A year of progress, in OpenAI's telling
OpenAI's November 18 post describes its models doing a specific, unglamorous job: rewriting SystemVerilog RTL, the code that describes a chip's circuits, and checking each rewrite with Synopsys tools to confirm the block still behaved the same way and to weigh the trade-offs. On AMD's flagship GPU IP, OpenAI reports up to 14% improvement in PPA (power, performance and area) against baselines it describes as heavily optimized. It calls the result roughly a year of progress achieved in less than five months.
Two parts of that account are checkable in principle. OpenAI says AMD has already included some of the rewrites in a design sent for manufacturing, and added others to production design files for future versions. It also says AMD verified single-digit percentage PPA gains across key blocks of an upcoming next-generation flagship CPU using model-written designs. Reported adoption into a design bound for manufacturing is a stronger signal than a benchmark score, because it means the rewrites passed the checks a production flow applies. It is still OpenAI's report of AMD's decision, not AMD's own account of it.
Where the 14% needs a discount
Start with the two words in front of it. "Up to" marks a maximum, and OpenAI does not explain how the reported PPA improvement was calculated: no per-metric breakdown between power, performance and area, and no statement of which block produced the best figure or whether the gain is block-level or chip-wide. A 14% area saving on one block and a 14% chip-level gain are very different claims wearing the same number.
Then note who is speaking. The post is OpenAI's, and as MLQ points out, it carries no statement from AMD. The verification described in it, including the CPU gains, is OpenAI reporting what AMD did. The two companies are also commercial partners, not strangers: OpenAI holds rights to buy up to 160 million AMD shares under a 2025 supply deal, per MLQ. None of that makes the claim false. It does mean the strongest version of it currently rests on one partner's account of the other's engineering.
Jalapeño is the cleaner evidence
The more concrete results come from OpenAI's own silicon. OpenAI began using GPT-4o on Jalapeño, its first custom inference chip, in early 2025. It estimates model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one, and OpenAI says the optimizations were proven in silicon during the chip's bring-up.
Those percentages are estimates from the team that built both the model and the chip, and the multiplier figure was measured in synthesis rather than on finished silicon. What the bring-up establishes is narrower and still useful: the model-proposed optimizations were in the design that reached silicon, and the chip worked with them in it.
The benchmark, and its own footnote
To compare model generations, OpenAI ran five models on two open-source 32-bit RISC-V cores, PicoRV32 and OpenTitan's Ibex, across 9,881 scored runs spanning two tasks, both cores and all token budgets. In the task of optimizing an existing core's RTL, at a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. Every model also scored higher when designing a core from written requirements than when optimizing an existing one.
OpenAI then supplies the reason to be careful with that second result. A core designed from requirements could omit features the original had, such as performance counters and additional instructions, that did little for throughput on the test workloads, which may explain part of the score advantage. OpenAI says requiring those features, tightening the area limit and checking timing after physical synthesis would make the test look more like production design. That is the right caveat, and it is doing real work: the from-spec scores should not be quoted without it.
The follow-through is GPT-Synopsys, a model OpenAI and Synopsys are working to train to operate electronic design automation tools the way expert designers do. Synopsys says customer design data will not be used to train it, via Unite.AI's account of Synopsys's materials. If AI-written RTL is going to move from one partner's flagship project into ordinary design flows, that tooling, and the verification wrapped around it, is where the claim will be settled.
Sources
- [1] OpenAI, "Advancing chip design with OpenAI models," November 18, 2026 — the primary source, read first-hand in full: the up-to-14% PPA claim, the manufacturing and production-file adoption, the CPU verification account, the Jalapeño figures, the RISC-V evaluation and its caveat, GPT-Synopsys.Read source
- [2] Unite.AI, "OpenAI Reports Model-Driven Gains in AMD Flagship Chip Designs," November 18, 2026 — a detailed account of the post, read first-hand; agrees with the post on every figure; also the route for Synopsys's customer-data assurance.Read source
- [3] MLQ, "OpenAI says its models found up to 14% gains on AMD's flagship GPU design in under five months," November 18, 2026 — the observation that the post carries no AMD statement, and the 160 million AMD share-purchase rights under the 2025 supply deal.Read source