Cerebras Q2 Reveals Inference’s New Bottlenecks

Cerebras is positioning fast inference as a distinct infrastructure layer, but its ambitious growth plan depends on executing a massive capacity ramp while reducing customer concentration.

• –More than 600 MW of data-center capacity is live or contracted, with manufacturing capacity expected to expand over 10x in 2026
• –Disaggregated inference with AMD and AWS could combine Cerebras’ low latency with higher throughput, improving utilization and margins
• –A $25.4 billion backlog and guidance for more than tripled 2027 revenue create significant upside if deployments stay on schedule
• –OpenAI remains a major customer, while AWS, coding platforms, security workloads, and enterprise applications are intended to diversify revenue
• –The key risk is execution: Cerebras remains loss-making, data-center buildouts are complex, and much of the growth thesis sits beyond the current quarter