Cerebras Executive Says Inference Speed Matters for AI Agents

At the AI Data Pipeline Forum on Oct. 15, Andy Hock, chief strategy officer at Cerebras Systems Inc., argued that inference speed matters more when a single request sets several AI agents to work.

“That’s where speed really makes the difference,” Hock said. He described agents taking on separate tasks and coordinating their work. His point was that speed matters in this setting, where multiple agents contribute to a response.

On selected open models, output-token generation on Cerebras systems reached up to 15 times the speed of NVIDIA GPUs. The company’s comparison uses its own measurements and an external benchmark from April 14, 2026. It measures token output on selected models, not the time required for a complete task involving several agents.

Inference is the stage when an AI model generates an answer; decoding produces that answer one token at a time. The distinction between token-generation speed and the time needed to coordinate a larger task is relevant to Hock’s remarks. Cerebras is also positioned for decoding in General Compute’s multi-vendor AI platform, which assigns different hardware to separate stages of inference.

Cerebras’ latest reported quarter ended June 30, 2026. Cloud and other services revenue reached $126 million, up 281% from a year earlier, while hardware revenue fell 23% to $54.1 million. Total GAAP revenue rose 74% to $180.1 million, and the company posted a GAAP net loss of $450.5 million.

Higher customer usage, new and expanded customer arrangements, and dedicated capacity deployments mainly drove the increase in cloud and other services revenue. Agent workloads were not identified among the drivers of that increase.

The company closed its IPO on May 15, selling 34.5 million shares at $185 each and reporting about $6.38 billion in gross proceeds. The final share count included the full 4.5 million-share underwriter option in addition to the initial 30 million shares.