CBRS Q2: Core Revenue Doubles, Rent-Back Squeezes Margin — AMD Deal Is Cerebras Capex

Cerebras beat on core revenue at $209.9M (+103%) and raised FY guidance to $880-890M, but GAAP missed on SBC and Q2 margin dipped 500bps on rented capacity. The AMD Helios deal is not a revenue share.

Cerebras reported Q2 after the close Aug 12 and the story is two numbers at once. Core revenue doubled to $209.9 million and the full-year guide went up, but GAAP revenue missed and the stock fell about 16% after-hours. The gap is stock-based compensation and a temporary margin squeeze from renting back its own systems to meet demand.

The call framed 2026 as a build year for a much larger 2027. Management said it expects to more than triple core revenue in 2027 off $25.4 billion in remaining performance obligations, with gross margin trending toward 60%+ over time.

What printed

Core revenue was $209.9 million, up 103% year over year, ahead of the $194 million core guide given on the Q1 call. Core Cloud and services was $127.7 million, up 287% year over year, nearly a fourfold increase. Core hardware was $82.1 million, up 17% year over year. Management said to focus on total core revenue, not mix, which swings with data-center timing.

GAAP revenue was $180.1 million versus about $191 to $194 million expected, and GAAP EPS was -$2.98 versus about -$0.18 expected, with the loss driven almost entirely by stock-based compensation. GAAP net loss was $450.5 million versus a profit a year ago. On a core basis, operating loss narrowed to $33.6 million.

Margins improved year over year but dipped sequentially. Core gross margin was 40.6%, up 940bps YoY but down from 46.5% in Q1. Core cloud gross margin was 41.8%, up 1,600bps YoY. Core hardware gross margin was 38.8%, up 510bps. Core operating margin was -16% versus -42% a year ago, a 2,600bps improvement while more than doubling revenue.

Backlog and balance sheet remain the foundation for the 2027 ramp. RPO was $25.4 billion at June 30 and does not include AWS or other hyperscaler backlog. Cash and investments were more than $8.6 billion with an $850 million revolver undrawn, after the $6.4 billion IPO in Q2.

Guidance raised, but Q3 is the trough

For Q3 2026, Cerebras expects core revenue of $214 to $216 million, core gross margin of 38% to 40%, and core operating margin of -25% to -23%.

For full-year 2026, the company raised all three core guides. Core revenue to $880 to $890 million from $855 to $865 million. Core gross margin to 41% to 43% from 38% to 41%. Core operating margin to -19% to -17% from -28% to -32%.

CFO Bob Komen said Q3 will be the low point for core gross margin before improving significantly in Q4 as more data centers come online filled with lower-cost Cerebras-owned systems. That mix shift will lift cloud gross margin back up.

Why margin dipped: the rent-back math

The sequential dip was not a pricing problem. Komen said core gross margin would have been about 500bps higher and more similar to last quarter without the rent-back. That implies an underlying margin near 45.6% versus 46.5% in Q1, essentially flat.

Cerebras rented back its own systems from cloud customers to meet demand from OpenAI and other cloud customers while new data centers were still ramping. Rented capacity carries higher cost than owned systems, so cloud margin compressed temporarily.

The path back to expansion has five drivers Komen listed:

• Pricing for fast tokens reflected in hardware and cloud deals to be recognized over the next several quarters
• Rolling off rented systems for owned systems
• Throughput gains of 20x over the next 18 months at the same speed, lowering cost per token per system and per watt
• Scale benefits on bill of materials
• Staying on 5nm wafers, cheaper and less constrained than 3/2nm, and avoiding HBM and CoWoS entirely

Longer term, management targets gross margin of 60%+ as these drivers compound.

AMD Helios: Cerebras capex, not a revenue share

The most asked question was how the AMD Helios deal splits economics. The answer was clear.

Prefill, which is parallel input processing, runs on GPUs and HBM where Helios is strong. Decode, which is sequential and memory-bandwidth heavy and represents the bulk of inference work, runs on the Cerebras Wafer-Scale Engine. The two are linked with standards-based I/O, which Cerebras said made integration easy.

