Cerebras chips run trillion-parameter AI model 7× faster than GPU clouds

Cerebras chips run trillion-parameter AI model at record speed

Less than a week after its blockbuster 2026 IPO, Sunnyvale‑based Cerebras announced it is serving Moonshot AI’s Kimi K2.6, a trillion‑parameter open‑weight model, to enterprise customers at 981 output tokens per second. Independent benchmarks from Artificial Analysis confirm a ↑ 6.7x advantage over the nearest GPU‑based cloud provider and a ↑ 23x lead versus the median. For a 10,000‑token prompt followed by a 500‑token answer, the response time fell to 5.6 seconds, compared with 163.7 seconds on the official Kimi endpoint.

“We’re proving we can handle the largest models at the speeds we’re known for,” said James Wang, Cerebras’ director of product marketing, in an exclusive interview with Reuters.

Why the Chinese‑built Kimi K2.6 matters

Released on April 20, K2.6 is a Mixture‑of‑Experts architecture with 384 experts, activating 32 billion parameters per token. It scores 58.6 on SWE‑Bench Pro, matching the performance of GPT‑5.4 on coding tasks and leading on agentic benchmarks such as Humanity’s Last Exam. Enterprises see it as a cost‑effective alternative to Anthropic and OpenAI APIs, especially where capacity constraints have caused outages.

Moonshot AI operates out of Beijing, raising compliance questions for U.S. firms in regulated sectors. Nevertheless, several Fortune 500 names in software, finance, and health care are piloting the service under confidentiality agreements.

Wafer‑scale advantage over GPUs

Typical AI inference relies on racks of Nvidia GPUs (e.g., NVL72). Data shuttles across many chips, and inter‑connect bandwidth becomes a choke point for trillion‑parameter models. Cerebras’ Wafer‑Scale Engine 3 condenses 44 GB of SRAM onto a single wafer‑size die, delivering on‑die bandwidth >200× that of Nvidia’s NVLink. All experts for a MoE layer reside on the same wafer, so routing occurs at SRAM speed.

Wang likened the architecture to “a queue of bagels”: each user occupies a distinct hardware slice, yet the system moves tokens so fast the perceived latency remains minuscule.

Market dynamics and competition

With a fresh ↑ $95 billion market cap, Cerebras signals intent to compete not just on speed but on model scale. Nvidia’s recent $20 billion acquisition of Groq underscores the premium placed on fast inference. Wang remains confident, noting both firms refresh hardware annually and hinting at an upcoming wafer‑scale iteration.

Pricing is kept private, but Wang claims it sits in the “mid‑upper range” of GPU cloud rates—comparable per‑token cost with an order‑of‑magnitude speed boost for latency‑sensitive workloads.

Looking ahead, Cerebras aims to support closed‑source frontier models from Anthropic and OpenAI, positioning itself as the go‑to provider for enterprises that need AI agents to think at hardware speed.

For further context on AI hardware trends, see the latest analysis from Bloomberg.