Third quarter 2026 earnings call, November 12, 2026

Cerebras reported a third quarter that beat its own guidance across the board and lifted its fourth quarter and full-year outlook. The call also contained the company's most direct admission yet that its OpenAI contract is behind schedule. Management said it is "running behind schedule on initial tranches" of the OpenAI deal, has deducted estimated delay credits from remaining performance obligations, and expects flattish sequential cloud revenue in Q4 because data center capacity is the binding constraint.

OpenAI: Delays, Credits and Cancellation Rights

CEO Andrew Feldman addressed investor speculation about OpenAI directly. He said OpenAI remains "a very large customer and a close partner," that OpenAI grew in Q3 versus Q2, and that more growth is expected. He also conceded that "Cerebras didn't have sufficient capacity to support ultrafast in a broad rollout," after the strong response to OpenAI's ultrafast preview built on Cerebras. Contracting for and bringing online the data center capacity the agreement requires took longer than the company would have liked. Feldman also said investors should expect OpenAI to build capacity across a range of hardware solutions, and that OpenAI will retire older models and deploy new ones across those options. He rejected rumors that Cerebras systems cannot support larger frontier-scale models, saying the CS4 and CS5 "are more than capable."

CFO Bob Komin described the contract mechanics. The OpenAI agreement, filed with the S-1, is a take-or-pay contract for 750 megawatts. If Cerebras is late delivering capacity, OpenAI is entitled to credits against future committed purchases. Cerebras calculated potential delay credits and deducted them from the $26.1 billion of RPO at September 30. Komin said the credits are estimated to be relatively small as a percentage of RPO and would be spread ratably over the next few years, so the impact on revenue and gross margin would be minimal. He said the company took a conservative approach and did not assume any delays would be excused for dependencies outside its control.

If a delay lasts 90 days or more, either party can cancel that portion of capacity. Komin said the amount that could be subject to cancellation this year is under 5% of total committed megawatts. Feldman confirmed in Q&A that no cancellations for delayed capacity have been received. The company now projects it will be back on schedule by the second half of 2027. Komin added that if a tranche were canceled, "we believe there is sufficient demand and we would find other customers to absorb it." Feldman described the RPO deduction as "the absolute worst case we could be hit with" and said Cerebras expects to do better. He noted that similar penalty provisions exist in the company's own agreements with data center builders.

On the cause, Feldman was specific: "Delays right now are exclusively land, power and shell. We have capacity. We have all materials. We are waiting on the completion of [the] power shell." He said data centers are late across the board, including those being built for OpenAI and hyperscalers. Asked whether OpenAI could switch part of the contract to hardware sales, Feldman said the contract gives them that right, but that OpenAI's own data centers are also behind. OpenAI also holds an option to buy an additional 1.25 gigawatts starting in 2029, which Komin said has not yet been discussed.

Q3 Results Beat; Q4 Guide Steps Up on Hardware, Cloud Flattens

Core revenue was a record $231 million, up 10% sequentially and 70% year over year, $16 million or 7% above the $215 million guidance midpoint. Core cloud and other services revenue was $145 million, up 272% year over year and 13% sequentially. Core hardware revenue was $86 million, up 5% sequentially. Cerebras posted positive non-GAAP net income for the first time as a public company, at $30 million of core net income.

For Q4, the company guided core revenue to $310 million to $330 million, implying sequential growth approaching 40% and year-over-year growth of about 85%. The increase comes from new hardware orders, including Oracle, that begin shipping in Q4 and continue through at least the first half of 2027. Cloud revenue, however, is expected to be flattish sequentially. Komin said this is not a pricing or demand issue: "The pricing environment is healthy, and demand for our solutions continues to outpace supply by a wide margin." Cloud growth is limited entirely by data center capacity, which begins to ramp in December, "but unfortunately, that's too late to affect Q4." Komin said the hardware mix in Q4 and Q1 will be closer to 50% of revenue than in recent quarters, and that a fuller 2027 outlook will come on the Q4 call.

The full-year 2026 outlook was raised to core revenue of $940 million to $960 million, core gross margin of 42% to 44%, and core operating margin of minus 10% to minus 8%. Management continues to expect revenue to more than triple in 2027. Komin said the pipeline has grown by much more than 10x over the last few months across hyperscalers, neo-clouds, sovereigns, enterprises and AI natives.

Margins: Q2 Was the Trough, Q4 Steps Up 700 Basis Points

Q3 core gross margin was 41%, up 40 basis points sequentially and 200 basis points above guidance. Komin said the low point came in Q2, a quarter earlier than the company had guided. Core hardware gross margin declined sequentially because of the final shipments under a years-old purchase agreement at lower pricing. Feldman said those shipments were covered by special pricing offered to an early customer. Cloud gross margin improved to 46.9%, up 510 basis points from Q2, on new business at higher pricing. That was partially offset by the practice of renting back some of Cerebras's own hardware from a customer at higher cost to bridge the capacity gap, which Komin expects to continue through at least mid-2027.

