Cerebras Won the Hyperscalers. Now It Has to Build the Buildings.
A blowout Q3 and three hyperscaler wins collided with delay credits and a 52% drawdown—the bottleneck has moved from wafers to land, power, and shell.
CBRS · Earnings Call · 2026-11-12
The beat the tape already discounted
Cerebras reported a genuinely strong Q3: core revenue of $231 million, up 10% sequentially and 70% year-over-year; cloud and other services revenue of $145 million, up 272% year-over-year; core gross margin of 41%; and positive non-GAAP net income of $30 million, ahead of schedule. Management raised Q4 core revenue guidance to $310 million–$330 million, implying roughly 40% sequential growth and about 85% year-over-year growth, with core gross margin stepping up to 47%–49% and core operating margin around breakeven. RPO stood at $26.1 billion, cash and investments at $7.8 billion. Yet the stock is down 40.1% since late June and 52.1% from its August 4 peak of $227.15. The market is not debating whether AI inference demand exists—it is debating whether Cerebras can deliver against it on schedule. That is the story of this call.The tape has been brutal for the whole AI-infrastructure complex. Global price action shows data center capacity down roughly 30% over 90 days, AI data centers down about 16% over 30 days, and High bandwidth memory down roughly 27% over 90 days. Cerebras is not being singled out; it is being thrown into the same bucket as every company whose revenue depends on physical build-out. But Cerebras has a company-specific wrinkle: the constraint is no longer theoretical. It is now embedded in RPO as delay credit risk.
The delay credit is the new four-letter word
The freshest keyword in Cerebras’s own trajectory is “delay credit,” and it sits right next to OpenAI contract. The OpenAI take-or-pay contract covers 750 megawatts of capacity. Cerebras has fallen behind on initial tranches because it could not bring data-center capacity online fast enough. Management deducted estimated delay credits from RPO, said less than 5% of total committed megawatts could be subject to cancellation this year, and now expects to be back on schedule by the second half of 2027. Andrew Feldman was blunt about the cause: “Our delays right now are exclusively land, power, shell. We have capacity. We have all materials. We are waiting on the completion of land, power, shell.” — Andrew Feldman, CEO · 2026-11-12That is a very different admission from the June call, when the same CEO said: “Demand is not the constraint. Supply is not the constraint. The constraint is data centers.” — Andrew Feldman, Chief Executive Officer · 2026-06-23 Back then it sounded like a boast about demand. Today it is a P&L and RPO issue. Bob Komin explained the accounting carefully:
The nuance matters: delay credits are spread ratably and are estimated to be relatively small as a percentage of RPO. But the market heard the word “delay” and marked the stock accordingly. The question for the next two quarters is not whether OpenAI wants Cerebras—it clearly does—but whether Cerebras can convert signed megawatts into revenue before the contract’s penalty clauses bite.Tom, in the RPO, what we've included for OpenAI is our estimate of the committed amount of revenue under the contract, under the take-or-pay contract, no extensions, no additional options, adjusted by the delay credits that we think would be -- could be due.
The moat argument shifts from software to hardware
The most intellectually interesting part of the call was Feldman’s answer on heterogeneous infrastructure and low-level software. The bear case against Cerebras has always been that Nvidia’s CUDA moat makes alternative accelerators too hard to program. Feldman inverted it: if Anthropic can bring up AMD GPUs in three days, then low-level software is losing value, and the remaining differentiation is hardware. “If the low-level software has little value, the thing that is left to provide value is innovative hardware.” — Andrew Feldman, CEO · 2026-11-12 That is a direct attack on the incumbency moat—and a new keyword, low level software, showed up at the top of the company’s keyword trajectory this quarter.He also took aim at the “low-batch GPU” threat with a specific cost curve: low-batch GPUs can get faster by shrinking batch size, but throughput collapses and cost per token explodes. “Around 250, 300 tokens per second, we are more economical... between 800 and 2,000 tokens, the GPU can't play.” — Andrew Feldman, CEO · 2026-11-12 The low batch debate is now central to the bull case, because fast inference is only valuable if it is also economical at scale.
Management also preemptively defended yields. Hardware gross margin fell to 31%, which bears have used as evidence of manufacturing problems. Feldman attributed it to the final shipments under an old customer agreement at legacy pricing, plus the cost of renting back its own equipment, not to fab yield or parametric variation. He expects hardware margins to improve markedly and eventually exceed the 60% long-term target. That claim will be tested in Q4, when hardware revenue steps up and the old pricing rolls off.
Capacity is the product now
Cerebras has now won three hyperscalers in six months—AWS, Oracle, and, announced today, Microsoft. That is a real strategic shift. The new hyperscalers broaden the customer base beyond OpenAI, and Oracle is not even in the September 30 RPO because the deal was signed after quarter-end. Management also says the pipeline has grown more than 10x, spanning neo-clouds, sovereigns, enterprises, and AI natives.But the bottleneck is physical. Cerebras has secured approximately 700 gross megawatts of data center capacity for delivery by the end of 2027, with more than 75% coming online outside the U.S. New facilities in Canada, France, and Norway are near-term, and the company is finalizing a site in São Paulo. Manufacturing is scaling too: manufacturing capacity is expected to exit 2026 at more than 10x the start of the year, a repeat of what Bob Komin told investors in August: “We're going to end the year well over 10x our manufacturing capacity. And we already have contracted facilities 3x or 4x more for growth in 2027.” — Robert Komin, Chief Financial Officer · 2026-08-12
The difference now is that manufacturing is no longer the binding constraint. Land, power, and shell are. That is why Q4 cloud revenue is guided flattish sequentially even as demand outstrips supply—not a pricing or demand problem, but a capacity-timing problem. It is also why the hardware mix is expected to rise toward 50% in Q4 and Q1: hardware ships when customers have data centers ready, while cloud revenue waits for Cerebras’s own capacity. Komin called Q4 “an important inflection point for our growth and core financial trends.” — Robert Komin, Executive (likely CFO or similar financial role) · 2026-11-12
The bull case is now a schedule bet. If the 700 megawatts land on time, the 2027 revenue tripling and margin expansion toward 60% are credible, and the current drawdown will look like a capacity-timing discount. If land, power, and shell slip again, delay credits and customer concentration stay in the narrative. The stock is already down more than 50% from its peak. That does not make it cheap automatically—but it does mean the market has stopped paying for the demand story and started pricing the execution risk. Cerebras just told investors exactly where that risk lives.