OpenAI Debuts First AI Model on Cerebras Chips, Diversifying from Nvidia
OpenAI has unveiled its inaugural artificial intelligence model powered by chips from semiconductor startup Cerebras Systems. This marks a strategic diversification from its previous reliance on Nvidia. The new model, dubbed GPT-5.3-Codex-Spark, is a superior, more efficient iteration of OpenAI’s Codex software, specifically designed for automating coding tasks.
Launched on February 13, 2026, Codex-Spark delivers over 1,000 tokens per second—approximately 15 times faster than its predecessors. The system operates on Cerebras’ Wafer Scale Engine 3, a dinner-plate-sized chip that offers dedicated low-latency inference capabilities. This enables software engineers to swiftly accomplish tasks such as editing code chunks and running tests, with the added benefit of interrupting or redirecting the model without the need for lengthy computing processes.
This collaboration is a result of a $10 billion-plus deal inked in January 2026 between OpenAI and Cerebras. An OpenAI spokesperson stressed that while Nvidia remains “foundational” to their infrastructure, the company is “deliberately expanding the ecosystem around it through partnerships with Cerebras, AMD and Broadcom.”
Codex-Spark is initially available to ChatGPT Pro subscribers at $200 per month, with a broader rollout planned in the coming weeks. With over 1 million weekly active users already utilizing Codex, this hardware diversification could significantly disrupt the AI chip market long dominated by Nvidia.