OpenAI has confirmed that It will have its first chip ready to power its AI systems throughout 2026. But unlike what was proposed a few months ago, it will not be developed internally, but rather it will be an external development. To do this, the company has teamed up con Broadcom y TSMC. In addition, according to Reuters, it will also add chips from AMD and Nvidia to its systems to cover its infrastructure needs to work with Artificial Intelligence.
The company has been evaluating options for some time to reduce costs and diversify the supply of chips it needs. At first they thought about developing everything internally, for which they wanted to obtain more financing. Their goal: build a network of factories to make chips. But they have finally abandoned their plans due to the high costs that internal manufacturing would entail.
OpenAI requires very high computing power to train and operate its systems and models. It is one of the main buyers of Nvidia GPUs, and uses chips both to train models in which AI learns from data, and for inference, applying AI to the predictions and decisions it makes based on the new information they receive.
Apparently, OpenAI has been working for months with Broadcom to make its first AI chip for inference, although now demand is greater for chips dedicated to training. But predictions for the future raise the need for inference chips, which in the medium term could overtake training chips in demand as more AI applications are deployed.
For now, OpenAI’s plans to train models involve using AMD chips through the Microsoft Azure cloud platform. Until now they have relied almost entirely on Nvidia GPUs to do so, but have changed systems due to chip shortages and infrastructure delays, as well as the high cost of training following only that system.
In addition, OpenAI is considering whether to develop or Buy other elements it needs to design more types of chip, and may reach agreements with more companies to acquire them. Meanwhile, your address has created a group dedicated to chipscomposed of about twenty people and engineers who have developed TPUs at the time (Tensor Process Units) at Google as responsible.
