OpenAI chip design work has produced what OpenAI describes as up to 14 percent power, performance and area improvement on AMD’s flagship GPU intellectual property, measured against baselines that AMD had already optimized heavily. The company said on November 18, 2026 that the model rewrote SystemVerilog RTL, using Synopsys tools to verify behavior and compare design trade-offs. The result is significant because it moves AI-assisted engineering beyond code generation and into measured circuit optimization.
OpenAI chip design: From model-written RTL to manufacturing
OpenAI said the GPU project represents roughly a year of progress achieved in less than five months. AMD has included some of the model-written changes in a recent design sent for manufacturing and added others to its production design files for future versions, according to OpenAI. The company also said AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using designs written by the model.
The central technique was a rewrite of SystemVerilog RTL, the code that describes a chip’s circuits. OpenAI said its model used Synopsys tools to check that every rewrite preserved the behavior of the affected block, then evaluated the resulting trade-offs in power, performance and area. Alphasignal reported that this verification step is essential because PPA objectives are coupled. A smaller design may consume less power but miss its timing target, while a faster design can require more energy or silicon.
The reported 14 percent figure requires careful interpretation. Alphasignal said it is a block-level PPA measure, with no stated equivalent for whole-GPU performance. MLQ reported that the post did not carry a statement from AMD and that OpenAI did not break the maximum down by power, speed or area. The available evidence therefore supports OpenAI’s claim of up to 14 percent improvement in the measured block, but it does not establish a corresponding gain across an entire GPU.
Jalapeño provides the silicon evidence
OpenAI’s own Jalapeño inference chip offers a separate set of results with stronger physical validation. The company began using GPT-4o to help design the chip in early 2025. Across the broader design effort, OpenAI’s team estimates that model-proposed changes reduced standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit. OpenAI said those optimizations were silicon proven during Jalapeño’s bring-up.
One multiplier redesign in the matrix unit 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 retaining its two-cycle latency. Unite.AI reported that the August 25, 2026 silicon results said AI assistance helped Jalapeño move from initial design to tapeout in nine months.
These Jalapeño figures relate to defined arithmetic and compute blocks. Alphasignal said OpenAI attributes the model’s role to block optimization and makes no claim that it generated the chip’s overall architecture. The commercial implications nevertheless extend beyond OpenAI. Chip design is a specialized process in which small improvements across repeated blocks can alter silicon cost, power consumption and performance, provided engineers retain control of functional requirements, timing and manufacturing constraints.
Benchmark results favor newer models
OpenAI tested five models on two open-source 32-bit RISC-V cores, PicoRV32 and the Ibex core used in OpenTitan. The evaluation covered optimization of existing RTL and specification-to-RTL design, with 9,881 scored runs across both tasks, both cores and all token budgets. Designs scored zero if they failed a correctness check or used more than 16 times the original core’s area.
At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, while GPT-6.1 Sol also outscored GPT-5.6 Sol on both. On the existing PicoRV32 core, GPT-6.2 Astra’s score near 112,000 tokens was about twice GPT-6 Astra’s. In specification-to-RTL work, GPT-6.2 Astra recorded a mean score above 0.7 on PicoRV32 and averaged about 0.75 on Ibex.
All five models scored higher when designing a core from written requirements than when improving the existing design on both cores. OpenAI itself qualifies that result, noting that a model can omit features found in the original cores, including performance counters and additional instructions. It said requiring those features, setting a tighter circuit-area limit and checking timing after physical synthesis would make the test more representative of production chip design.
GPT-Synopsys targets the wider workflow
OpenAI is partnering with Synopsys to train GPT-Synopsys, a model designed to use Synopsys electronic design automation tools in the way expert chip designers do. Unite.AI reported that the partnership builds on a multi-year agreement announced on September 30, 2026 covering joint development, go-to-market collaboration and revenue sharing. It said the intended workflow ranges from PPA optimization to timing and verification closure, with agents operating tools, interpreting results, implementing changes and iterating toward verified outcomes for engineer review.
Synopsys said customer design data would not be used to train GPT-Synopsys. Unite.AI reported that the data would be encrypted at rest and in transit, with configurable retention, audit and permission controls. Those protections matter because chip designs are proprietary and because automated optimization must remain accountable to conventional verification. The proposed partnership therefore extends the model from producing candidate RTL to interacting with the broader toolchain.
AMD relationship adds strategic weight
The hardware collaboration also sits inside a wider commercial arrangement. MLQ reported that under a deal announced in October 2025, OpenAI agreed to use six gigawatts of AMD GPUs and received the right to buy up to 160 million AMD shares at a low fixed price. MLQ said that purchase right could amount to a stake of roughly 10 percent, with vesting tied to deployment milestones and AMD share-price targets.
The immediate technical claim is still narrower than the business arrangement. OpenAI reported up to 14 percent block-level PPA improvement on AMD flagship GPU IP, and AMD reportedly verified single-digit percentage gains in key CPU blocks. The industry significance comes from the workflow: a frontier model rewrites RTL, established tools check equivalence, and AMD incorporates selected changes into designs that progress toward manufacturing. As OpenAI chip design moves from claims toward manufacturable products, the critical tests will be independent confirmation, disclosure of the PPA components and whether the same method generalizes across more blocks without degrading timing, functionality or manufacturability.

