A big investment
The path to creating an adapted AI chip requires substantial sources. Experts from the industry said Reuters that designing a single version of such a processor could cost no less than $ 500 million, with extra costs for developing supporting software and hardware that possibly doubles that amount.
According to Reuters, the current OpenAi chip project, led by former Google chip designer Richard Ho, includes a team of 40 engineers who work with Broadcom on the processor design. The Taiwanese company TSMC, which also produces Nvidia's chips, will produce the chips from OpenAi with the help of its 3-nanometer process technology. The chips are said to record memory and network functions with high bandwidth that are comparable to those in the Processors of Nvidia.
Initially, the first chip of OpenAi will focus primarily on performing AI models (often called “inference”) instead of training them, with limited commitment throughout the company. The timeline suggests that the mass production could start at TSMC in 2026, although the first tape-out and production run are confronted with technical risks that extra solutions may require and the project can delay for months.
The switch from OpenAi to AI hardware is because large technology companies issue record amounts on the AI infrastructure. Microsoft is planning to invest $ 80 billion in 2025, while Meta has reserved $ 60 billion for the following year, Reuters notes. Last month, OpenAi (together with Softbank, Oracle and MGX) announced a new “Stargate” infrastructure project of $ 500 billion, aimed at building new AI data centers in the US.