For Q4, core gross margin is guided up by 700 basis points to 48% at the midpoint, with management expecting to exceed 50% in the near term and move toward its 60% long-term target. Core operating margin was a loss of 14.3% in Q3, nearly 10 points better than the guidance midpoint, and is guided to between minus 1% and 1% in Q4. Management targets a 40% long-term operating margin and expects significant progress toward it in 2027.

Yields: Management Takes on the Bear Case

Feldman said there are persistent misconceptions in the financial community, namely that Cerebras systems cannot support large models and that manufacturing yields are low. He said both are false and addressed yields at length. Fab yields from TSMC and parametric yields are both high, he said. The design uses roughly 1 million identical tiles with a small portion held in reserve, so the hardware routes around defects, in the fab and in the field.

For parametric variation across the wafer, Cerebras removes the weakest outer edges by using the largest square that fits in a 300-millimeter wafer, validates timing across a broader range of operating conditions than others use, and tunes core voltage independently across 84 wafer sections. "We know of no other chip with this capability," Feldman said. He attributed the low hardware gross margins to the ongoing manufacturing ramp and the final shipments under special pricing, not to yields, and expects hardware margins to exceed the 60% long-term target eventually.

Customers: Microsoft Becomes the Third Hyperscaler

Cerebras announced a cloud deal with Microsoft, which joins AWS and Oracle for three hyperscaler wins in six months. Feldman said more details will come in the first half of next year, when revenue from the deal is expected to begin. No size metrics were given. Oracle is purchasing hardware to deploy in its U.S. data centers as part of OCI, with a Cerebras-powered inference service expected to launch in the summer. Feldman said, "we do not think this is the last deal we do together." Cerebras is also adding neo-cloud channels, including Fireworks, and cited cloud customers including Cognition, Sigma, Block and a consumer assistant platform. Feldman said customers tell the company that fast tokens are "the most profitable tokens they sell."

Capacity: 700 Megawatts Contracted, Mostly Outside the U.S.

Cerebras has secured about 700 gross megawatts of data center capacity for delivery by the end of 2027, with more than 75% of 2027 capacity coming online outside the U.S. New facilities in Canada, France and Norway come online in the next several months, additional capacity was contracted in Norway and Finland, and the company is in final stages on a site in Sao Paulo. Feldman said Cerebras is in advanced negotiations for gigawatts of additional capacity and cited the evolving U.S. policy landscape as a potential challenge for domestic development. Komin said the bulk of the capacity ramp starts in mid-2027, and that new large data centers can take 15 to 18 months to bring online.

On manufacturing, Feldman said Cerebras expects to exit 2026 with more than 10x the manufacturing capacity it had at the start of the year, working with Flex and Sanmina, and plans to add factories in 2027. TSMC has committed sufficient wafers for 2027. The company does not use HBM, and Feldman expects memory shortages and price increases to continue through 2028, raising competitors' costs.

Roadmap: WSE-5 Back From the Fab, Disaggregated Inference Live

The WSE-5 taped out June 15, is back from the fab, running models and in in-chassis testing, which Feldman called exceptional execution for moving from tape-out to system tests in under 5 months. The CS5 will more than double the performance of the CS4, is built on TSMC's 5-nanometer node where wafer competition is less intense and costs are much lower, and should enter production in early summer 2027.

Cerebras also launched production-scale heterogeneous disaggregated inference, using AMD GPUs for prefill and Cerebras systems for decode. Feldman said this raises the number of high-speed tokens a data center can deliver by up to 5x per megawatt, and he expects it to become a default offering in major cloud deployments. Until a few weeks ago all deployments were CS3 alone. Cerebras was a day 0 launch partner for Moonshot's Kimi K3.1, a $2.8 trillion parameter model that Feldman said runs at up to 2,000 tokens per second on Cerebras.

Responding to a question from Barclays' Tom O'Malley, Feldman argued that frontier labs' ability to bring up models on different hardware in about 3 days shrinks the value of low-level software that historically favored incumbents. In his view, that leaves differentiation to hardware and is "a tremendous advantage to innovative hardware companies like Cerebras."

Competition From Low-Batch GPUs

Asked about fast inference from low-batch GPUs, Feldman said GPUs trade throughput for speed, so cost per token jumps as speed increases. He placed the crossover where Cerebras becomes more economical at around 250 to 300 tokens per second, with Cerebras "vastly more" economical beyond roughly 500 tokens per second, and said that in the range of 800 to 2,000 tokens per second the GPU cannot compete. Even at batch size 1, he said, GPUs cannot match Cerebras's speed.

Balance Sheet

Cerebras ended Q3 with $7.8 billion in cash, equivalents, restricted cash and marketable securities, plus an undrawn revolving credit facility of up to $850 million. The only debt is the $1 billion OpenAI working capital loan received in January. Komin said Cerebras pays only bill-of-materials cost on the equipment it manufactures, rather than third-party gross margins of 70%, which lowers capital expense per megawatt. He also noted that the IPO lockup releases were fully complete as of November 9. Management will attend the Tomorrow Summit in Austin on November 17 and the Barclays Global Technology Conference on December 9.