November 19, 2026, (Inside AI) ā OpenAI's frontier models are now making measurable improvements to the chips that power AI, according to a company update. The models have reduced standard-cell area by about 13% in the warp engine and 10% in the matrix unit of OpenAI's first custom inference chip, JalapeƱo. They have also delivered up to 14% improvement in power, performance, and area (PPA) on AMD's flagship GPU IP relative to heavily optimized baselines.
The results, which OpenAI says were silicon-proven during JalapeƱo's bring-up, represent roughly a year of chip design progress achieved in less than five months. AMD has already included some of the model-generated changes in a recent design sent for manufacturing and added others to production design files for future versions.
The news signals a shift in how AI is being applied to hardware engineering. Instead of merely assisting with routine tasks, frontier models are now reasoning about circuit design at the level of an expert microarchitect, rewriting SystemVerilog RTL and verifying behavior with Synopsys tools. This could compress design cycles and lower barriers for custom silicon, but it also raises questions about verification rigor and the changing role of human engineers.
From JalapeƱo To AMD: Model-Driven PPA Gains
OpenAI began using GPT-4o to help design JalapeƱo in early 2025. Even that early model found nontrivial PPA improvements. In the warp engine, savings came from floating-point adders and multipliers, conversion functions, reduction logic, and the special-function unit. The models explored alternative logic implementations and how to divide computations across pipeline stages, finding opportunities to share hardware and reduce intermediate storage.
Read: SMIC: AI Chip Supply Chain to Stay Tight Through 2027
In the matrix unit, improvements accumulated across floating-point adders, multipliers, and low-precision dot-product circuits. One multiplier redesign used about 15% less cell area in synthesis while reducing latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while retaining its two-cycle latency.
The AMD collaboration extended this approach to third-party hardware. On AMD's flagship GPU IP, the model identified further improvements by rewriting the code that describes the chip's circuits. AMD also verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU using designs written by OpenAI's model.
"We've observed the model reason about RTL at the level of an expert microarchitect," OpenAI stated in the update. The model can rewrite circuits that perform calculations, control operations, or move data, and it used Synopsys tools to verify that each rewrite preserved the block's behavior and to evaluate trade-offs in power, performance, and area.
OpenAI is also working with Synopsys to integrate frontier intelligence with electronic design automation (EDA) tools that engineers already use. The partnership aims to train GPT-Synopsys, a new model designed to use Synopsys EDA tools the way expert chip designers do.
The broader implication is that AI models are becoming active participants in the design loop, not just passive assistants. If models can reliably propose and verify RTL changes that pass manufacturing checks, the economics of chip design could shift. Smaller teams might tackle custom silicon projects that previously required large engineering organizations. However, the approach still depends on human oversight and robust verification flows, and the long-term reliability of model-generated designs across diverse workloads remains an open question.
RISC-V Tests Reveal Model Progression
To measure model capability, OpenAI evaluated successive generations on two open-source 32-bit RISC-V cores: PicoRV32, a compact core for FPGAs, and OpenTitan's Ibex embedded processor. Each model received the original hardware description and was asked to increase throughput, measured as instructions per second. Designs scored zero if they failed simulation checks or used more than 16 times the original core's area.
Read: UK AI Chip Startup Fractile Hits $7 Billion Valuation with $720 Million Raise
With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. GPT-6.1 Sol also outscored GPT-5.6 Sol on both. The full evaluation included 9,881 scored runs across both tasks, both cores, and all token budgets.
In a second task, models designed a RISC-V core from written requirements and templates. All five models scored higher when designing from scratch than when optimizing an existing design. Starting from written requirements gives models room to omit features like performance counters and additional instructions, which may explain part of the score advantage. OpenAI acknowledged that requiring those features, setting a tighter area limit, and checking timing after physical synthesis would make the test more representative of production chip design.
The RISC-V evaluations suggest that model-driven design is improving generation over generation. But the gap between benchmark scores and production readiness remains significant. OpenAI's own caveats point to the need for more rigorous, production-grade benchmarks before these results can be generalized.
Looking ahead, OpenAI expects further PPA improvements on top of GPU and CPU relative to the most optimized baseline. The Synopsys partnership and the GPT-Synopsys model could broaden the design challenges addressed. For now, the JalapeƱo and AMD results offer a concrete demonstration that frontier models can contribute net new improvements to sophisticated production designs, not just toy problems.
If a research paper dropped this week, Mahesh has already read it. He covers AI tools and vibe coding for builders who need to know what is worth their time and what is not.