Magazine
OpenAI no longer wants to depend on NVIDIA: the company is building its own AI chips

Technology

OpenAI no longer wants to depend on NVIDIA: the company is building its own AI chips

OpenAI is developing its own AI chip, Jalapeño, together with Broadcom and TSMC. The company wants to reduce its dependence on NVIDIA and gain greater control over AI infrastructure.

· Updated · 3 min read

Translated with AI. Original: Українська

OpenAI is starting to play by NVIDIA's rules. The company is developing its own processors for artificial intelligence and has already shown how far it has come in this race. OpenAI's first custom chip, called Jalapeño, has reached tape-out, the moment when a processor design is finalized and handed to a manufacturer for production. According to OpenAI Chief Financial Officer Sara Friar, the company managed to reach this stage in less than nine months.

For OpenAI, this is much more than a technical experiment. Today ChatGPT runs on a huge amount of computing infrastructure, and the AI chip market is largely controlled by NVIDIA. Its processors have become one of the main resources of the global AI race. A custom chip gives OpenAI the ability to gradually control not only the model users see, but also part of the technology chain beneath it.

Jalapeño is being developed by OpenAI together with Broadcom and is designed for AI inference, that is, processing requests and generating responses with already trained models. TSMC is to manufacture the processors. Back in June, OpenAI confirmed the development of its first custom AI chip, and now the company has revealed new details about the pace of its creation.

Interestingly, OpenAI used its own AI models to design the chip itself. In other words, artificial intelligence effectively helped build the infrastructure on which the next generation of AI will run. For the technology industry, this is an important shift: AI is gradually ceasing to be just a software product and is beginning to influence the physical foundation of its own ecosystem.

At the same time, OpenAI is strengthening its cooperation with Samsung Electronics on developing the next generation of semiconductors. The company said that joint work on chips has become one of the areas of cooperation where the greatest progress has already been achieved. Samsung and SK Hynix also intend to supply memory for OpenAI's large-scale Stargate project.

The reason for this race is obvious: the cost of AI computing is becoming one of the main factors of business. OpenAI is simultaneously trying to make its models cheaper and to grow its corporate segment. According to the company, the number of users of its AI coder Codex has already reached 25 million, and corporate revenue rose by 32% in the period from June to July.

Custom chips could give OpenAI more control over performance, cost and scaling. But this also means the company is entering a territory where NVIDIA, AMD, Broadcom, Google, Amazon and Qualcomm have already gathered. Qualcomm alone this week signed a deal with Amazon that provides for the possibility of purchasing up to $60 billion worth of AI chips and related products.

So the story of Jalapeño is not just about a new processor. OpenAI is gradually transforming from a company that creates the best-known AI models into a company that wants to control a growing share of the infrastructure around them. If this plan succeeds, the next stage of competition in artificial intelligence will no longer take place only between models. It will take place between chips, data centers, memory and energy, that is, between those who control the physical foundation of the AI economy.

And this is where OpenAI is directly entering a field that for years was considered NVIDIA's territory.

Read also: $13 billion for the future of AI: NVIDIA buys Hugging Face and steps behind the scenes of artificial intelligence

Text: Arina Blyzniuk

Comments

No comments yet — be the first.

Your comment appears after the editors review it. No links or phone numbers. No email needed.

Discuss on Telegram

Read also