Barclays asked directly whether Cerebras buys a Helios rack for its cloud and keeps all customer revenue or shares it. CEO Andrew Feldman said yes to the first part. Cerebras buys and stands up Helios racks in its own cloud and sells the combined service. No revenue share was disclosed, and Feldman said additional parts of the AMD arrangement will be announced over time.

The combined system keeps Cerebras speed, pegged at a 15x lead and 10x faster on GPT-5.6 Sol, and adds about 5x more throughput than Cerebras alone. Feldman called it vastly faster than Helios alone and vastly more throughput than Cerebras alone, and said there are already buyers.

Needham asked if using Helios to serve OpenAI is incremental revenue. Feldman declined to detail OpenAI terms but framed the math. At the same speed, more throughput means more tokens per system, lower cost per token, lower power per token, higher margin, and more dollars per megawatt of data center. In a power-constrained world, that makes each data-center investment more valuable.

The deal also extends the life of installed GPUs. Feldman said Helios is better than older GPUs and Tranium 3 is better than Tranium 2, but even minus-one generation is vastly better in a disaggregated solution than not. Cerebras has now tied up about half the leading chip makers between AMD and AWS.

On AMD's recent inference-hardware acquisition, Feldman called the target interesting and innovative but said he does not see the first application being data-center inference.

Capacity, capability, customers

Capacity remains the bottleneck. Over the last seven months Cerebras secured more than 600MW of data-center capacity live or under contract by end of 2027, with a pipeline measured in gigawatts. Sites include Alabama, Dallas, Denver, Minneapolis, Santa Clara, Stockton, plus France, Finland, Manitoba, Montreal, Norway, Saskatchewan and Toronto. Inference does not need gigawatt training footprints, which gives more flexibility.

Manufacturing is scaling with Flex and Sanmina. Output is already 4x the first half of 2025 and will be more than 10x for full-year 2026, with 3 to 4x more contracted for 2027. TSMC 5nm wafers are secured, which avoids constraints facing 3nm and HBM supply.

Capability is about speed and throughput. GPT-5.6 Sol now runs at 10x faster than alternatives on Cerebras, which management said proves software maturity for large frontier models. The roadmap is to double speed each year from the current lead and increase throughput by more than 20x over the next 18 months. CS4 will be unveiled next week at Supernova, CS5 is planned for H2 2027, with stacked memory and wafer-scale optics in development with US government partnerships.

Customers are broadening beyond the $20B+ OpenAI deal signed Dec 24, 2025. Q2 included six deals over $30 million. Coding customers include Figma, Cognition and Lovable in Europe. Agentic customers include Block, AlphaSense and GSK. A new market is security with CrowdStrike, where fast AI enables inline LLM inspection that is invisible to users. AWS Trainium plus Cerebras will be deployed in Amazon data centers and delivered via Bedrock, live in Q1 2027, with other hyperscalers in discussion for mid-2027 revenue and 2028 ramp.

On concentration, UBS asked if two customers will be two-thirds of revenue next year. Feldman gave history, from government to sovereign cloud at $1B to frontier lab to hyperscaler, and said OpenAI will stay large but AWS and others including coding and security will be a larger portion. An analyst floated about $1B from AWS next year, which management did not confirm.

This preview from earlier today framed the $194M core guide and margin setup heading into the print CBRS Q2 Earnings Preview: What to Expect After the Close. For context on how backlog translates to revenue, see CoreWeave Q2: $104B Backlog, $25B More in Early Q3, and on data-center scale see SemiAnalysis Reframes H2: Microsoft's 10GW Awakening.

What to watch

• Q4 owned-system ramp versus rented mix, and whether core gross margin steps up as guided from the Q3 trough
• Conversion of the $25.4B RPO and timing of AWS Bedrock revenue starting Q1 2027
• Throughput gains flowing to cost per token and dollars per megawatt, and how that shows up in pricing
• Cash burn for data-center fit-out, where Cerebras said it is reimbursed for a meaningful portion as pass-through from its largest customer
• CS4 details at Supernova next week and any update on AMD Helios commercial terms

Sources: Bargo transcript ircall://CBRS/2026-08-12 (40,110 chars, filed 2026-08-12 22:16 ET); SEC 8-K EX-99.1 filed 2026-08-12 20:09 ET; Investing.com and 247wallst after-hours coverage.