Artificial intelligence is starting to design the chips that run it. THE November 18, 2026, OpenAI published a detailed report of work carried out with AMD: his models rewrote parts of circuits of flagship graphics processors from the manufacturer, with gains ranging up to 14%. Some of these modifications are already in a chip sent to manufacturing. A result which illustrates a dizzying loop: an AI which contributes to improving the hardware on which the next AIs will be trained.
What OpenAI did
A chip is not designed transistor by transistor. Engineers first describe its circuits in a hardware description languagehere the SystemVerilog, in a form called RTL (register transfer level). This code describes calculations, control operations and data movements. Specialized software then transforms it into physical plans ready to be engraved.
OpenAI models have worked at this level. They proposed rewriting circuit blocks, each modification then being verified by Synopsys toolsa leading chip design software company, to ensure that circuit behavior remained exactly the same. The same tools measured the effect of each change. The engineers could thus accept or reject the proposals block by block.
This systematic verification is the key to the method. A language model can be wrong or invented, and an error in an engraved circuit is extremely expensive: a defective chip can force a manufacturing process to take several months to restart. By only retaining modifications whose equivalence has been demonstrated by proven tools, OpenAI and AMD limit this risk. The model plays the role of an engineer who comes up with lots of ideas very quickly, and the verification software plays the role of a controller who doesn’t miss anything.
Gains on already highly optimized circuits
According to OpenAI, the modifications proposed by its models have improved up to 14% these three criteria on blocks of intellectual property from AMD’s flagship graphics processors, compared to versions that the manufacturer had already heavily optimized. OpenAI estimates that this work represents approximately one year of progress made in less than five months.
The results were quickly integrated:
- AMD has included some changes in a chip recently sent to manufacturing;
- others have been added to its design files for future versions;
- AMD has confirmed, according to OpenAI, single-digit earnings on key parts of its next flagship central processor.
This collaboration is part of the partnership established between the two groups in October 2025, which provides for the deployment of AMD graphics processors for OpenAI, as we reported when the alliance was announced at 6 gigawatts. According to OpenAI’s announcement at the time, the deployment of AMD Instinct MI450 GPUs for OpenAI was expected to begin in the second half of 2026. Improving the design of its own supplier’s chips therefore directly serves the interests of OpenAI, which seeks to get more computing power for every dollar and watt spent.
Jalapeño, the OpenAI chip designed with AI
OpenAI applied the same method to Jalapeñoits first in-house chip intended forinferencethat is to say the execution of the models once trained. AI-assisted work began in early 2025 with the o3-mini model, and the chip moved from design to manufacturing in nine monthsaccording to The Next Web.
| Jalapeño Element | Gain obtained thanks to OpenAI models |
|---|---|
| Vector unit | Around 13% less surface area |
| Matrix unit | Around 10% less surface area |
| A multiplier circuit | Approximately 15% less surface area, with reduced latency |
| An adder circuit | Around 15% less surface area, at equal latency |
Important point: these gains were confirmed on siliconwhen commissioning the chip. They therefore do not remain at the simulation stage.
More and more talented models
OpenAI also compared five generations of its models on the design of two small open-source processors, PicoRV32 and Ibex. Its most recent model, GPT-6.2 Astraobtained the best results, both for improving existing designs and for building a processor from written specifications. OpenAI recognizes, however, that this second exercise was easier, because the models could omit certain functions present in the originals.
The company also works with Synopsys on a model called GPT-Synopsystrained to directly use design software that engineers already use. It is not alone in this niche: start-ups like Architect Labs, which raised $24 million in June 2026, are targeting the same market.
What still needs to be checked
These results call for some caution. They are announced by OpenAIwhich has every interest in demonstrating the value of its models in the industry. According to The Next Web, OpenAI’s report does not name any AMD products and does not include a manufacturer’s statement: the 14% figures and gains on the central processor are reported by OpenAI alone.
The figure of 14% is also a maximum: it does not apply to an entire chip, but to certain blocks. An improvement of a few percent on an entire chip would nevertheless remain significant, as the margins for progress are narrow on circuits already optimized by teams of experts.
An accelerating loop
The stakes go beyond AMD and OpenAI. If AI models can improve the chips that drive them more quickly, each generation of hardware could arrive sooner and cost less, which in turn would accelerate advances in AI. This phenomenon ofself-acceleration is at the heart of industry debates: Anthropic itself warned in its prospectus that its AI could soon greatly accelerate its own research, and independent evaluator METR estimated that AI had almost doubled the pace of Anthropic’s progress.
For the semiconductor industry, where the design of a cutting-edge chip mobilizes hundreds of engineers for years, the promise is considerable: shortening cycles, reducing costs and exploring more avenues. For engineers in the sector, the profession could evolve towards the supervision and verification of AI proposals. The coming months will tell whether other manufacturers confirm comparable gains, and especially whether chips designed with the help of AI keep their promises once in production.

