AMD and Cerebras have announced a collaboration to build a disaggregated AI inference platform, combining AMD Instinct GPUs with Cerebras’s wafer-scale accelerators. The partnership is intended to deliver faster, lower-latency responses for AI agents and strengthen AMD’s position against Nvidia, whose acquisition of Groq addressed a similar inference-focused capability.

AMD’s GPUs would handle compute-intensive prompt processing, while Cerebras’s SRAM-based Wafer Scale Engine would generate tokens, a memory-heavy task. The companies claim the combination could improve tokens generated per watt by up to five times, though they have not provided detailed benchmarks; it is expected to reach Cerebras Cloud later this year. Cerebras says its approach may require only dozens of accelerators for a trillion-parameter model, compared with roughly 2,000 Groq LPUs cited for Nvidia’s alternative.