OpenAI says its AI improved AMD's next flagship chip designs by up to 14%. The chips haven't shipped yet
OpenAI says its AI models have found real improvements in chips designed by AMD, the main challenger to Nvidia. If the claim holds up, AI could start to speed up the slowest and most expensive part of the chip business: design.
In a research post published on November 18, OpenAI said its models improved AMD's flagship graphics chip designs by up to 14%. The measure combines power use, speed and the physical size of the circuits, which chipmakers call PPA. OpenAI said the work amounted to roughly a year of normal progress, done in under five months.
OpenAI said AMD has already put some of the changes into a design it has sent to manufacturing, and added others to files for future versions. OpenAI also said AMD had verified smaller, single-digit percentage gains in key parts of its next flagship CPU, the main processor in computers and servers.
What OpenAI's AI actually did
Chips start life as code that describes their circuits. OpenAI's models rewrote that code for AMD's designs. Software from Synopsys, the largest maker of chip design tools, then checked that each rewritten block still behaved exactly as before.
OpenAI also reported results on its own first chip, an AI processor called Jalapeño built with Broadcom. It estimated that changes proposed by its models shrank two key parts of the chip by about 13% and 10%. Unlike the AMD figures, OpenAI says these changes have been proven on working chips.
The post included a test of several generations of OpenAI models on two small, freely available processor designs. Its newest model, GPT-6.2 Astra, scored highest. OpenAI itself noted that one of the two tests was easier than real work, because models could leave out features the original designs had.
Why buyers and investors should read the fine print
Every headline number comes from OpenAI. Tom's Hardware and The Next Web both reported the figures as OpenAI's claims, and neither cited independent testing.
According to 24/7 Wall St., AMD has issued no formal statement. It reported that AMD's only public reaction was a post by chief executive Lisa Su, who quoted OpenAI co-founder Greg Brockman's note on the results and called them fantastic work by engineers at both companies.
The 14% is a maximum, not an average. OpenAI did not say how much came from lower power, higher speed or smaller size. Those matter differently to a buyer. A smaller circuit can mean cheaper chips; lower power can mean lower data center electricity bills.
Sending a design to manufacturing is also a long way from a product on sale. 24/7 Wall St. pointed out that no measured gains on finished AMD chips have been reported. It noted that AMD speaks at investor conferences hosted by UBS on December 1 and Barclays on December 9, and suggested that if neither talk mentions the work, investors treat the 14% figure as marketing. Tying the changes to a named product with a ship date would give the claim more weight, it said.
There is a conflict of interest to weigh. Under a deal announced in October 2025, OpenAI agreed to deploy six gigawatts of AMD chips and received a warrant, the right to buy up to 160 million AMD shares at one cent each. The shares unlock in stages as OpenAI's deployments grow and as AMD's stock hits price targets. That means OpenAI benefits when AMD looks strong.
According to 24/7 Wall St., AMD shares closed up 5.3% at $530.14 on November 18, the day of the post. It noted that no source has tied that move to OpenAI's research.
The bigger prize is selling the tool to everyone
The AMD results are partly a sales pitch for a product still in development. In late September, Synopsys and OpenAI agreed to build GPT-Synopsys, an AI model trained to run the design software chip engineers already use, Reuters reported.
The business terms are unusual. OpenAI pays Synopsys a fee to train on its tools. When Synopsys customers use the model, the two companies split revenue based on how much it improves a chip's design, according to Reuters. Neither company has disclosed dollar figures or a launch date.
That model would not be exclusive to AMD. 24/7 Wall St. argued that Nvidia and other chipmakers could gain the same edge once it is on sale, which would shrink any AMD advantage to months. On that view, Synopsys, which would earn from every customer, has the broader exposure.
OpenAI's own engineers are cautious. Richard Ho, who leads its hardware team, said in an interview with More Than Moore published September 30 that he still doesn't fully trust the models. Every AI-suggested change still goes through full checks, The Next Web reported.
Outside experts have given OpenAI some credit for speed. Speaking to IEEE Spectrum in September about Jalapeño, UC San Diego professor Andrew Kahng called its design pace likely best in class. David Chin, co-founder of chip design startup Verkor.io, said Broadcom's help was essential and doubted a team starting from scratch could match it.
For chip buyers and investors, the takeaway is narrower than the headline. AI is now producing design changes that a major chipmaker is willing to send to manufacturing. Whether those changes make AMD's next chips meaningfully cheaper or faster will only be known once they ship and are tested.
Written by AI from the public sources linked below.