Cerebras Q2 revenue up 74% YoY, partners with AMD/AWS on decoupled inference for 5x throughput

On August 12 after market close, AI chip company Cerebras Systems reported Q2 2026 results. GAAP revenue was $180.1 million, up 74% YoY; on the company's "core" basis, quarterly revenue reached $209.9 million, up 103% YoY, exceeding prior guidance.

Q2 hardware revenue was $54.1 million, down 23% YoY; cloud and other services revenue was $126 million, up 281% YoY (on a core basis $127.7 million, up 287%). Cerebras was originally positioned as an Nvidia challenger in AI chips, but its largest revenue source is now cloud services. The company's core operating margin remained in negative territory at approximately -16%; Q2 net loss was $450.5 million, primarily impacted by approximately $386.6 million in stock-based compensation expense.

The company raised guidance: Q3 core revenue forecast of $214-216 million, above consensus of $212 million; core gross margin forecast of 38%-40%; 2026 full-year core revenue forecast of $880-890 million, above prior guidance of $855-865 million. Shares fell as much as 17% after-hours on hardware segment concerns.

Another key topic on the earnings call was "decoupled inference": Cerebras is partnering with AMD and AWS to split inference into prefill and decode stages— GPUs handle the parallel-friendly prefill, while Cerebras handles the memory-bandwidth-heavy decode. Combined with AMD Helios racks, throughput can increase 5x while maintaining speed. Cerebras plans to go live on AWS Bedrock in Q1 2027.

CEO Andrew Feldman said that AI coding, agents, and security applications have higher requirements for low-latency inference; Cerebras signed six deals over $30 million in Q2, with new agreements from Figma, Cognition, lovable, Block, AlphaSense, and GSK. OpenAI's Ultrafast mode is built on Cerebras inference compute.