Cerebras revenue misses estimates; shares plunge 16.61% after hours
AI chipmaker Cerebras Systems reported its second quarterly earnings report since going public Wednesday after the market close. Although the company raised its full-year revenue guidance, quarterly revenue fell short of market expectations, sending its shares down 16.61% in after-hours trading. Second-quarter total revenue rose 74.3% year over year to $180 million, below analysts’ expectations of $194 million. Adjusted loss per share was $0.05, better than the expected loss of $0.17. Cerebras raised its full-year adjusted revenue guidance to $880 million-$890 million from $855 million-$865 million. It also lifted its full-year adjusted gross-margin outlook to 41%-43% from 38%-41%. However, adjusted gross margin for the second quarter was 40.6%, down from 46.5% in the previous quarter. Chief Financial Officer Bob Komin said the five-percentage-point decline was mainly because computing resources previously deployed at other customers now carry higher computing-rental costs. As a closely watched AI chipmaker, any deterioration in Cerebras’ performance could prompt investors to question whether the company can take market share from Nvidia, which has accelerated its push into the fast-growing AI-inference market through methods including technology licensing. Cerebras, based in Sunnyvale, California, sells computing hardware and provides access to computing capacity through cloud services. Revenue from its cloud business nearly tripled year over year to $126 million, while hardware sales declined to $54.1 million from $70.3 million a year earlier. Cerebras raised $6.4 billion at a $185 IPO price when it listed on the Nasdaq in May. Its shares fell after reaching a high in May, but still closed Wednesday at $262.06, up 42% from the IPO price. Cerebras’ flagship Wafer-Scale Engine is a single chip roughly the size of a serving platter and contains several trillion transistors, according to the company. Cerebras says the design is more efficient than Nvidia’s approach of connecting thousands of smaller GPUs. Unlike other AI chips that use high-bandwidth memory, or HBM, Cerebras integrates storage directly into the chip, helping it avoid the pressure from the recent surge in HBM prices. CEO Andrew Feldman said: “Because HBM prices have risen, Nvidia’s costs have increased substantially ... This is a competitive focal point. If they can’t deliver products, or if component prices rise sharply, that would of course benefit us.” Production of advanced AI chips is currently highly concentrated at Taiwan Semiconductor Manufacturing Co. TSMC faces a shortage of advanced-process silicon-wafer capacity as demand for advanced processes from companies including AMD and Nvidia continues to grow. Cerebras’ chips, however, are manufactured using TSMC’s 5-nanometer process. Feldman said that, compared with the 2-nanometer and 3-nanometer processes used by AMD and Nvidia, the supply pressure for 5-nanometer chips is lower, giving Cerebras greater flexibility in securing capacity. Cerebras also recently announced a partnership with AMD, with products expected to enter production by the end of the year. OpenAI has adopted Cerebras chips to run its GPT-5.6 Sol model.
Second-quarter total revenue rose 74.3% year over year to $180 million, below analysts’ expectations of $194 million. Adjusted loss per share was $0.05, better than the expected loss of $0.17.
Cerebras raised its full-year adjusted revenue guidance to $880 million-$890 million from $855 million-$865 million. It also lifted its full-year adjusted gross-margin outlook to 41%-43% from 38%-41%.
However, adjusted gross margin for the second quarter was 40.6%, down from 46.5% in the previous quarter. Chief Financial Officer Bob Komin said the five-percentage-point decline was mainly because computing resources previously deployed at other customers now carry higher computing-rental costs.
As a closely watched AI chipmaker, any deterioration in Cerebras’ performance could prompt investors to question whether the company can take market share from Nvidia, which has accelerated its push into the fast-growing AI-inference market through methods including technology licensing.
Cerebras, based in Sunnyvale, California, sells computing hardware and provides access to computing capacity through cloud services. Revenue from its cloud business nearly tripled year over year to $126 million, while hardware sales declined to $54.1 million from $70.3 million a year earlier.
Cerebras raised $6.4 billion at a $185 IPO price when it listed on the Nasdaq in May. Its shares fell after reaching a high in May, but still closed Wednesday at $262.06, up 42% from the IPO price.
Cerebras’ flagship Wafer-Scale Engine is a single chip roughly the size of a serving platter and contains several trillion transistors, according to the company. Cerebras says the design is more efficient than Nvidia’s approach of connecting thousands of smaller GPUs. Unlike other AI chips that use high-bandwidth memory, or HBM, Cerebras integrates storage directly into the chip, helping it avoid the pressure from the recent surge in HBM prices.
CEO Andrew Feldman said: “Because HBM prices have risen, Nvidia’s costs have increased substantially ... This is a competitive focal point. If they can’t deliver products, or if component prices rise sharply, that would of course benefit us.”
Production of advanced AI chips is currently highly concentrated at Taiwan Semiconductor Manufacturing Co. TSMC faces a shortage of advanced-process silicon-wafer capacity as demand for advanced processes from companies including AMD and Nvidia continues to grow. Cerebras’ chips, however, are manufactured using TSMC’s 5-nanometer process.
Feldman said that, compared with the 2-nanometer and 3-nanometer processes used by AMD and Nvidia, the supply pressure for 5-nanometer chips is lower, giving Cerebras greater flexibility in securing capacity.
Cerebras also recently announced a partnership with AMD, with products expected to enter production by the end of the year. OpenAI has adopted Cerebras chips to run its GPT-5.6 Sol model.