OpenAI, the company behind ChatGPT, is considering entering the world of custom artificial intelligence (AI) chips, according to sources familiar with the matter. Although the company has not yet made a final decision, it has been exploring various options to address the shortage of expensive AI chips it relies on. These options include building its own AI chip, collaborating more closely with chipmakers like Nvidia, and expanding its supplier base beyond Nvidia.
OpenAI’s CEO, Sam Altman, has identified the acquisition of more AI chips as a top priority for the company. Altman has publicly expressed frustration with the scarcity and high costs associated with the graphics processing units (GPUs) that power OpenAI’s software. Nvidia currently dominates the GPU market, controlling over 80% of it.
To tackle these challenges, OpenAI has been developing its generative AI technologies on a supercomputer built by Microsoft, with the assistance of 10,000 Nvidia GPUs. However, this setup is extremely costly. According to an analysis by Bernstein analyst Stacy Rasgon, each query with ChatGPT amounts to approximately 4 cents. If ChatGPT scales to a similar level as Google search, it would require billions of dollars’ worth of GPUs annually.
Building its own AI chips would place OpenAI in a small group of major tech players, such as Google and Amazon, that design their own chips. However, developing custom chips would be a sizable and risky investment, potentially costing hundreds of millions of dollars annually. Even if OpenAI commits to the task, success is not guaranteed.
Alternatively, OpenAI has explored the possibility of acquiring a chip company to expedite the process. Similar acquisitions have been made by companies like Amazon, which acquired Annapurna Labs in 2015 to develop its own chips.
However, even if OpenAI proceeds with its custom chip plans, it could take several years to bring them to fruition. During this time, the company would remain dependent on commercial providers like Nvidia and Advanced Micro Devices.
While some big tech companies have attempted to develop their own processors, they have encountered challenges along the way. Meta, for example, faced issues with its custom chip efforts and had to scrap some AI chips. It is now developing a newer chip that can handle various types of AI work. Microsoft, OpenAI’s main backer, is also reportedly developing a custom AI chip for OpenAI to test, indicating a potential distancing between the two companies.
The demand for specialized AI chips has surged since the launch of ChatGPT. These chips are crucial for training and running the latest generative AI technology. Nvidia remains one of the few chipmakers that produce AI chips suitable for these applications and currently dominates the market.
In conclusion, while OpenAI is considering venturing into the domain of custom AI chips, it has not yet made a final decision. Building its own chips or acquiring a chip company are among the options being explored. However, the journey toward custom chips could take several years, and in the meantime, OpenAI would continue to depend on existing commercial providers.