Overview
On November 19 Beijing time, OpenAI released a technical report titled Advancing chip design with OpenAI models, detailing its progress in using AI models to assist chip design. The company said the models not only optimized circuits for its in-house AI inference chip Jalapeño, but also, in collaboration with AMD, delivered up to a 14% improvement in power, performance and area (PPA) metrics on key IP blocks of a flagship GPU. Part of the work has been incorporated into designs recently sent to fabrication by AMD, while other improvements are in production design files for future versions.
For its own chip, OpenAI began using GPT-4o in Jalapeño’s design in early 2025. In the vector compute unit, the models adjusted floating-point adders, multipliers, data-conversion logic and pipeline structure to reduce hardware resource use and intermediate data storage. In the matrix compute unit, they redesigned floating-point operations and low-precision dot-product circuits. One multiplier optimization cut the synthesized standard-cell area by about 15% while shortening latency from two clock cycles to one; an adder optimization achieved a similar 15% area reduction while holding latency at two cycles.
OpenAI estimates the model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, with the gains confirmed during silicon debug. In the AMD collaboration, the models rewrote SystemVerilog RTL describing the circuitry and improved PPA on already highly optimized designs — work OpenAI says accomplished in under five months what normally takes about a year.
Background & Interpretation
1. Technical Background and Evolution
AI-assisted chip design is not new; electronic design automation has long applied machine learning to discrete tasks such as place-and-route and yield prediction. What is distinctive here is that the models moved beyond helping engineers write code into the circuit and microarchitecture layer itself — directly rewriting RTL and restructuring adders and multipliers with quantifiable area and latency gains. This marks large models beginning to touch the most specialized, experience-dependent parts of the design flow.
2. Core Drivers and Underlying Mechanics
The underlying logic is that chip design contains vast, formally describable and automatically verifiable optimization space. Floating-point units, data-conversion logic and pipeline structures all have clear area-power-latency trade-offs, and models can efficiently search combinations that human engineers cannot exhaustively enumerate. Where objectives are quantifiable and verification automatable, the high-throughput trial-and-error capability of large models translates into engineering gains. That AMD still saw single-to-double-digit percentage gains on already optimized designs suggests the method also works for marginal improvements on mature designs.
3. Market Structure and Supply-Chain Impact
If these capabilities prove scalable, they could reshape R&D efficiency and competitive dynamics. On one hand, they may shorten design cycles and reduce reliance on senior designers, helping latecomers catch up on specific IP blocks. On the other, collaboration patterns among EDA toolchains, IP vendors and foundries may shift. The report also notes OpenAI applying the approach to next-generation flagship CPU blocks and open-source RISC-V processor design, indicating an ambition to platformize AI design capability rather than serve a single in-house chip.
Implications & Outlook
Near term, the results still require full validation through tape-out, mass production and yield, with real gains to be disclosed in future products. Over the medium term, whether AI-assisted design moves from case study to standard flow depends on reproducibility and portability across process nodes. Key variables include when AMD products using the design reach production, how transferable the gains are to other processes and architectures, and whether mainstream EDA vendors integrate similar capabilities.
Takeaways for Industry Participants
Chip design teams should pilot AI-assisted optimization on IP blocks with clear objectives and automated verification, treating model output as a complement to — not a replacement for — engineers. EDA and IP firms should focus on accumulating design data and building verification loops, which determine scalability. For domestic supply chains operating under compute constraints, design-efficiency gains are a soft but strategic lever worth incorporating into long-term efforts to shorten R&D cycles.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.