OpenAI says its models did not just write code about chips. They rewrote the circuits inside AMD flagship GPU design, and AMD liked some of the rewrites enough to send them toward manufacturing.
In a post published on November 18, OpenAI reported gains of up to 14 percent in power, performance and area on AMD flagship GPU intellectual property, measured against heavily optimized baselines. OpenAI calls that roughly a year of progress achieved in less than five months. AMD has already included some of these changes in a recent design it sent for manufacturing, OpenAI said, and added others to its production design files for use in future versions.
The announcement
The work sits at the RTL level. RTL, short for register transfer level, is the code that describes what a chip circuit does before that circuit is laid out in physical form. AMD GPU code for this work is written in SystemVerilog, a language built for describing and checking hardware.
OpenAI said its model reasoned about that code at the level of an expert microarchitect, a designer who decides how a processor block is organized inside. The model rewrote circuits that perform calculations, control operations, or move data. For each rewrite, it used Synopsys tools to verify that the changed block still behaved as the original did, and to weigh the trade offs in power, performance and area. That trio is what chip teams shorten to PPA.
That checking step matters more than the rewriting step. A clever shortcut that changes the answer a circuit gives is not an improvement. It is a bug with good public relations. OpenAI account is that every rewrite had to pass the Synopsys verification flow before its PPA numbers counted.
Important details
OpenAI gave no breakdown of the 14 percent figure. The post does not say whether the best result was mostly power, mostly performance, or mostly area, and it does not name the GPU product or the blocks where the gains were found. Readers should not read up to 14 percent as 14 percent faster. It is the best reported PPA improvement on selected GPU intellectual property, as OpenAI describes it, not a whole chip result and not a measurement from finished AMD silicon.
The CPU result is smaller and, in one way, clearer. OpenAI said AMD verified single digit percentage gains in PPA across key blocks of its upcoming next generation flagship CPU using designs written by the model. That verification is attributed to AMD through OpenAI post. AMD had published no statement of its own on this work as of November 19 in the sources checked for this report, and AMD has not published its own figures for either the GPU or the CPU result.
| Detail | What OpenAI reported |
|---|---|
| GPU result | Up to 14 percent PPA improvement on flagship GPU intellectual property |
| Baseline | Heavily optimized baselines, in OpenAI wording |
| Time claim | Roughly a year of progress in less than five months, in OpenAI wording |
| Method | Model rewrote SystemVerilog RTL, Synopsys tools verified each rewrite |
| Production use | Some changes in a recent design sent for manufacturing, others in production design files for future versions |
| CPU result | Single digit percentage PPA gains across key blocks of the next flagship CPU, verified by AMD as reported by OpenAI |
| Product named | None |
| Breakdown of the 14 percent | Not given |

Where the method came from
The AMD work grew out of Jalapeño, the first custom inference chip OpenAI designed to run its own models. OpenAI said it started using an early frontier model to help design Jalapeño in early 2025, and that even that early model found nontrivial PPA improvements.
The gains sat in the parts of the chip that do the maths. In the vector unit, which manages and executes parallel computing tasks, savings came from floating point adders and multipliers, conversion functions, reduction logic, and the special function unit. In the matrix unit, one multiplier redesign used about 15 percent less cell area in synthesis while reducing latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15 percent while keeping its two cycle latency.
Across the broader effort, OpenAI estimates that model proposed changes reduced standard cell area by about 13 percent in the vector unit and 10 percent in the matrix unit. Those optimizations were silicon proven during Jalapeño bring up, the stage where a new chip is first powered on and checked in the lab. OpenAI first shared the Jalapeño story in August, when it said model help let the team go from initial design to tapeout in nine months. Tapeout is the point where a finished design is sent for manufacturing.
That history gives the AMD claim a useful anchor. The Jalapeño numbers are tied to a real chip that reached the lab. The AMD numbers, by contrast, are tied to design files and to one design sent for manufacturing. That is further than a benchmark demo and shorter than a product on sale that buyers can test themselves.
The benchmark, and its own caveat
To track how this skill changes from one model generation to the next, OpenAI also tested models on two open source 32 bit RISC V cores. PicoRV32 is a compact core designed for FPGAs, the reprogrammable chips used in many embedded systems. Ibex is the embedded processor used in the OpenTitan security chip.
