Cerebras Systems endured its worst week since going public in May, with shares tumbling nearly 20 percent after a widely followed semiconductor research firm said OpenAI's fastest new AI model tier is running on Nvidia hardware rather than Cerebras systems.
The stock closed Friday at $166.43, its lowest level since the IPO, before recovering about 3 percent in extended trading to $171.51 after OpenAI's Sam Altman pushed back on what he called speculation about the partnership. At the close, the AI chipmaker's market value sat at just over $39 billion — a steep fall from the $95 billion valuation it commanded on its first day of trading in May.
The trigger was a post on X on Wednesday from SemiAnalysis, which said OpenAI's "Ultrafast" mode for the GPT-6.1 Sol model is being served on Nvidia graphics chips, not on Cerebras hardware. Neither OpenAI nor Cerebras has publicly confirmed or denied the hardware configuration behind the speed tier, which is why a single social post from a research shop could move a public stock by double digits.
A single post moves a $40 billion stock
The report stung because the symbolism was impossible to miss. When OpenAI unveiled its Ultrafast service tier at DevDay on September 29, it promised speeds up to eight times faster than standard — around 300 tokens per second — with raw speed as the entire selling point. In August, Cerebras had been publicly named as the technology partner powering Ultrafast mode for the prior GPT-5.6 Sol tier, at speeds up to 750 tokens per second.
Cerebras built its identity on exactly this kind of problem. Its Wafer-Scale Engine is a dinner-plate-sized ASIC that packs billions of processing cores and vast memory pools onto a single silicon wafer, avoiding the latency of shuffling data between hundreds of separate GPUs. For a company whose IPO was priced on the promise of taking inference share from Nvidia, a report that its flagship customer chose the incumbent for the speed tier itself cut straight at the investment thesis.
The slower 300-tokens-per-second figure for GPT-6.1 Sol Ultrafast, versus 750 for the prior tier, has also raised questions about whether the larger model is testing the limits of Cerebras' on-chip memory design — a technical detail that normally stays in the background.
Lockups, insider sales, and the January lifeline
The SemiAnalysis post landed at the worst possible moment. On Wednesday, the largest wave yet of a staggered post-IPO lockup expiration released about 19.4 million insider-held shares — roughly 8 percent of the company's shares outstanding — adding fresh supply to an already skittish tape.
Regulatory filings added to the pressure: CEO Andrew Feldman and CTO Sean Lie sold more than $240 million of Class A stock between August 20 and September 25 under trading plans adopted shortly after the IPO, and other executives also sold holdings worth millions.
What has not changed, at least on paper, is the January agreement under which Cerebras said it would supply OpenAI with 750 megawatts of compute through 2028 in a deal valued at more than $10 billion. OpenAI remains Cerebras' largest customer by revenue backlog — which is precisely why the question of who powers the Ultrafast tier carries outsized weight for the stock. With analysts expecting inference to overtake training as the dominant AI data-center task by the end of the decade, every contested inference workload between Cerebras and Nvidia matters for the company's narrative.
