OpenAI, the developer of ChatGPT, unveiled an artificial intelligence (AI) model powered by chips from the semiconductor company Cerebras, touted as a "Nvidia rival."
OpenAI on the 12th (local time) released GPT-5.3-Codex-Spark, a lightweight version of its coding-focused AI model "Codex (GPT-5.3-Codex)." The new model prioritizes fast task throughput over complex computation.
OpenAI said the model can process more than 1,000 tokens per second, providing an "ultra-low latency" environment in which users barely notice lag. A token is a unit of data processed by an AI model, and in English data it generally corresponds to one word per token.
To achieve the new model's fast inference speed, OpenAI used Cerebras' Wafer-Scale Engine 3 (WSE-3) instead of Nvidia's graphics processing units (GPU). Unlike other semiconductor manufacturers that dice wafers into small pieces to make chips, Cerebras has wafer-scale engine (WSE) technology that turns an entire wafer into a single massive chip.
This makes it possible to use both compute and memory on a single chip without needing to link a compute chip to a memory chip. The company said this approach saves power otherwise spent moving data between chips and prevents bottlenecks, improving data processing speed. Cerebras claims its chip's response time is up to 20 times faster than conventional GPUs.
OpenAI called this "the first milestone to come out of our partnership with Cerebras," adding that "(Cerebras chips) are particularly strong for workflows that demand ultra-low latency." It added, however, that "(Nvidia's) GPUs remain a critical foundation across our training and inference workloads."
On the 12th, Cerebras signed a deal to supply OpenAI with a total of 750 megawatts (MW) of computing power from this year through 2028. The companies did not disclose specific terms, but foreign media reported the contract is worth $10 billion (about 14 trillion won). U.S. media reported earlier this month that OpenAI has been seeking alternatives after being dissatisfied with ChatGPT responses using Nvidia chips.
Analysts say OpenAI's announcement also aligns with moves to shift away from Nvidia. Reports that Nvidia would scale back its investment in OpenAI have fueled speculation of friction between the two. In addition to Cerebras, OpenAI is working with U.S. semiconductor manufacturer Broadcom to develop its own AI chips.