Each model got the original hardware description and was asked to increase throughput, meaning the number of instructions the core can execute per second. OpenAI checked each design in simulation, scored designs at zero if they failed a check or used more than 16 times the original area, and measured the rest on workloads hidden from the models. Across both tasks, both cores, and all budgets, the full evaluation included 9,881 scored runs. At a comparable budget of about 112,000 tokens per run, GPT 6.2 Astra had the highest average score on both cores.
OpenAI then gave models written requirements instead of existing code, and every model scored higher that way. OpenAI is direct about why that result should be read with care. Starting from written requirements gives a model room to omit features found in the original cores, such as performance counters and additional instructions, and those features may do little to improve throughput on the test workloads. The company said requiring those features, setting a tighter limit on circuit area, and checking timing after physical synthesis would make the test more representative of production chip design.
That caveat does not undo the AMD result, which was measured against production designs rather than a benchmark core. It does show the difference between the two kinds of evidence in the same post. The benchmark is useful for comparing models with each other. The AMD and Jalapeño work is the evidence that matters for real silicon, and it is also the part with the fewest public details.

- Early 2025OpenAI starts using a frontier model to help design its Jalapeño inference chip
- October 6, 2025AMD and OpenAI announce a strategic partnership to deploy 6 gigawatts of AMD GPUs, starting with 1 gigawatt of Instinct MI450 systems in the second half of 2026
- August 2026OpenAI shares how models accelerated Jalapeño from initial design to tapeout in nine months
- September 30, 2026OpenAI and Synopsys announce GPT Synopsys, a model trained to use Synopsys design tools
- November 18, 2026OpenAI reports up to 14 percent PPA gains on AMD flagship GPU design and single digit gains on CPU blocks
What is next
The confirmed next step is a product, not just more research posts. OpenAI and Synopsys announced on September 30 that they will train GPT Synopsys, a specialized model built to use Synopsys electronic design automation tools the way expert chip designers do. Electronic design automation, shortened to EDA, is the software engineers use to design and verify chips. Under that multi year agreement, OpenAI licenses Synopsys EDA tools to develop the model, and the two companies will collaborate on research and development and on joint go to market work under a shared revenue framework. Synopsys said early technology engagements are already underway with leading semiconductor customers, and that GPT Synopsys will run on OpenAI hosted infrastructure and work with Synopsys own agentic platforms.
OpenAI also said it expects further PPA improvements on top of GPU and CPU designs relative to the most optimized baseline. That is an expectation, not a result, and this report treats it that way.
The open questions are concrete. Which AMD products will carry model written changes that buyers can identify? Will AMD publish its own account of the GPU and CPU gains, with a breakdown that OpenAI post does not give? And will the same method keep finding gains once the easy rewrites in a heavily optimized design have already been taken? A second round on the same blocks would tell readers more than a first round on a new block, because it would show whether this is a one time harvest or a repeatable step in the design flow.
There is also a wider race worth watching. Anthropic recently told investors that its Claude Mythos Preview model wrote almost all the low level code needed to get a model running on AMD GPUs in about three days, a claim we examined in Anthropic Claude Mythos Preview: AMD Kernels Claim. That work was about kernels, the software layer that runs a model on a chip. The OpenAI work reported here is one level deeper, about the circuits of the chip itself. Both point the same way. Models are moving from writing software that runs on hardware to shaping the hardware that runs the models, and the labs that sell the models are now also, in a small but real way, designing the silicon. Our earlier report today on Nvidia Halts Approved Chip Shipments To China shows the other side of that shift, where access to advanced chips remains a political and commercial fight even as the design of those chips starts to change hands.
Facts checked November 19, 2026, against OpenAI November 18 post and the Synopsys announcement below. The 14 percent GPU figure, the single digit CPU figure, and the year of progress estimate are OpenAI reported claims, attributed as such in this report. AMD had published no separate statement on this chip design work in the sources checked as of November 19.
Sources
- OpenAI, Advancing chip design with OpenAI models, November 18, 2026: https://openai.com/index/advancing-chip-design-with-our-models/
- Synopsys, OpenAI and Synopsys Announce GPT Synopsys, September 30, 2026: https://news.synopsys.com/2026-09-30-OpenAI-and-Synopsys-Announce-GPT-Synopsys-Frontier-Intelligence-to-Revolutionize-Chip-Design
- OpenAI, AMD and OpenAI announce strategic partnership to deploy 6 gigawatts of AMD GPUs, October 6, 2025: https://openai.com/index/openai-amd-strategic-partnership/
- Hero and body images: Dirty Static illustrations
END OF TRANSMISSION
Stay curious. Question everything.


