How OpenAI built an AI model without NVIDIA chips
What's the story
OpenAI has released a lightweight version of its coding tool, Codex. The latest model, GPT-5.3-Codex-Spark, is designed for faster inference and is powered by a dedicated chip from Cerebras. This marks a new level of integration in OpenAI's physical infrastructure as it reduces reliance on NVIDIA hardware. The partnership between OpenAI and Cerebras was announced last month with a multi-year agreement worth over $10 billion.
Chip integration
Spark is the first milestone in this partnership
The integration of Cerebras into OpenAI's computing solutions is aimed at significantly improving the speed of AI responses. Now, OpenAI has dubbed Spark as the "first milestone" in this partnership. The new tool is designed for quick, real-time collaboration and rapid iteration. It will be powered by Cerebras's Wafer Scale Engine 3 (WSE-3), a third-generation waferscale megachip with an impressive 4 trillion transistors.
Tool features
Spark is available in a research preview
OpenAI describes the new lightweight tool, Spark, as a "daily productivity driver, helping users with rapid prototyping." The company says it is not designed for longer, heavier tasks that the original Codex model was built for. Currently, Spark is available in a research preview for ChatGPT Pro users in the Codex app.
Future plans
OpenAI envisions Codex-Spark as the first step toward dual-mode Codex
OpenAI envisions Codex-Spark as the first step toward a dual-mode Codex: one for real-time collaboration and rapid iteration, and another for long-running tasks requiring deeper reasoning. The company also noted that Cerebras's chips are particularly good at enabling workflows that require extremely low latency. Cerebras has been in the game for over a decade but has gained more prominence in the tech industry with the rise of AI.
Financial growth
Cerebras raises $1 billion, reveals IPO plans
Just last week, Cerebras raised $1 billion in fresh capital at a valuation of $23 billion. The company has also revealed plans for an initial public offering (IPO). Sean Lie, CTO and co-founder of Cerebras, expressed excitement over the new model saying it will open up new interaction patterns and use cases. "This preview is just the beginning," he added.