Here is a sentence I did not expect to write this year: an AI model redesigned the chip it runs on, and the result is already heading to a factory. On 18 November, OpenAI published a technical report, Advancing chip design with OpenAI models, showing its frontier models rewriting the circuit code of real production chips and improving them by amounts that would normally take human engineers about a year.
The headline numbers
The biggest result came from AMD's flagship GPU. OpenAI's models rewrote SystemVerilog RTL, the code that describes a chip's actual logic, and delivered up to 14 percent better power, performance and area, the combined measure chip engineers call PPA, against a baseline AMD had already heavily optimised. OpenAI describes that as roughly a year of design progress achieved in less than five months. AMD has already folded some of the model-written changes into a design sent for manufacturing, with the rest written into production design files for future products.
- AMD flagship GPU: up to 14 percent PPA improvement on selected circuits, against heavily optimised baselines.
- AMD next-generation flagship CPU: AMD verified single-digit percentage PPA gains across key blocks using model-written designs.
- OpenAI's own Jalapeño inference chip: model-proposed changes cut standard-cell area by about 13 percent in the vector unit and about 10 percent in the matrix unit, proven in silicon during bring-up.
- One multiplier redesign used about 15 percent less cell area while cutting latency from two clock cycles to one.
- Jalapeño went from initial design to tapeout in nine months with AI assistance, starting with GPT-4o in early 2025.
- 9,881 scored tests across two open-source 32-bit RISC-V cores, PicoRV32 and Ibex, with GPT-6.2 Astra scoring highest on both.
Every rewrite was checked with Synopsys electronic design automation tools to confirm the circuit still behaved exactly as specified, and to measure the trade-offs. OpenAI also announced a partnership with Synopsys to train a new model called GPT-Synopsys, designed to use those tools the way expert chip designers do.
The angle nobody is spelling out: cheaper chips mean cheaper AI for everyone else
Most coverage of this story treats it as an engineering curiosity, but the economics are the real story. Chips are the single biggest cost behind every AI model you use. If AI can squeeze 14 percent more performance out of the same silicon, or the same performance out of a smaller chip, then every model gets cheaper to run, and the savings eventually reach the people paying per token.
Think about what that means for a small team. I wrote this week about OpenAI GPT-6.2 Astra and Sol, where the flagship got 20 percent cheaper and the $2 Sol model nearly matched it on coding. That price war is fought in dollars per million tokens, but it is won or lost in silicon. A model that designs better chips is a model that lowers its own running costs, and that is a flywheel: cheaper inference, more users, more data, better models, better chips again. The companies that own both sides of that loop, the model and the chip, will be very hard to catch.
There is also a supply angle that matters far beyond Silicon Valley. Chips are the choke point of the whole AI race, and I covered this week how Nvidia just stopped shipping AI chips to China, even the ones Washington approved. When the best hardware is scarce and political, anything that gets more work out of each chip matters. For users in Pakistan, where the newest GPUs are expensive and hard to get, a 14 percent efficiency gain is not an abstract number: it is the difference between a model being affordable to run locally or only available through an expensive foreign API.
The honest caveats
OpenAI is careful about the limits, and so should we be. The company gave no breakdown of the 14 percent figure between power, speed and area, named no AMD product, and reported no measurements from finished AMD silicon. All the figures are self-reported by OpenAI, and AMD has not published its own numbers. The RISC-V benchmark also flatters the models a little: OpenAI admits that designing from a written specification lets models drop features such as performance counters, which may explain part of the higher scores, and timing was not checked after physical synthesis.
And this capability is not coming to your laptop any time soon. The post says nothing about whether any of this is available outside OpenAI and its partners, and GPT-Synopsys is a joint project with one EDA giant, not an open tool. For now, this is a story about what the biggest labs can do, not what a freelancer can download.
Bottom line
AI writing code was the first shock. AI designing the silicon that runs the code is the second, and it may matter more. A 14 percent gain on a flagship GPU, achieved in under five months and already heading to manufacturing, is not a demo. It is the start of a loop where better models build better chips that run better models, and the price of intelligence keeps falling for everyone downstream.
Sources: OpenAI, Advancing chip design with OpenAI models, 18 November 2026; figures as reported by Unite.AI and i10X. Note: the direct openai.com URL for this report was not returned by search, so no OpenAI link is given here; all figures are OpenAI self-reported and AMD has published no separate figures.
