Cerebras: The Fast-Inference Bets Are Coming Due
A record quarter, a $25B backlog, and a new AMD partnership suggest the company is shifting from promise to scale.
CBRS · Earnings Call · 2026-08-12
The Quarter of Inflection
Cerebras Systems reported a blowout Q2 2026: core revenue of $209.9 million, up 103% year-over-year, and core cloud revenue nearly quadrupling to $127.7 million. But the deeper story is that the company is finally converting its architectural bet on fast inference into a scalable business. The key change is not just the revenue beat — it's the disaggregated inference partnerships and the unprecedented capacity buildout that set the stage for 2027. As CEO Andrew Feldman put it: “we expect to more than triple our core revenues in '27 and continue to grow at multiples in the years following.” — Andrew Feldman, Chief Executive Officer · 2026-08-12 This is a confidence borne of a $25.4 billion RPO, up from essentially nothing a year ago.Disaggregation: The New Growth Engine
The most novel development is the company's embrace of disaggregated inference, where GPUs handle prefill and Cerebras's wafer-scale engine handles decode. Feldman announced partnerships with both AMD and AWS, noting: “the joint solution of Helios racks in front of Cerebras systems ... is an extremely strong offering.” — Andrew Feldman, Chief Executive Officer · 2026-08-12 The economics are compelling: a combined system can maintain Cerebras's speed while boosting throughput by 5x, and the roadmap promises a 20x throughput increase by the end of 2027. Feldman explained the strategic logic: “Whenever you increase throughput while keeping your performance the same, you increase your opportunity for revenue.” — Andrew Feldman, Chief Executive Officer · 2026-08-12 This is a direct counter to the GPU-centric world, and it positions Cerebras not as a competitor to NVIDIA but as a complement that unlocks underused GPU capacity. The fast inference market is the company's TAM, and disaggregation makes it more accessible to price-sensitive customers.Capacity Becomes the Moat
The company's biggest shift is in execution on capacity. Feldman highlighted that they have secured over 600 megawatts of data center capacity, with a pipeline measured in gigawatts, and manufacturing capacity will increase more than 10x in 2026. This is a dramatic acceleration from the prior quarter, when the binding constraint was clear. As he said on the last call:Now they're building the muscle to remove that constraint, using their advantage of not needing gigawatt-scale training footprints. The new data center wins span geographies, and the company is also diversifying beyond OpenAI — signing deals with Figma, Cognition, Block, and CrowdStrike, the last of which is a new security use case enabled by speed. Bob Komin emphasized the margin path: core gross margin expanded 940 bps year-over-year to 40.6%, with a clear trajectory toward 60%+ as rented systems roll off and owned data centers come online. The company is also using its balance sheet ($8.6B in cash) to fund this expansion without dilutive pressure. The manufacturing capacity and disaggregated inference efforts are not incremental — they are the foundation for a step-change in scale. The prior quarter's Fast AI narrative is now backed by concrete contracts and a visible roadmap.Demand is not the constraint. Supply is not the constraint. The constraint is data centers.