Cerebras Systems Inc. (CBRS)

News Summary

• Cerebras disclosed disaggregated inference partnerships with AMD (Helios racks for prefill, CS systems for decode) and AWS (Trainium) targeting Q4 2026 commercial deployment, claiming the combined architecture delivers 5x more tokens per CS system—materially reducing power and cost per token—while also confirming it is serving OpenAI's GPT-5.6 Sol at 10x faster speed than alternatives, giving Cerebras compounding road map visibility Feldman called a structural moat.
• The CS4 will be unveiled at the Supernova Conference imminently and CS5 is on track for H2 2027; the road map targets doubling speed annually and more than 20x throughput improvement over 18 months from a claimed 15x performance lead, backed by manufacturing capacity scaling more than 10x in 2026 and contracted facilities for an additional 3x-4x in 2027 across 600 MW of live or contracted data center capacity across 14 global locations.
• New Q2 enterprise wins included Figma, Cognition, Lovable, Block, AlphaSense, GSK, and CrowdStrike—the CrowdStrike deal was highlighted as a net-new market enabled purely by speed, where LLMs inspect enterprise traffic inline invisibly, a use case Feldman said only exists because AI is fast; the $25.4 billion RPO backlog and management guidance for more than 3x revenue growth in 2027 reinforce accelerating adoption.
• Q2 core revenue reached $209.9M (up 103% YoY), core gross margin expanded 940 bps YoY to 41%, and the adjusted operating loss narrowed sharply; GAAP revenue was $180M versus $194M consensus, and the $450.5M GAAP net loss was almost entirely driven by $386.6M in IPO-related stock-based compensation; full-year core revenue guidance was raised to $880M-$890M with Q3 guided to $214M-$216M.
• CBRS fell roughly 14-17% in after-hours and premarket trading as markets focused on the GAAP revenue miss versus the $194M consensus, though some analysts characterized the selloff as a buying opportunity given the core beat, margin expansion, and 2027 tripling outlook; headlines were split between 'plunges on miss' and 'slides but analysts see opportunity,' reflecting genuine investor uncertainty about GAAP versus core accounting conventions for a newly public company.

Financial Highlights

Earnings Call Highlights

Cerebras delivered a strong Q2 with record core revenue of $209.9M (up 103% YoY) and beat guidance across all metrics, projecting a tripling of core revenue in 2027 backed by $25.4B in RPO. Management's tone was confident and expansive, framing 2026 explicitly as a foundation-building year while telegraphing aggressive capacity, product, and partnership moves that set up 2027-2029 as the payoff period. Analysts probed customer concentration, disaggregation economics, and the AWS ramp with genuine curiosity rather than skepticism, suggesting the investment community is broadly constructive on the growth narrative. The dominant theme was that speed is a structural differentiator that unlocks new markets and new architectural paradigms—disaggregation being the most concrete near-term expression of that thesis.

• Cerebras announced disaggregated inference partnerships with both AMD (Helios racks for prefill, CS systems for decode) and AWS (Trainium), claiming the combined solution delivers Cerebras-level speed with 5x higher throughput per CS system—Andrew Feldman stated this 'means up to 5x more tokens are produced per CS system, so power and cost are significantly reduced per token,' with commercial deployment targeted for Q4 2026.
• The company disclosed it is now serving OpenAI's GPT-5.6 Sol at 10x faster speed than alternatives, which Feldman framed as a strategic moat: 'Serving these models allows us to see into the future and to prepare for it,' giving Cerebras compounding road map advantages previously exclusive to NVIDIA.
• Cerebras secured over 600 megawatts of data center capacity (live or under contract through end of 2027) across 14 locations globally, with a pipeline now 'measured in gigawatts'; manufacturing capacity is on track to increase more than 10x in 2026, with contracted facilities for an additional 3x-4x growth in 2027.
• The CS4 is set to be unveiled at the Supernova Conference next week, and CS5 is on track for H2 2027 launch; the road map targets doubling speed annually and increasing throughput by more than 20x over the next 18 months, starting from a claimed 15x performance lead over the rest of the industry.
• New enterprise customer wins in Q2 included Figma, Cognition, Lovable, Block, AlphaSense, GSK, and CrowdStrike—the CrowdStrike deal was highlighted as a net-new market created by speed, where LLMs inspect enterprise traffic inline invisibly, a use case Feldman said 'only exists if AI is fast.'

Sources

• Cerebras (CBRS) Q2 earnings report 2026
• Cerebras Systems Fast Inference Cloud Business Nearly ...
• Cerebras Q2 2026 earnings: stock falls after guidance raise