OpenAI said on November 18, 2026, that its models have moved from assisting chip design to shipping changes in production silicon, reporting gains of up to 14 percent in power, performance, and area on AMD's flagship GPU intellectual property and smaller but real reductions in the size of its own Jalapeño inference chip. The announcement matters because it is the first well-documented case of a frontier lab's models writing hardware description code that a major chipmaker included in a design sent for manufacturing, and because it shows how quickly the loop between model and machine is tightening.
From Research Demo to Production Silicon
OpenAI started using GPT-o3 mini to help design Jalapeño in early 2025, when the chip was still a blank-slate architecture project built with Broadcom. Even that early model found nontrivial improvements in the vector unit, which manages parallel computing tasks, and in the matrix unit, which handles arithmetic. The models explored alternative logic implementations and ways to split computations across pipeline stages, looking for chances to share hardware and cut intermediate storage.
The company estimates that model-proposed changes reduced standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit. One multiplier redesign used about 15 percent less cell area while cutting latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15 percent while keeping its two-cycle latency. OpenAI says these optimizations held up during Jalapeño's silicon bring-up, the point where paper gains either survive contact with manufactured hardware or disappear.
The AMD Results
The more significant claim involves a third-party vendor. Working with AMD, OpenAI reports up to 14 percent improvement in power, performance, and area relative to heavily optimized baselines on the flagship GPU intellectual property. The model found those gains by rewriting SystemVerilog RTL, the code that describes a chip's circuits, then using Synopsys tools to verify that each rewrite preserved the block's behavior while measuring the trade-offs.
OpenAI describes the result as roughly a year of progress achieved in less than five months. AMD has already included some of the changes in a recent design it sent for manufacturing and added others to production design files for future versions. AMD also verified single-digit percentage gains across key blocks of its upcoming next-generation flagship CPU using designs written by the model.
Those numbers come from OpenAI's account of the collaboration. AMD's own figures were not published alongside the post, and the 14 percent figure applies to selected circuits rather than the whole chip. The claim is still notable because a flagship GPU design carries years of human optimization, and because some of the output has already passed the only test that counts in semiconductors: it went to the fab.
Benchmarks on Open Cores
To measure progress from one model generation to the next, OpenAI tested five models on two open-source 32-bit RISC-V cores. PicoRV32 is a compact core for reprogrammable chips called FPGAs. Ibex serves as the embedded processor in the OpenTitan security chip. The models received the original hardware description and were asked to raise throughput, measured in instructions per second.
Designs scored zero if they failed a correctness check or used more than 16 times the original core's area. Across 9,881 scored runs, GPT-6.2 Astra posted the highest average score on both cores at a comparable budget of about 112,000 tokens per run, and GPT-6.1 Sol outscored GPT-5.6 Sol on both. On PicoRV32 at that budget, GPT-6.2 Astra scored about twice GPT-6 Astra.
OpenAI is candid about the limits of this test. Starting from a written specification rather than an existing core let the models drop features such as performance counters and extra instructions, which may explain part of the advantage. Requiring those features, tightening the area limit, and checking timing after physical synthesis would make the benchmark look more like production design work.
Why the Gains Are Real but Narrow
The honest reading is that models have become genuinely useful at a narrow, well-verified slice of chip design. They rewrite blocks of RTL, and independent EDA software confirms the rewritten block still behaves the same way. That is a real engineering workflow, not a demo. It is also a long way from a model designing a flagship processor on its own.
OpenAI's own chip chief, Richard Ho, said recently that he still does not fully trust AI design changes, so every one gets checked. That is the right posture and probably the permanent one for the next several years. Verification is the bottleneck, and verification is where the trusted tools sit.
What This Means for the Industry
The strategic signal is bigger than the benchmarks. OpenAI is training GPT-Synopsys with Synopsys, a partnership the two companies announced on September 30, 2026, that pairs OpenAI's models with Synopsys' electronic design automation tools under a revenue-sharing agreement. The goal is a specialized model that uses the same tools expert designers use, which would spread the capability beyond a handful of frontier labs.
For businesses watching AI costs, the chain runs in one direction: better models design better chips, better chips run models more cheaply, and cheaper inference lowers the price of every AI product built on top. The 14 percent figure will not show up in your API bill next quarter. The direction it points will.
If you are tracking how AI reshapes the hardware it runs on, more analysis and practical guides are published at Crowe Endeavors.
Sources
"OpenAI Models Accelerate Chip Design – Jalapeno and AMD GPU/CPU Achieve Up to 14% PPA Improvement" (AI Radio (YouTube), November 18, 2026) — https://www.youtube.com/watch?v=4N3rUaLCWCE
"Advancing chip design with OpenAI models" (OpenAI, November 18, 2026) — https://openai.com/index/advancing-chip-design-with-our-models/
"AI models deliver up to 14% chip design gains in AMD collaboration, OpenAI says" (ANI, November 19, 2026) — https://aninews.in/news/business/ai-models-deliver-up-to-14-chip-design-gains-in-amd-collaboration-openai-says20261119082527
"OpenAI says its AI found 14% gains in AMD's flagship GPU design" (The Next Web, November 18, 2026) — https://thenextweb.com/news/openai-models-amd-gpu-chip-design-14-percent