Cerebras Missed on Revenue, Raised Its Own Forecast, and Still Dropped 14%

A miss, a raised forecast, and a drop that outran both

Cerebras’s dossier traces a company whose IPO quarter turned its coverage from “chip launcher” to “Nvidia’s most-cited rival,” and whose revenue has run unusually concentrated in relationships that double as shareholdings — G42, its largest customer at 87% of first-half-2024 revenue by the company’s own IPO filing, also holds a stake above 5%. That figure is two years old and the record doesn’t confirm G42 still holds that share; what it does confirm is that Cerebras has since built at least one more relationship of the same shape, detailed below. Concentration, current or historical, is the backdrop the Q2 print landed against.

Cerebras reported Q2 revenue up 74% year over year to $180 million — below the $194.23 million analysts had modeled, a miss of roughly 7%. In the same release, the company raised its own annual revenue and gross-margin forecasts. Shares dropped more than 14% after hours.

Why a raised forecast didn’t stop the drop

A revenue miss paired with raised full-year guidance is not, on its face, a bearish combination — it says the company believes the shortfall is timing, not demand. Whether a 14% drop on a 7% miss is unusually large for this specific company or this sector isn’t something the record here establishes; no comparable Cerebras print or peer chip-company reaction is available to benchmark it against, and that comparison would need to be run before calling the size of the move a finding rather than an observation. What the record does support is a plausible mechanism for why a miss this size could move the stock this much: Cerebras’s dossier establishes concentration in structural terms — one customer accounting for the large majority of first-half-2024 revenue while also holding an equity stake large enough to draw a CFIUS review of the IPO itself. A quarterly miss at a company with that customer profile reads less as “demand is soft” and more as “the timing of one or two large orders moved,” which is a harder story for a market to price confidently than a broad-based demand slowdown would be.

A week later, Cerebras unveiled CS-4 on August 19, a server rack built on a new “Nexus” architecture with 50% fewer components than its predecessor and powered by three WSE-3 Turbo chips, available in the third quarter of 2026. A server-chip launch planned on a quarters-long cycle is not a reaction to a stock drop seven days earlier — nothing in the record ties CS-4’s timing to the Q2 print — but the launch does bear on the same concentration question the print raises. A lower-component-count, faster-to-ship rack is the kind of product that could shorten the sales cycle concentrated, lumpy orders make painful. Whether it actually does, or simply gives the same small set of large customers a faster machine to order, is a question CS-4 raises and doesn’t itself answer. OpenAI is the other name in that small set. It agreed in April to pay Cerebras more than $20 billion for chip access, doubling a prior deal from January, with equity in Cerebras attached to the commitment — on top of the roughly $2.6 billion in combined CoreWeave and Cerebras stock OpenAI already holds from similar buy-commitment structures. One day after the Q2 miss, OpenAI previewed Ultrafast, an API tier powered by Cerebras chips that runs GPT-5.6 Sol up to fourteen times faster. A concentrated customer base cuts both ways: G42’s 87% revenue share is what drew a CFIUS review of the IPO itself, and OpenAI’s stake-plus-purchase-commitment structure is the same shape of relationship on a second name. The print that missed by 7% landed inside a customer relationship the market has already shown it treats as load-bearing.

Sources