Cerebras Systems, Inc. Q3 FY2026 earnings call

• Customer Expansion & Hyperscaler Wins: Cerebras secured deals with three hyperscalers in the last six months, including Microsoft and Oracle. Oracle’s cloud service powered by Cerebras is scheduled to launch in summer 2027. New NeoCloud partners like Firework and consumer app "Built" have also been onboarded.
• Capacity Strategy: The company faces significant delays in data center readiness due to Land, Power, and Shell (LPS) constraints, not silicon or manufacturing issues. Approximately 75% of capacity coming online in 2027 will be outside the U.S., with new facilities in Canada, France, Norway, Finland, and Brazil. Total contracted capacity stands at ~700 gross megawatts for delivery by end-of-2027.
• Manufacturing & Supply Chain: Manufacturing capacity is expected to increase more than 10x by the end of 2026, leveraging partners Flex and Sandmina. Cerebras maintains high yields through redundancy (routing around defective tiles) and parametric variation management (independent voltage tuning across 84 wafer sections). The supply chain avoids HBM memory, mitigating industry-wide shortages.
• Product Roadmap: CS4 systems are in full production, doubling performance over CS3. The next-generation WSE5 processor (CS5) returned from the fab ahead of schedule, taped out June 15, and is undergoing system testing. CS5 is expected to enter production in early summer 2027, doubling CS4 performance.
• OpenAI Partnership Status: OpenAI remains a key customer but has experienced delays in bringing initial capacity online due to LPS issues. Cerebras projects being back on schedule by H2 2027. Technical rumors regarding support for large models are false; CS3/CS4/CS5 can support Astra-scale models.
• Disaggregated Inference: Cerebras launched production-scale disaggregated inference, combining AMD GPUs for pre-fill and Cerebras CS3 for decode. This solution delivers higher token throughput per megawatt and supports day-zero launches for large open-weight models like Kimi K3.1.

Cerebras reported record total core revenues of $231 million in Q3 2026, representing a 70% year-over-year increase and exceeding the guided midpoint by 7%. The revenue mix consists of two primary segments: Core Cloud and Other Services, which generated $145 million (approximately 63% of total core revenue), up 272% year-over-year; and Core Hardware, which contributed $86 million (approximately 37% of total core revenue). While hardware revenue declined sequentially due to final shipments under legacy low-price agreements, it is expected to rebound significantly in Q4.

• Q4 2026 Core Revenue: Raised to $310–$330 million (previously lower guidance).
• Q4 2026 Core Gross Margin: Raised to 47%–49% (midpoint 48%), reflecting a step-function improvement of ~700 basis points.
• Q4 2026 Core Operating Margin: Raised to -1% to +1%, indicating a return to near break-even profitability.
• Full Year 2026 Core Revenue: Raised to $940–$960 million.
• Full Year 2026 Core Gross Margin: Raised to 42%–44%.
• Full Year 2026 Core Operating Margin: Raised to -10% to -8%.
• Long-term Outlook: Management expects revenue to triple in 2027 with expanding gross margins targeting 60% long-term and operating margins targeting 40%.

• Data Center Construction Delays: Significant reliance on third-party Land, Power, and Shell completion creates bottlenecks. Delays in OpenAI’s capacity deployment risk triggering delay credits/cancellations, though management estimates less than 5% of committed MW could be cancelled this year.
• Gross Margin Pressure: Short-term margin compression persists due to running back leased data center capacity and hardware at higher costs until owned capacity scales up mid-2027.
• Execution Risk on Capacity Ramp: Failure to secure sufficient data center capacity could limit the ability to meet surging demand from hyperscalers and AI natives.
• Competitive Landscape: Incumbent GPU vendors may attempt to compete via low-batch inference strategies, though Cerebras argues this approach becomes economically inefficient at higher speeds (>250 tokens/sec).

Q: Analyst asked about the relative size of new deals and current RPO composition for OpenAI and Oracle. / A: Andrew stated details on Microsoft will be announced in H1 2027 when revenue starts. Bob clarified that OpenAI’s RPO includes take-or-pay commitments adjusted for estimated delay credits, plus pass-through costs for data center setup. He confirmed Oracle is not yet in RPO as the deal was signed after Q3 close.

Q: Analyst asked how heterogeneous infrastructure trends (e.g., Anthropic using AMD) impact Cerebras’ moat. / A: Andrew explained that tools reducing low-level software friction diminish incumbents' software moats, shifting competitive advantage to hardware innovation. He believes this trend benefits Cerebras as differentiation increasingly relies on superior hardware architecture rather than software compatibility.

Q: Analyst asked about RPO confidence regarding OpenAI delays and causes thereof. / A: Andrew noted the RPO deduction represents a conservative worst-case estimate for delay credits, and they expect to perform better than this. He confirmed delays are exclusively due to Land, Power, and Shell (LPS) construction issues, not silicon or material shortages.

Q: Analyst asked about CS4 vs CS3 crossover timing and if CS5 will eliminate need for GPUs. / A: Andrew indicated CS3/CS4 crossover is in T1 2027. He emphasized that while CS5 doubles CS4 performance, disaggregated inference (combining Cerebras and other GPUs) currently delivers superior throughput and efficiency for constrained environments, suggesting hybrid architectures remain valuable.

Q: Analyst asked about cost-per-token economics of Cerebras versus low-batch GPU configurations. / A: Andrew explained that while GPUs can achieve modest speed via low batching, their cost per token jumps sharply as throughput drops. Cerebras becomes more economical around 250-300 tokens/sec and vastly more so at 800-2,000 tokens/sec, where GPU architectures cannot compete effectively.