(Corrects paragraph 1 to drop reference to company missing revenue estimates following a revision of comparable revenue number by LSEG; adds core revenue in the last paragraph; the error appeared in a previous version of the story)
By Anhata Rooprai and Max A. Cherney
Aug 12 (Reuters) - AI chip designer Cerebras Systems slumped 16% in extended trading on Wednesday after its quarterly results, suggesting that Wall Street was ready to punish richly valued AI stocks if they slipped up on any key metric.
Shares of the company have gained more than 41% from its IPO price, fueled by optimism around demand for its specialized chips that it has pitched will rival market leader Nvidia's pricey processors. They closed up 11.6% in regular trading.
This is Cerebras' second quarter reporting results as a public company, and the post-market plunge shows that investors are still trying to rightly value a firm that is under pressure to prove its business can scale profitably.
Any sign of slowing customer adoption could deepen doubts about Cerebras' ability to gain share from Nvidia, which has stepped up efforts to capture the fast-growing AI inference market with technology licensed from startup Groq.
Cerebras' adjusted gross margin in the second-quarter ended June 30 was lower by about 5 percentage points because it had incurred higher costs to rent computing capacity that it previously deployed at other customers, finance chief Bob Komin said on a post-earnings conference call, even as the company expects to increase manufacturing capacity by more than 10-fold this year.
Second-quarter adjusted gross margin was 40.6%, down from 46.5% in the preceding three months.
Still, Cerebras raised its annual revenue and gross margin forecasts and said it was not experiencing supply shortages as severe as other chip designers.
Cerebras' flagship wafer-scale engine is a single chip the size of a dinner plate containing trillions of transistors, a design that it says is more efficient than connecting thousands of smaller graphics processors together, as Nvidia does.
The Sunnyvale, California-based company sells both compute hardware and access to its chips through a cloud business, which roughly quadrupled to $126 million, compared with the year-ago quarter. But its hardware sales declined to $54.1 million from $70.3 million a year ago.
ON-CHIP MEMORY BOON TO SUPPLY
Placing memory directly on the chip has lessened the impact of surging memory prices and placed Cerebras in a better position to compete with Nvidia, CEO Andrew Feldman told Reuters in an interview.
"Nvidia's prices have gone through the roof because of HBM prices," Feldman said, referring to the high-bandwidth memory included with AI processors. "This is a battleground, and if they can't deliver or they're having significant component price increases, of course that helps."
Advanced AI chip production is concentrated at chipmaker TSMC, which faces capacity shortages for the silicon wafers it sells at its most advanced manufacturing processes used by the likes of AMD and Nvidia.
Cerebras uses TSMC's 5-nanometer process to make its chips, which means the company can more readily access supply because it has "less pressure on it than the two and three nanometer nodes," Feldman said.
The company is also racing to expand chip volumes to support a $20 billion multiyear agreement to provide AI compute to OpenAI, a deal viewed as key to justifying its valuation.
It expects 2026 adjusted revenue to be between $880 million and $890 million, higher than its previous forecast of $855 million to $865 million. Komin said the company plans to more than triple revenue in 2027.
Cerebras also raised annual adjusted gross margin projection to 41%-43%, from 38%-41% earlier.
Second-quarter sales rose 74.3% to $180.11 million. Core revenue was $209.87 million, while analysts expected $194.23 million, according to data compiled by LSEG. Adjusted loss was $6.91 million, narrower than the $40.5 million loss a year ago.
(Reporting by Anhata Rooprai in Bengaluru and Max A. Cherney in San Francisco; Editing by Shilpi Majumdar and Sayantani Ghosh)
