Cerebras Systems Delivers 13x Speed Advantage on Trillion-Parameter Models, Signs Definitive AWS Agreement
Q1 2026 Earnings Call, June 23, 2026
Cerebras Systems reported its first quarter as a public company with core revenue of $191.3 million, up 92% year-over-year, while demonstrating what CEO Andrew Feldman called "the fastest AI in the world, bar none, not by a little bit, but by an order of magnitude." The company showcased a live demonstration running the trillion-parameter Kimi K2 model that completed in 21 seconds on Cerebras hardware versus 4 minutes and 37 seconds on leading GPUs, a 13x speed advantage that underscores the company's fundamental thesis that fast tokens are more valuable tokens.
OpenAI and AWS Partnerships Reshape Revenue Mix
The company has moved from definitive agreement to production deployment with remarkable speed. Cerebras signed its $20 billion-plus OpenAI contract on December 24, 2025, and achieved production deployment by February 1, taking just 35 days from signature to running models the company had never seen before. OpenAI's GPT 5.4 is now running on Cerebras infrastructure and available to OpenAI engineers and select customers, with GPT 5.5 expected to follow shortly.
More significantly, Cerebras announced this week that it has completed a definitive agreement with AWS, having signed a binding term sheet in March. The solution will combine AWS's Trainium 3 chips for prefill with Cerebras' CS-3 systems for decode in what Feldman described as an "order of magnitude faster" disaggregated architecture. This partnership provides Cerebras access to enterprises where they already store data and have existing agreements, addressing a critical go-to-market challenge.
When asked about timing for AWS deployments, Feldman indicated investors "should expect to see AWS' impact in 2027," noting that TSMC has provided supply "for our plan and beyond in 2026." The disaggregated approach represents a significant opportunity, though Feldman cautioned that specializing hardware for prefill and decode embeds assumptions about traffic shape, making the solution particularly valuable for hyperscalers like AWS that can intelligently route traffic.
Financial Performance Shows Operating Leverage Despite Near-Term Margin Pressure
Core cloud and services revenue reached $79.8 million, up 167% year-over-year, while core hardware revenue contributed $111.6 million, up 60%. CFO Robert Komin noted that hardware revenue will likely decrease in coming quarters as the mix shifts toward deploying hardware in Cerebras Cloud to fulfill significant contracts, though this could change quickly based on customer deployment preferences.
Core gross margin reached 46.5% compared to 42.1% in the prior year, with cloud and services margin hitting 52.9%. However, Komin warned that margins will compress temporarily by 10 to 15 percentage points as the company rents back capacity from existing customers to accelerate service delivery while aggressively building out its own data center infrastructure. The company expects to return toward its target margin of 60%-plus as it transitions away from rented systems.
Core hardware margin of 42% benefited from timing of performance-based incentive pricing but is expected to normalize to the low 30s going forward. Komin explained that when hardware systems are sold upfront, they include support and services with significantly higher margins that are recognized over time as cloud and other services revenue, making total profitability over contract life closer to target overall gross margin.
Operating expenses of $92.6 million grew 51%, roughly half the 92% core revenue growth rate, demonstrating meaningful operating leverage. The company reached near breakeven with core operating loss of just $3.5 million and operating margin of negative 2%, a substantial improvement from negative 19% a year ago.
Wafer-Scale Architecture Sidesteps Industry Supply Constraints
Feldman made a compelling case for why Cerebras avoids the binding constraints plaguing competitors. The company doesn't use HBM memory, which is in short supply and expensive, instead relying on SRAM printed directly on the logic wafer. "As long as you can make the chip, you can make SRAM. Its supply is approximately infinite," Feldman stated. Cerebras also doesn't use TSMC's constrained CoWoS process or 3-nanometer capacity, operating instead at the 5-nanometer node where there is less contention and lower manufacturing costs.
The company has expanded manufacturing partnerships, adding Sanmina as a second major contract manufacturer alongside Flextronics, and has added hundreds of thousands of square feet of U.S.-based manufacturing and clean room space. Feldman noted that Cerebras is "the only accelerator maker to manufacture exclusively in the U.S."
Data Center Capacity Emerges as Primary Constraint
When asked directly about constraints on growth, Feldman was unequivocal: "Demand is not the constraint. Supply is not the constraint. The constraint is data centers." The company has added facilities across the U.S., Canada, France, and the Nordics, with early discussions underway for Israel, UAE, Australia, Singapore, India, and Indonesia.
The company ended Q1 with $3.3 billion in cash, cash equivalents, restricted cash, and marketable securities after raising $1 billion in Series G in September 2025, another $1 billion in Series H in February 2026, securing an $850 million credit facility in April, and completing what Komin called "the largest semiconductor IPO in history" at $6.4 billion. This war chest provides flexibility to accelerate data center sourcing and deployment.
Market Philosophy Centers on Speed as Fundamental Value Driver
Feldman articulated a clear vision for why AI fundamentally expands the addressable compute market. "AI gives computers purchase on whole swaths of the world that had previously been foreclosed," he explained. "Computers have historically been good at math, very good, but they were relatively poor at everything else. AI opens up the world of human experience to computers. As a result, the size of the market increases exponentially."
On the question of total addressable market, Feldman pushed back against GPU makers' characterization that fast inference represents only 25% of the opportunity. "We look out into technologies and can't find examples of where slow has owned meaningful portions of the market over medium periods of time," he said, citing search and dial-up internet as analogies. "When we look out at the space, we see the entire inference market as available to us for fast inference. Who doesn't want answers in less time? And who doesn't want more productive agents?"
The company also highlighted how speed enables safety, noting that guardrails add computational overhead that takes time on slow infrastructure. "Traditionally, guardrails force the trade-off between safety and user experience, between safe and fast. Cerebras eliminates this trade-off," Feldman stated.
Guidance Reflects Backend-Loaded Ramp
For Q2, Cerebras expects core revenue of approximately $194 million (88% year-over-year growth), core gross margin of 36% to 38%, and core operating margin of negative 30% to 32%. The margin compression reflects the temporary impact of renting back capacity.
For full year 2026, the company projects core revenue of $855 million to $865 million (69% growth at midpoint), core gross margin of 38% to 41%, and core operating margin of negative 28% to 32%. Komin noted the forecast assumes hardware will decline sequentially for the rest of the year while cloud revenue ramps in the backend as data center capacity comes online. He characterized the second half outlook as conservative given the early stage of the year and ongoing data center deployment timing.
The guidance framework excludes data center pass-through costs from the OpenAI contract (reported at 3% markup), noncash amortization of customer warrants, stock-based compensation, and one-time items, with reconciliations available in earnings materials. Komin noted these pass-through amounts start small in Q1 but will become more significant, though OpenAI can choose whether to receive capacity in its own data centers, which would eliminate future pass-through amounts for that capacity.