OpenAI says its AI redesigned parts of AMD’s top AI chip and found a year’s worth of gains in under five months OpenAI said on November 18 that its models found up to 14% improvement in power, performance and area on AMD's flagship GPU circuit designs relative to heavily optimized baselines, work OpenAI president Greg Brockman described as "roughly an additional year of progress on AMD's chip achieved in less than 5 months using our models." AMD has included some of the changes in a recent design sent for manufacturing and added others to production design files for future chips, and AMD chief executive Lisa Su said she was "Proud of these fantastic results." OpenAI said it used Synopsys software to verify each SystemVerilog RTL rewrite behaved identically, and that model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit of its own Jalapeño inference chip, confirmed on finished silicon. OpenAI president Greg Brockman left and AMD chief executive Lisa Su, in 2019 file photos. Image: TechCrunch, CC BY 2.0 Brockman, TechCrunch Disrupt 2019 ; Genevieve Martin, OLCF at ORNL, CC BY 2.0 Su, Oak Ridge National Laboratory 2019 OpenAI says its AI models squeezed roughly a year of chip-design progress out of AMD’s flagship GPU in less than five months, and that AMD has already put some of the changes into a chip it has sent to the factory. The company set out the results https://openai.com/index/advancing-chip-design-with-our-models/ on Wednesday, November 18, claiming gains of up to 14% on designs AMD had spent years refining. What OpenAI says its models did The work targeted PPA, chip designers’ shorthand for power, performance and area: how much energy a circuit uses, how fast it runs and how much silicon it takes up. Improving one usually costs another. OpenAI says its models found up to 14% improvement “relative to heavily optimized baselines on the flagship GPU IP”, meaning the circuit building blocks AMD reuses across its graphics chips. The model got there by rewriting the SystemVerilog RTL, the code that describes a chip’s circuits before they are turned into a physical layout. OpenAI says it used Synopsys software to check that each rewrite still behaved the same and to weigh the trade-offs. “We’ve observed the model reason about RTL at the level of an expert microarchitect,” the company writes. According to OpenAI, AMD has included some of the changes in “a recent design it sent for manufacturing”, the step chipmakers call tape-out, and added others to its production design files for future chips. AMD also verified single-digit percentage gains in key blocks of its upcoming flagship CPU, the post says. “Unprecedented results at the frontier of chip design” Greg Brockman, OpenAI’s president, made the headline claim in an X post https://x.com/gdb/status/2123130414323888463 that afternoon: roughly an additional year of progress on AMD’s chip achieved in less than 5 months using our models Greg Brockman, president, OpenAI He added that some of the results were in a design that had “recently been taped out” and called them “unprecedented results at the frontier of chip design”. AMD’s chief executive, Lisa Su, replied https://x.com/LisaSu/status/2123141323096261077 that she was “Proud of these fantastic results”, crediting engineers at both companies with “pushing the frontiers of chip design together”. Neither company has named the chip. OpenAI refers only to AMD’s “flagship GPU” and, in its closing section, its “next generation flagship GPU and CPU”. AMD’s current top data-centre GPU is the Instinct MI455X, which powers the first phase of the 6 gigawatt deal https://ir.amd.com/news-events/press-releases/detail/1260/amd-and-openai-announce-strategic-partnership OpenAI signed with AMD in October 2025, and in July the two companies said their work together was extending toward the MI500 generation https://newsroom.amd.com/news/aai-2026-openai-update/ . The same approach on OpenAI’s own chip OpenAI says it started using its models on chip design in early 2025, with o3-mini, on Jalapeño, its first custom inference chip, whose first results https://openai.com/index/jalapeno-first-results/ it shared in August. The team estimates that changes proposed by the models cut standard-cell area by about 13% in Jalapeño’s vector unit and 10% in its matrix unit, and says those savings were confirmed on the finished silicon. It also tested five generations of its models on two small open-source RISC-V processor cores, asking them to speed up the existing designs and to build new ones from written requirements. GPT-6.2 Astra, its newest, scored highest on both. OpenAI concedes the second test flattered the models, because they could leave out features such as performance counters that the original cores include. Next, OpenAI says it is training GPT-Synopsys, the model it announced with Synopsys in September https://madrobot.blog/2026/09/30/openai-synopsys-gpt-synopsys-ai-model-chip-design/ , to use Synopsys’s chip-design software the way expert engineers do. What hasn’t been shown Every figure comes from OpenAI, and AMD has not issued a statement of its own beyond Su’s post. Neither company has published the redesigned circuits, which blocks were improved or how the 14% splits between power, performance and area, and “up to” describes the best result, not an average. The “year of progress” is OpenAI’s own estimate, with no published baseline for how long AMD’s engineers would have taken. Unlike Jalapeño, OpenAI has not reported measurements from finished AMD silicon, and a taped-out design is still some way from a chip on sale. The post landed the same day as OpenAI’s report that its AI agents now finish nearly half https://madrobot.blog/2026/11/22/openai-agents-half-research-tasks-jagged-frontier/ of its researchers’ multi-hour tasks, and a week after Anthropic said its Mythos model got Claude running on AMD chips in three days https://madrobot.blog/2026/11/11/anthropic-mythos-claude-amd-chips-three-days/ . Why it matters AI labs are now pointing their models at the hardware their models run on, and AMD is the first big chipmaker to say it is shipping some of that work. If the gains hold up in silicon, better models could mean faster chip upgrades, which in turn train better models. For now, the evidence is OpenAI’s word and one taped-out design.