Cerebras's weekly slide erased roughly the headline value of its OpenAI contract

A drop of nearly 20% that leaves a company worth just over $39 billion means it began the week near $48 billion to $49 billion. Roughly $9 billion to $10 billion of market value went in a week ^[17]. OpenAI's January contract with Cerebras was initially valued at over $10 billion ^[4]. A research firm's post about one model's launch hardware moved the stock by about the headline value of that contract, or rather, the post and an insider unlock did ^[2][11].

Up to 19.4 million shares held by directors, officers, employees and other holders became free to sell on Wednesday, equal to 8% of shares outstanding, CNBC reported from the prospectus ^[11]. At Friday's close of $166.43, that tranche is worth about $3.2 billion ^[18]. It came on top of up to 14.6 million shares unlocked every two weeks since August 19 ^[12]. SemiAnalysis posted its claim that OpenAI would run GPT-6.1 Sol's Ultrafast mode on Nvidia GPUs on the same Wednesday ^[2].

OpenAI's commitment is to buy up to 750 megawatts over three years ^[3]. At the initial valuation that works out to about $13 million per megawatt, or about $3.3 billion a year if spread evenly ^[19][20]. Later reports put the total above $20 billion once extensions and related financing are counted, according to CryptoBriefing ^[4]. The companies did not disclose a minimum purchase. Cerebras leases its dinner plate-sized chips from its own data centers as a cloud service ^[6], so what the contract is worth to Cerebras depends on how much capacity OpenAI takes as the phases arrive through 2028 ^[3].

Cerebras sells its hardware on ultra-low-latency inference ^[7]. The reports say the first deployment of the tier OpenAI named Ultrafast went to Nvidia GPUs ^[1]. OpenAI launched its fastest product without waiting for the Cerebras phases due through 2028 ^[1][3]. Both companies have described Cerebras as a complement to Nvidia and AMD ^[14].

Several outcomes fit the Sol report. The Ultrafast mode could move to Cerebras as later phases come online, since the reports describe only the initial deployment ^[1]. OpenAI could draw the full ceiling, and the $20 billion version could arrive ^[4]. It could also take well under 750 megawatts, and the contract wording allows that ^[3].

I think the selloff is a fair discount on that last outcome, because the Sol choice is the only public evidence so far of how OpenAI splits fast inference between suppliers ^[1]. CryptoBriefing's view is that the market overstated the risk, since one model's initial deployment does not rewrite a three-year purchase commitment ^[22]. That argument holds only if the commitment has a floor. If OpenAI moves Sol's fast tier onto Cerebras, or Cerebras discloses a contracted minimum, I am wrong.

Altman's own money is small by comparison. On the share count CryptoBriefing reports, his roughly 89,373 shares ^[5] are worth about $14.9 million at Friday's close ^[16], while Cerebras's market value is down about $56 billion from its first-day close ^[21]. CryptoBriefing wrote that his dual role, CEO of the customer and longtime shareholder in the supplier, will likely keep governance questions in the conversation ^[15].