Under this production architecture, standard graphics processors handle the initial prefill phase while Cerebras machines dedicate their resources entirely to the decode phase. The setup is already operating on the Cerebras Inference Cloud using AMD Instinct GPUs alongside Cerebras CS-3 systems. Decode processing benefits from wafer-scale chips equipped with tens of gigabytes of on-chip SRAM and tens of petabytes per second of memory bandwidth. Cerebras plans to scale this hybrid deployment across hundreds of megawatts of infrastructure through 2027.
Key points
Disaggregating inference tasks between GPUs and wafer-scale engines relieves critical pow…: Disaggregating inference tasks between GPUs and wafer-scale engines relieves critical power constraints for compute-heavy AI applications.
Third-party infrastructure partners can now deploy this disaggregated architecture direct…: Third-party infrastructure partners can now deploy this disaggregated architecture directly within their own data center facilities.
Cerebras intends to utilize this design to scale token output across 700 megawatts of tot…: Cerebras intends to utilize this design to scale token output across 700 megawatts of total planned capacity alongside its upcoming CS-4 system.
Sources checked by our analysis pipeline: 1
We’ve deployed Cerebras Wafer Scale Engines with GPUs to deliver lightning-fast tokens at scale
The world is hungry for ultrafast tokens. In 2026, Cerebras saw an insatiable demand for fast inference, both on our inference cloud and via our partners. To meet demand, we began developing disaggregated inference: GPUs handle prompt processing (prefill) and Cerebras systems handle output generation (decode).
This solution is now in production and delivers five times more tokens per megawatt at the same ultrafast Cerebras speeds.
We can now serve more ultrafast tokens via Cerebras Inference Cloud to meet market demand. Infrastructure partners can also leverage the Cerebras disaggregated solution to achieve more ultrafast tokens in their own datacenters.
Alongside our recently announced CS-4, our dramatically scaled manufacturing, and our planned 700MW of data center capacity, disaggregated inference will enable Cerebras and partners to serve many more ultrafast tokens in 2027 and beyond.
In this blog we cover:
- The need for more ultrafast inference capacity
- Why Cerebras is advantaged for decode
- Cerebras disaggregated inference production results
- Cerebras disaggregated inference for infrastructure partners
The tech giant's latest strategic shift repositions its enterprise software suite directly into an integrated artificial intelligence environment. By embedding essential productivity applications into this single hub, the company aims to fundamentally change how employees navigate daily digital tasks. Future updates will also incorporate persistent automation to handle tasks continuously across the platform.
This new routing strategy enhances real-time large language model inference by reusing prompt data stored directly in GPU memory. By avoiding redundant computations during long shared prompts or multi-turn dialogues, the system optimizes instance throughput while balancing compute loads across available hardware. Organizations can implement this capability without altering existing client-facing application interfaces.
Relying on large language models for routine operations like input validation, routing, and formatting introduces hundreds of milliseconds of latency and output variability. Engineering organizations are curbing these production expenses by adopting a structured evaluation framework that reserves expensive calls exclusively for ambiguous generation tasks. Shifting rule-based logic to deterministic code restores auditable, stable behavior while driving response times down to single-digit milliseconds.