OpenAI, the company behind ChatGPT, is taking big steps to secure its chip supply and reduce operating costs by designing its first in-house chip for artificial intelligence systems. Partnering with Broadcom, OpenAI has initiated custom chip development focused on inference, an area expected to see soaring demand as AI technology becomes more widely implemented.
To bring the project to life, OpenAI has secured a manufacturing commitment with Taiwan Semiconductor Manufacturing Company (TSMC) to produce its chip, targeting a 2026 release. Assembling a team of 20 engineers, including former experts from Google’s Tensor Processing Unit (TPU) team, OpenAI aims to create chips tailored specifically to its AI workloads, with Broadcom providing expertise on information transfer and design.
In response to ongoing chip shortages and steep costs associated with Nvidia’s graphics processing units (GPUs), OpenAI is also diversifying its chip sources. The company, one of Nvidia’s largest customers, plans to add AMD's MI300X chips to its infrastructure, which it accesses through Microsoft's Azure platform.
The move to include AMD chips offers an alternative to Nvidia, whose GPUs currently hold over 80 per cent of the AI chip market but remain in short supply amid high demand. With this diversified approach, OpenAI aims to mitigate supply risks and costs while benefiting from the advancements of AMD’s latest chip technology.
OpenAI’s need for chip power is rising rapidly, driven by the costs of training large AI models and running services like ChatGPT. The company expects compute costs—covering hardware, electricity, and cloud services—to be its biggest expense, contributing to a projected USD 5 billion loss on anticipated revenue of USD 3.7 billion this year. While OpenAI initially considered building its own manufacturing network, it ultimately decided to forgo this ambitious plan due to the high costs and time constraints, choosing instead to leverage industry partnerships and internal designs.
The partnership with Broadcom, which already provides custom chip solutions to companies like Google, is part of OpenAI’s broader strategy to optimise AI infrastructure and meet escalating demand. Although committed to collaborating with Nvidia for its upcoming Blackwell chips, OpenAI is cautiously balancing its supplier relationships as it explores new avenues.
(Inputs from Reuters)