OpenAI Says Its Models Cut Up to 14% From AMD Designs
OpenAI published on 18 November 2026 a research post saying its models found up to 14% improvement in power, performance and area (PPA) on AMD's flagship GPU IP, measured against heavily optimized baselines. The model rewrote SystemVerilog RTL, the code that describes the chip's circuits, and used Synopsys tools to check that each rewrite kept the block's behavior. OpenAI calls the result roughly a year of progress in under five months. AMD has already put some changes into a recent design sent for manufacturing and added others to its production design files.
The Numbers, and Who Measured Them
Most figures in the post come from OpenAI, and only one is attributed to AMD's own checking. The same post describes work on Jalapeño, OpenAI's first custom inference chip built with Broadcom, where model-proposed changes cut standard-cell area and were, in OpenAI's words, silicon proven during bring-up.
| Block | Reported gain | Who stands behind it |
|---|---|---|
| AMD flagship GPU IP | Up to 14% PPA vs optimized baseline | OpenAI; some changes sent to manufacturing |
| AMD next flagship CPU, key blocks | Single-digit % PPA | Verified by AMD |
| Jalapeño vector unit | About 13% less standard-cell area | OpenAI estimate, silicon proven |
| Jalapeño matrix unit | About 10% less standard-cell area | OpenAI estimate, silicon proven |
| One multiplier redesign | About 15% less area, 2 cycles to 1 | OpenAI, in synthesis |
Separately, OpenAI and Synopsys announced on 30 September 2026 a multi-year partnership to build GPT-Synopsys, a model that operates Synopsys design tools. Synopsys chief executive Sassine Ghazi told Reuters its results will be rechecked with conventional Synopsys tools, which is the right control and also a sign the output is not yet trusted unchecked.
What a Buyer of AI Hardware Should Do
Treat a gain of this kind as a signal about design speed and not as a price cut you can bank today. A faster design loop can shorten the time between GPU generations and widen what second-source vendors such as AMD can offer, but a block-level saving of 14% does not translate into 14% lower system cost or power draw, and the post gives no such figure. If a vendor quotes AI-designed improvements in a tender, ask which baseline it beat, which block it applies to and who verified it. European buyers planning capacity for 2027 should ask suppliers whether the quoted generation includes these changes before they commit to a delivery date.
Read next: Capital and Labor Just United Against Data Center Bans | OpenAI Cuts Astra 20% as Sol Nearly Matches It



