OpenAI is advancing in a plan to reduce NVIDIA’s dependence (NVDA) for the supply of chips, internally developing its first microprocessor focused on artificial intelligence applications.
ChatgPT’s creator should finish the project of her first chip in the coming months and plans to send it for production by Taiwan Semiconductor Manufacturing, sources told the sources Reuters. The process of sending a first project to a chip factory is called “taping out“. Openai and TSMC declined to comment.
One tape-out Typical costs tens of millions of dollars and will take about six months to produce a finished chip, less than OpenAi pay substantially more to speed up manufacturing. There is no guarantee that the chip will work and a failure will require the company to diagnose the problem and repeat the step tape-out.
Within OpenAi, the AI mode training chip is seen as a strategic tool to strengthen OpenAi’s negotiation relationship with other chip suppliers, the sources said. After the initial processor, OpenAi engineers plan to develop increasingly advanced chips with wider features with each new version.
If the initial release occurs without problems, this will allow OpenAI to produce its first AI chip and potentially test an alternative to the expensive processors of NVIDIA later this year. OpenAi’s plan to send your project to TSMC this year demonstrates that the startup has made a rapid progress in the first project, a process that usually takes years.
Large technology companies, such as Microsoft and Meta, have been struggling to produce satisfactory chips, despite years of effort. The emergence of the Chinese AI deepseek startup has also raised questions about the need for fewer chips for the development of powerful models in the future.
Continues after advertising
The OpenAi chip is being designed by the company’s internal team led by Richard Ho, who has doubled in recent months to 40 people, in collaboration with Broadcom. Ho joined OpenAi over a year ago from Google, where he helped lead the custom of the search for the search giant.
The HO team is smaller than the efforts of large companies like Google or Amazon. A new chip design for an ambitious and large -scale program can cost $ 500 million for a single version of a chip, according to sources in the sector with knowledge of chip design budgets. These costs may double if necessary software and peripherals are included for the project.
General AI models developers, such as OpenAi, Google, and Meta, have shown that an increasing number of chips connected to data centers makes models smarter and, as a result, they have insatiable demand for chips.
Continues after advertising
The goal said it will invest $ 60 billion in AI infrastructure next year and Microsoft predicted disbursements of $ 80 billion in 2025. Currently, Nvidia’s chips are the most popular in the sector and have a market share of about 80% . OpenAi itself is participating in the $ 500 billion Stargate infrastructure program announced by US President Donald Trump last month.
But the increase in costs and dependence on a single supplier has led major customers to explore internal or external alternatives to Nvidia’s chips.
The OpenAI AI processor, although able to train and execute AI models, will initially be implemented on a limited scale, the sources said. The chip will have a limited function in the company’s infrastructure.
Continues after advertising
To develop as comprehensive an effort as Google or Amazon’s AI chip programs, OpenAi would have to hire hundreds of engineers.
TSMC is manufacturing OpenAi AI chip using advanced 3 nanome process technology. The chip features a systolic matrix architecture commonly used with high bandwidth memory (HBM) – also used by Nvidia in its chips – and broad network features, the sources said.
