Industry

Cerebras CFO claims GPT-5.4 and GPT-5.5 running on wafer-scale chips

Cerebras Systems' CFO said during the company's IPO coverage that it is testing OpenAI's unreleased GPT-5.4 and GPT-5.5 models on its wafer-scale inference chips, with public availability planned soon.

May 16, 2026

Cerebras Systems is running OpenAI's unreleased GPT-5.4 and GPT-5.5 models internally on its wafer-scale inference chips and plans to make them publicly available soon, according to the company's CFO during a CNBC interview on May 14.

Watch: Cerebras CFO claims GPT-5.4 and GPT-5.5 running on wafer-scale chips

The statement emerged during coverage of Cerebras's IPO, which closed as the largest AI hardware offering of 2026. The CFO did not specify a release timeline or pricing structure, but emphasized the performance advantage of pairing next-generation language models with Cerebras's CS-3 chip architecture. The CS-3 packs 850,000 cores on a single silicon wafer and has been used to serve models like Llama 3.1 405B at speeds an order of magnitude faster than GPU clusters.

More in Industry

OpenAI has not publicly announced GPT-5.4 or GPT-5.5. The company's most recent confirmed release is GPT-4o, which shipped in mid-2024. If the CFO's remarks are accurate, Cerebras would be among the first inference providers to host GPT-5-series weights outside OpenAI's own infrastructure. The combination of frontier-model capability and wafer-scale throughput would position Cerebras in direct competition with hyperscale API providers that rely on NVIDIA H100 and H200 farms.

Cerebras has historically focused on training acceleration, but the company pivoted to inference in late 2024 after demonstrating sub-200ms time-to-first-token on 70B-class models. Running GPT-5.x internally suggests the architecture scales to models well above 100B parameters without the memory bottlenecks that plague multi-GPU setups.