AMD, Cerebras Unveil Disaggregated AI Inference Platform Combining Helios and Wafer-Scale Engine

The CODEW | AI Infrastructure Watch

The two chipmakers say splitting inference into separate stages, each run on the hardware best suited for it, can cut latency while boosting energy efficiency up to fivefold.

AMD and Cerebras Systems announced a technical partnership Thursday to build a new AI inference platform that combines AMD's Helios rack-scale infrastructure with Cerebras' Wafer-Scale Engine, aiming squarely at the growing market for real-time, latency-sensitive AI applications like coding assistants, live agents, and autonomous workflows.

The companies unveiled the collaboration at AMD's Advancing AI 2026 event.

Splitting inference into two jobs

The core idea is "disaggregated inference" — breaking the AI inference process into its two main stages and running each on the hardware best suited to it, rather than forcing a single chip architecture to handle both.

AMD's Helios platform, built on Instinct GPUs, handles the first stage: ingesting prompts and processing large context windows at high throughput. Cerebras' Wafer-Scale Engine then takes over for the second stage — generating tokens in response, a job that is heavily bandwidth-bound and where the Cerebras chip's architecture is designed to minimize latency.

The two companies say the combined system is designed to deliver up to five times higher tokens per second per watt than a comparable setup running on Cerebras hardware alone. That figure comes from internal modeling by AMD's performance labs and Cerebras, measured on the Kimi 2.6 model at a comparable interactivity level, and both companies have attached the same qualifier to the claim.

AMD CEO Lisa Su framed the partnership as an extension of Helios' reach into a more demanding segment of the market, saying it is meant to power "a new platform for real-time agentic applications."

Following — and inverting — a broader industry trend

The approach mirrors a concept Nvidia has also pursued with its now-cancelled Rubin CPX design, which similarly split inference workloads into prefill and decode stages run on different hardware. AMD and Cerebras have effectively inverted that assignment: where Nvidia's plan used a compute-optimized chip for the prefill stage and a memory-heavy chip for decode, AMD's GPU-based Helios handles prefill here while Cerebras' wafer-scale chip handles decode.

Industry observers see the move as part of a broader shift in AI infrastructure competition — away from a single chip's raw performance and toward how well a system can orchestrate multiple, specialized types of hardware working together. Rivals including Nvidia and Groq are pursuing similar heterogeneous approaches to inference architecture.

Rollout plans

Cerebras plans to deploy AMD Helios systems directly in its own data centers, integrating them with its existing Wafer-Scale Engine racks. The joint platform is expected to become available first through Cerebras Cloud in the second half of 2026.

The announcement gave a lift to both companies' shares in early trading, with AMD stock recovering intraday losses following the news, and Cerebras Systems shares touching a one-month high.

Sources: AMD, Cerebras Systems, Tom's Hardware, Stocktwits, Yahoo Finance, Invezz, TECHi, XenoSpectrum

AI Infrastructure Watch is a recurring editorial series that tracks the hardware, cloud platforms, networking technologies, and systems architecture powering modern artificial intelligence. Rather than focusing on AI models themselves, it covers the infrastructure enabling AI at scale.