to do.
Cerebras did not name the customer in its earnings press release or on the call. That omission matters. A rent-back to a strategic hyperscaler reads very differently from a rent-back to a financial buyer, and the market's interpretation depends on which one it is. CEO Andrew Feldman, in a CNBC interview after the print, told investors they had "misunderstood" the margin guide. That is a characterization, not a consensus. It is worth reporting because Feldman said it, not because it settles the question.
There is also no prior public-quarterly comparable to lean on. This was Cerebras's first earnings report since its IPO, so any "beat" or "miss" framing rests on prospectus-era or sell-side estimates rather than on the company's own historical guidance. Yahoo Finance and Blockonomi pegged the move closer to 14%, TechCrunch closer to 20%. What the intraday drop confirmed matters more than the exact number: a public market that is willing to pay for AI hardware scarcity but is not yet willing to accept a sale-leaseback as the price of that scarcity on a single quarter's earnings.
The next test is whether the rent-back stays small and short. If the next quarter lets Cerebras book higher-margin system sales again as its data centers come online, the full-year guide resets quickly and the "misunderstood" framing earns some retrospective weight. If the company has to extend the arrangement, or sign a second one, the 38% to 41% range becomes a ceiling rather than a temporary trough, and the rent-back stops being a story about one company and starts being a template for how AI compute is going to be financed through the rest of this cycle.