Cerebras details CS-4 design and previews faster successors

Cerebras has moved the AI inference contest from chip specifications to the design of a complete data-centre rack. In a technical account published on 25 August after Hot Chips 2026, the company detailed how its CS-4 system packages compute, power, cooling and networking into the Nexus rack-scale platform.[1]

Cerebras says CS-4 is the first Nexus-based product and that one rack can carry three compute backpacks. Each backpack contains one wafer-scale engine together with the equipment needed to power, cool and connect it.[1]

The engineering details are concrete, but the competitive conclusions come from the vendor. Cerebras also used the presentation to set targets for CS-5 in 2027 and to describe a CS-6 concept using stacked memory. Buyers and cloud users still need independent workload tests, availability terms and prices.[1]

Cerebras describes Nexus as a reusable rack platform whose power, cooling and input-output systems can improve separately. The compute sits in three removable backpacks at the rear, while shared high-voltage equipment remains at the front. The stated goal is faster deployment and replacement without rebuilding the whole rack around every processor generation.[1]

Each backpack has its own water-conditioning system. Flow and temperature meters monitor the cooling loop, an actuator adjusts water flow, and dry quick-disconnect valves let technicians remove a unit without draining the shared system. Leak and condensation sensors can place a backpack into a safe state and cut power to its supply modules.[1]

For CS-5, targeted for 2027, Cerebras says it aims for up to 10,000 output tokens per second per user on named open models. It also targets up to 5,000 tokens per second per user on much larger frontier models and three million tokens per second per megawatt. These are internal targets attached to an unreleased system.[1]

CS-6 is further out. The company says it began work in 2024 on combining wafer-scale SRAM and compute with three-dimensionally stacked DRAM. The aim is to keep more of a large model close to the processor and reduce the surrounding system footprint. Cerebras's account gives no launch date, public price or independent benchmark for that design.[1]

Token speed can matter when an AI agent makes many sequential model calls, because every slow step compounds. It does not by itself describe accuracy, queueing under load, prompt processing time, uptime or the total bill. A useful comparison must run the same model, context length, batch size and end-to-end task across each service.[1]

Malaysian developers are more likely to encounter this competition through an API or managed cloud service than by operating a wafer-scale rack. That makes access terms important: supported models, region, data handling, service levels and ringgit-equivalent cost per completed task can outweigh a peak tokens-per-second figure in a production decision.[1]