At the start of the year, several major economic media revealed that OpenAI had begun to lay the foundations of a new strategy, with the objective of producing its own hardware to train its AI models. Ten months later, Sam Altman’s company seems to have made good progress; according to Reutersit now hopes to produce its first chips in the near future.
This initiative, initially revealed by Bloomberg and the Financial Times, is built on a simple observation: like all companies that have made machine learning their business, OpenAI is today largely dependent on Nvidia. The green giant dominates this market head and shoulders thanks to its GPUs specialized in high-performance computing, and simply has no serious competitor in this segment at the moment.
In the current context of the rush towards AI, all of Big Tech is competing for these products, to the point that Nvidia is no longer able to meet demand – a very annoying situation for these companies which are constantly seeking to strengthen their computing capacity. OpenAI therefore began to explore other avenues: sourcing from other companies, but also producing its own chips.
The contours of this strategy are beginning to become clearer in a new report from Reuters; the creator of ChatGPT is currently build close ties with several big names in the industry to continue moving forward.
To begin with, the news agency’s sources indicate that OpenAI has approached AMD. The Orange team has considerably strengthened its AI hardware catalog this year, and the American startup therefore plans to start using its chips in parallel with those of Nvidia.
Broadcom in the design and TSMC in the kitchen
But above all, it intends to launch the production of its own proprietary chips. For this, it counts on two major players in the semiconductor industry. The first is the American Broadcom, which specializes in the design and sale of many types of chips. According to Reuters, the two partners have already been working on the design of chips specifically designed for inference for several months. OpenAI has notably assembled a crack team made up of renowned experts. Reuters mentions in particular two engineers who played a decisive role in the creation of Tensor Processing Units, the chips specialized in machine learning from Google.
But Broadcom’s prerogatives end there, because the company does not have its own production infrastructure. To transform its concepts into real products, it relies on service providers like TSMC, the Taiwanese semiconductor giant which plays a hub role in this industry. The latter maintains privileged relationships with most of the hardware giants, who rely heavily on it for the manufacturing part. And according to Reuters, Broadcom played a matchmaker role in allowing OpenAI to secure the necessary production capacity from TSMC.
The foundations of the program appear to be well in place, and OpenAI apparently expects to see concrete results fairly quickly. Still according to Reuters, the first OpenAI proprietary chips should leave TSMC factories by 2026. Even if the exact timetable is still far from being set in stone, Sam Altman’s troops are therefore on a good path . OpenAI has never been closer to producing this hardware which, in theory, will allow it to increase its ability to train models while providing a degree of independence from Nvidia.
As this deadline approaches, it will also be interesting to see if other AI big names will follow OpenAI’s example, in which case the standoff that already exists in the software field could also spill over to the side hardware.
🟣 To not miss any news on the WorldOfSoftware, , .
