OpenAI plans to build AI chips on TSMC’s forthcoming 1.6 nm process node.

OpenAI, the developer of ChatGPT, has been mulling over the idea of building its own AI chip for some time. However, it appears that the project will be moving forward as United Daily News reported the company is now paying TSMC to produce the new chips. OpenAI has booked production slots on the 1.6 nm A16 process node, rather than its current N4 and N3 process nodes.

The report from UDN, via Wccftech, doesn't provide any evidence for this claim. However, the Taiwanese media agency is usually pretty accurate with tech forecasts. OpenAI is currently spending a lot of money on ChatGPT due to the high cost of Nvidia AI servers.

Nvidia's superchips and Hopper H100 dominate the industry. Alphabet, Amazon and Microsoft have all spent hundreds of millions on Nvidia hardware. The cost of developing and designing a competitive AI chip may be the same, but once you have the product, the ongoing costs will be much lower.

UDN reports that OpenAI originally planned to use TSMC’s relatively low-cost process node N5 to manufacture its AI chips. However, that plan has been abandoned in favor of a system still in development - A16 will be the next-generation N2, which is not yet being mass-produced.

TSMC claims that A16 is a 1.6 nm process node, but the number is now meaningless. It will use the gate-all-around nanosheet transistors (GAAFET) as N2, but will be the first TSMC Node to employ Super Power Rail, a backside power delivery system.

Why would OpenAI use something that is still a few more years away from being ready to be ordered in bulk? UDN's report indicates that OpenAI has approached Broadcom to handle the development for the AI chips, but neither company has any experience with TSMC cutting-edge nodes as far as I am aware.

It is possible that Apple is collaborating on the project, as it uses ChatGPT for its AI system. This is currently powered by Google's AI servers, but given that Apply is increasingly using its own technology, I wouldn’t be surprised if they were also looking to develop a new AI chip.

OpenAI's future is uncertain with mounting losses and a crowded AI market. Rumours of Apple and Nvidia investment may help turn the tide. Its stock value is maintained by the regular infusion of millions of dollars.

If OpenAI is eventually bought by Microsoft, Meta or even Nvidia, then it's unlikely the OpenAI chip will ever be completed, as Nvidia would not want to lose any sales.

Even if the chip is developed, it won't be able to compete with other ASICs in terms of performance. ASICs are designed to perform one task very well, but that's all they're good for. OpenAI's chip may be great for OpenAI, but few other companies will be interested.