Cerebras Systems stock fell nearly 20% this week to its lowest level since the company went public in May, closing Friday at $166.43. The drop combined two hits at once: a research firm’s report that OpenAI’s fastest new model tier runs on Nvidia chips rather than Cerebras hardware, and the largest wave yet of a staggered lockup expiration that released about 19.4 million insider shares. In extended trading the stock recovered somewhat after OpenAI’s Sam Altman pushed back on what he called speculation about the partnership, trading after hours at $171.51, up about 3%.
The report came from SemiAnalysis, which posted on X on Wednesday that OpenAI’s "Ultrafast" mode for the GPT-6.1 Sol model runs on Nvidia graphics chips, not on Cerebras systems. OpenAI positioned GPT-6.1 Sol as a model that nearly matches its flagship GPT-6 Astra on agentic coding and professional work at roughly one-fifth of the price, with speed as the selling point. Neither OpenAI nor Cerebras has publicly confirmed the hardware configuration for the speed tier, which is why a single social post from a research shop could move a public stock by double digits.
For a company that priced its IPO on the promise of taking inference share from Nvidia, the symbolism was hard to miss. The model tier is called Ultrafast, and the story, unconfirmed as it is, says the ultrafast part is someone else’s silicon. Cerebras has built its identity on raw inference speed, and its marketing leads with captures like tokens per second. When a benchmarking firm reports that the flagship customer chose the incumbent for exactly that tier, the discount applied to the stock reflects doubt about the customer’s commitment, not just about one workload.
The deal that explains the weight
The market’s reaction is easier to understand with the deal on the table. In January, Cerebras agreed to supply OpenAI with 750 megawatts of compute through 2028 in an arrangement valued at more than $10 billion. Cerebras builds dinner plate sized waferscale chips tailored for inference and sells them as a cloud service from its own data centers. Its pitch depends on doing inference faster and cheaper than anything built on Nvidia’s fleet. Losing one of the highest profile inference workloads in the industry, for a speed tier literally marketed on being fast, takes air out of the alternative-to-Nvidia trade.
Cerebras said this week that early results of a disaggregation technique raised inference throughput 5x with the same number of systems and no loss in token generation speeds, so the technical story has not stood still. Third party benchmarking from Artificial Analysis in September listed Cerebras first among commercial inference providers for a large open model, at 1,669 output tokens per second, ahead of SambaNova at 708, Groq at 475, Microsoft Azure at 319, Nebius at 294 and Baseten at 293. But investors now have gauge readings as well as claims. Cerebras went public at $185 on May 14, popped soon after to a market capitalization around $95 billion, and has spent recent sessions below its IPO price, closing October 1 at $169.52, 8.4% under the offer price.
A lockup schedule built for bad weeks
The supply side compounded the damage. Under the IPO terms, shares unlock in 11 separate waves rather than one block, a structure designed to smooth selling pressure that in practice can stretch it across a quarter. Beginning Wednesday, up to 19.4 million shares held by directors, officers, employees and other holders became tradeable, about 8% of shares outstanding. Smaller blocks of 14.6 million shares had already released every two weeks since August 19. Two further larger tranches arrive October 14 and October 28, and the whole lockup ends no later than November 9, which falls two trading days after the third quarter earnings report. The structure keeps supply on the schedule on a weekly basis into November, which limits how quickly a rebound can clear the overhang.
| Unlock wave | Shares | Timing |
|---|---|---|
| August waves | 14.6M each | Every two weeks since Aug 19 |
| Large tranche | up to 19.4M | Oct 1 |
| Large tranche | pending | Oct 14 |
| Large tranche | pending | Oct 28 |
| Full lockup ends | remaining | No later than Nov 9 |
Executives are not waiting for a better tape. Chief operating officer Dhiraj Mallick completed the sale of 396,000 shares under a June 30 Rule 10b5-1 trading plan, a scheduled sale rather than a reaction to the week, but it landed in the same filing cycle as the bad headlines, and optics count when a quarterly report approaches. Institutional holders watching insider behavior between unlock waves knew the sale was preplanned, yet the disclosure still fed the narrative that people close to the company are reducing exposure.
What moves the stock from here
Cerebras has real points in its column: cloud revenue growth claimed at 281% on a GAAP basis, 600 megawatts of data center capacity under contract, manufacturing capacity set to scale more than 10x in 2026, partnerships with AMD and AWS, and OpenAI as the launch partner for its frontier model GPT-5.6. The company also counts a fresh 165 megawatt data center project in Finland among recent milestones. But the week’s two events cut at the same belief, that Cerebras takes Nvidia’s inference share for the biggest customer in AI. McKinsey expects inference to overtake training as the dominant data center workload by the end of the decade, which makes an inference tier, rather than a training chip, the battleground that matters most.
Watch two dates. October 14 and October 28, when the remaining large unlock waves hit an already thin tape, and the third quarter earnings report set to land inside days of the full lockup expiring. If earnings land near the November release with a credible OpenAI revenue number and clear language on which workloads run where, the thesis could recover from this level. If OpenAI clarifies the other direction, or stays quiet while the remaining tranches unlock, Friday’s close will look less like a bottom and more like a checkpoint. For now the market is treating the Nvidia claim as probable, the lockup as certain, and耐心 nothing in between.