- Cerebras and AMD announced a disaggregated inference solution at Advancing AI 2026 that combines AMD Helios rackscale systems with the Cerebras Wafer-Scale Engine, targeting up to 5x higher tokens per second per watt versus a WSE-only setup.
- On May 20, 2026, Cerebras served Moonshot AI’s Kimi K2.6 at 981 tokens per second, independently verified by Artificial Analysis as 6.7x faster than the next-fastest GPU cloud, running on the WSE-3’s 900,000 cores and 44 GB of on-chip SRAM.
- Cerebras is running OpenAI’s GPT 5.4 models under a $20 billion multi-year deal covering 750 megawatts of wafer-scale deployment, while its adjusted gross margin forecast of 38-41% for 2026 trails Nvidia’s mid-70% range.
Cerebras Systems just landed its biggest proof point yet: OpenAI‘s GPT 5.4 is already running on its wafer-scale chips as part of a $20 billion multi-year deal, and this week the company added AMD to its stack. The new disaggregated inference architecture pairs AMD’s Helios rackscale systems with the Cerebras Wafer-Scale Engine, each doing what it does best, with AMD handling prefill and large context windows and the WSE driving token generation. The joint solution is slated for initial availability through Cerebras Cloud in the second half of 2026.
Why Inference Speed Matters Now
Inference has become the bottleneck that determines whether an AI application is actually useful. As AI agents and real-time copilots move into production, response latency directly limits what those systems can do, a slow token stream breaks the conversational loop that makes agents feel capable. Cerebras’ wafer-scale architecture was designed from the ground up to attack this problem.
The core insight is about data movement. Conventional multi-chip GPU systems spend significant time shuffling model weights and activations between chips and off-chip memory. Cerebras eliminates most of that by keeping weights and experts directly on the wafer. Where a typical GPU features 100 to 150 self-contained cores, a Cerebras chip uses roughly 900,000 smaller, uniform cores, each with its own logic and SRAM. Less off-chip traffic means faster token generation, and the benchmarks back it up.
The Benchmark Numbers
On May 20, 2026, Cerebras reported serving Moonshot AI’s Kimi K2.6, a trillion-parameter sparse model, at 981 output tokens per second. Artificial Analysis independently verified that figure as 6.7x faster than the next-fastest GPU cloud, attributed to the WSE-3’s 900,000 cores and 44 GB of on-chip SRAM. Vellum’s Open Source LLM Leaderboard, updated July 24, 2026, ranks Cerebras as fastest for throughput at 1,828.8 tokens per second and records a time-to-first-token of 0.54 seconds. In 2025, Cerebras also recorded gpt-oss-120B running at 3,000 tokens per second.
Cerebras claims its CS-3 system can exceed Nvidia’s inference speed by up to 15x, according to the company’s own benchmarks, independent testing has not confirmed the figure, and real-world results will vary by workload and model architecture.
The AMD Partnership
The disaggregated design announced at Advancing AI 2026 is a pragmatic engineering choice. Prefill, processing the input prompt and managing large context windows, is a workload that maps well to AMD Helios. Rapid autoregressive token generation is where the WSE excels. Splitting the pipeline means neither chip is operating outside its sweet spot.
Based on modelling conducted in July 2026, the combined system is projected to deliver up to 5x higher tokens per second per watt compared with a WSE-only configuration, according to Cerebras. Cerebras plans to deploy AMD Helios systems across its data centres to support the joint offering. For teams building latency-sensitive agents, the kind where a 500ms delay breaks the user experience, that efficiency gain matters as much as raw speed. The broader trend toward heterogeneous compute is visible across the industry; this partnership is one concrete example of how that plays out at the rack level. The infrastructure requirements for self-evolving AI agents make this kind of specialised, low-latency hardware increasingly relevant.
From Training Roots to Inference Focus
Cerebras’ current inference push has a clear lineage. In March 2023, the company released Cerebras-GPT, a family of seven open-weight language models ranging from 111 million to 13 billion parameters, trained on the EleutherAI Pile dataset using DeepMind’s Chinchilla scaling rules. The models shipped under Apache 2.0 licences on Hugging Face and GitHub. The stated goal was compute-optimal training: maximise accuracy for a given compute budget rather than simply scaling parameters. That same efficiency-first philosophy now runs through the hardware’s inference design.
The Financial Picture
Cerebras entered public markets in May 2026, raising $6.4 billion in its IPO. The company reported 94% year-over-year revenue growth in Q1 2026, with cloud and services revenue up 178% over the same period. The OpenAI deal, 750 megawatts of wafer-scale deployment starting in 2026, is the anchor contract. A separate partnership with Amazon combines Cerebras CS-3 systems with Trainium chips for inference workloads.
The margin picture is harder. Cerebras is guiding for an adjusted gross margin of 38% to 41% in 2026, below the 47% it reported in Q1 and well below Nvidia’s mid-70% range. The gap reflects heavy capital investment in computing capacity and data centre infrastructure, the cost of building out the supply side before the revenue scales to match. The $20 billion OpenAI contract and the Amazon partnership suggest the revenue will come; the question is how long the margin compression lasts. Sparse models, which activate only a fraction of parameters per inference pass, also suit the WSE’s architecture particularly well, which may give Cerebras a structural advantage as that model class grows.
