AMD and Cerebras partner on Helios platform for ultra-fast AI inference

The partnership creates a heterogeneous compute architecture addressing the two biggest bottlenecks in generative AI: prompt processing and token generation. By pairing AMD’s high-throughput Helios racks with Cerebras’s low-latency wafer-scale chips, the companies are targeting the emerging market for real-time agentic workflows and live AI assistants. For the broader ecosystem, this signals a move away from uniform GPU clusters toward disaggregated, specialized hardware stacks. The primary signal to track is the performance of Cerebras Cloud vs. Nvidia-based services during initial enterprise trials in early 2026.

The Helios system represents AMD's first fully integrated rack-scale AI platform, designed to compete with Nvidia’s upcoming Vera Rubin NVL72 architecture. According to Forbes and TechWire Asia in July 2026, a single Helios rack combines 72 Instinct MI455X GPUs and 18 6th Gen EPYC 'Venice' processors to deliver approximately 2.9 exaflops of peak FP4 performance. This configuration has already secured massive commitments from hyperscalers and labs; for instance, Microsoft recently signed a deal to deploy Helios on Azure for cloud-based inference services (per Tom's Hardware, July 2026). This follows a separate strategic partnership where AMD committed to investing $5 billion in Anthropic to supply up to two gigawatts of AI compute capacity.

AMD's pivot toward integrated rack systems is part of a broader strategy to erode Nvidia's dominant market share, which Introl estimated at over 90% at the start of 2026. By utilizing open-standard interconnects like UALink over Ethernet rather than proprietary protocols, AMD is positioning Helios as a more flexible alternative for tier-one cloud providers. Per TrendForce in July 2026, Helios is expected to hold a notable advantage in memory-intensive tasks due to its 31TB of HBM4 capacity, which is roughly 50% more than competing flagship racks. This memory density is critical for running the mixture-of-experts (MoE) models that currently dominate the high-end inference market.