OpenAI is working on its own AI accelerator and plans to finalize the design this year. TSMC will bake the chips.
ChatGPT maker OpenAI wants to power future AI developments with its own chips. Right now, OpenAI, like just about every other AI giant, is very dependent on Nvidia. Nvidia’s accelerators are very powerful, but also pricey. Proprietary chips should reduce OpenAI’s dependence and strengthen its bargaining power with Nvidia.
OpenAI is said to have its design almost ready to send it to TSMC for manufacturing. With that, volume production is not yet underway, but if OpenAI meets its own deadlines, the chip could roll off the assembly line as early as 2026. That’s what Reuters knows on the authority of anonymous sources. Once an initial design goes to the factory, it takes on average about six months before mass production is an issue.
Microsoft and Meta
Microsoft also wants to build its own AI chips, and it does so under the name Athena. Not much has come of that plan as yet: Microsoft remains dependent on Nvidia. Meta has also been working on its own accelerator since 2023, but that company too has not yet been able to break away from Nvidia. Google and AWS are already enjoying the benefits of proprietary hardware, though.
Proprietary chips are key assets for companies looking to develop large LLMs. Today, Nvidia is the court supplier of the necessary hardware. The latest generation of Blackwell chips was supposed to succeed the powerful Hopper series, but Blackwell is facing problems. Moreover, Nvidia is charging heavily as far as its accelerators are concerned, even for the older Hopper generation. AMD wants to offer an alternative, but has yet to make a very big impact. Either way, at best customers can only hope that Nvidia’s monopoly becomes a duopoly.
Proprietary chips change the position of customers. OpenAI would become less dependent on Nvidia or AMD and its prices and production volumes. Moreover, with an alternative in hand, the company can negotiate prices better. Before that happens, OpenAI must first finish and implement its chip. In any case, that will take some time.