In a market where computing power is in short supply, Cerebras’s latest artificial-intelligence hardware promises ultrafast inference speeds. Can it deliver what customers and investors want to see?
During the company’s Tuesday Supernova event, Cerebras unveiled the CS-4, an AI accelerator built from its proprietary wafer-scale engine, which is a type of AI processor. The company says that the CS-4 features three wafer-scale engines with modular system components that together can perform up to 30 times faster than traditional graphics processing units on inference, or the process by which AI models generate live responses to user queries.
Inference speeds have become a key bottleneck for agentic AI workflows, which consume large amounts of tokens, or AI data units. In order to be commercially viable, AI agents need to process those tokens and output responses quickly. The new CS-4 aims to help speed up these workflows by increasing the amount of tokens generated per megawatt tenfold relative to its CS-3 predecessor. It also features fewer components and more efficient assembly, which the company says makes it faster to deploy.
The CS-4 is now in early access and will be generally available later in the third quarter, management said.
When Cerebras made its public debut on the Nasdaq in May, Wall Street cheered the upstart chipmaker as a challenger to AI giant Nvidia. Unlike small, general-purpose GPUs, Cerebras’ wafer-scale architecture is built from an entire 300-millimeter silicon wafer that gives the chip massive memory bandwidth.
However, it’s been a bumpy post-IPO ride for Cerebras shares, which have fallen over 30% from their first-day close. Last week, the company’s second-quarter earnings report triggered a selloff for the stock as revenue fell short of consensus estimates.
“The company’s technology, what they’ve done with their innovations in semiconductors, may be the most impressive that we’ve seen in a long, long time, so I give them that,” Paul Meeks, head of technology research at Freedom Capital Markets, wrote in a note following the earnings report.
However, Meeks believes that it may be more difficult than investors originally thought for Cerebras to ramp up capacity and deliver solutions at scale. The company’s hardware products and cloud-services offerings face steep competition. On the earnings call, Cerebras flagged that it was continuing to work on scaling wafer supply, manufacturing capacity and data-center capacity.
“There might be a little bit more of a gap between the buildout and the revenue than people expect,” Meeks said.
Cerebras shares were hit especially hard amid a sector-wide chip selloff Tuesday, dropping 13%.
The latest CS-4 product should boost performance, “in turn providing Cerebras with an opportunity to lift pricing and margin (with benefits from any shift, in our view, likely not captured in Cerebras expectations),” Wedbush analyst Matt Bryson wrote in a note last week.
Additionally, the company is expanding its customer base following a high-profile multi-year agreement with OpenAI in January. Last month, Cerebras and AMD entered a partnership to create a joint AI inference solution, and Cerebras also struck a deal with cybersecurity company CrowdStrike.
Bryson believes “it’s unlikely that Cerebras incorporated either deal into its initial guidance, meaning both partnerships should create some further room for upside in future quarters vs. prior management expectations.”
Comments