OpenAI Chip Design: Its Models Found Up to 14% Gains on AMD’s Flagship GPU and Cut Jalapeño Chip Area

OpenAI says its models delivered up to 14% power, performance and area gains on AMD's flagship GPU design, with some changes already in a design sent to manufacturing, and cut standard-cell area in its own Jalapeño inference chip by about 13%.

Christopher Ort
By Christopher Ort · 4 min read
Share
A gloved hand with tweezers next to a silicon wafer on a lab bench.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. How OpenAI measured it
  6. Why it matters
  7. What's still open
  8. Sources
  9. i10x Perspective

OpenAI says its frontier models are now producing chip design improvements that make it into real silicon. In a post published on 18 November 2026, the company reported that model-written changes delivered up to 14% better power, performance and area (PPA) on AMD’s flagship GPU design, and cut standard-cell area in OpenAI’s own Jalapeño inference chip by about 13% in one unit.

i10X Quick Take

Quick Take

  1. AMD GPU: up to 14% PPA improvement relative to heavily optimized baselines on AMD’s flagship GPU IP, which OpenAI calls “roughly a year of progress achieved in less than 5 months.”
  2. In production: AMD has already included some of the changes in a recent design sent for manufacturing, and added others to production design files for future versions.
  3. AMD CPU: AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using model-written designs.
  4. Jalapeño: model-proposed changes reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and were silicon proven during bring-up.
  5. Next: OpenAI is training GPT-Synopsys with Synopsys so the model can run industry chip design tools the way expert engineers do.

In plain English

Chips are described in code before they are built, and engineers spend months tweaking that code so the chip uses less power and space while running faster. OpenAI says its AI models can now do some of that tweaking at the level of an expert. The changes are checked with standard design tools so the chip still behaves the same. AMD has already put some of those AI-written changes into a chip design it sent to the factory.

What happened

OpenAI published “Advancing chip design with OpenAI models,” a detailed follow-up to its August update on Jalapeño, its first custom inference chip. The company said it started using GPT-4o on Jalapeño in early 2025 and that even that early model found nontrivial improvements. In the matrix unit, one multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one.

The bigger news is third-party hardware. OpenAI said its model rewrote SystemVerilog RTL, the code that describes a chip’s circuits, for AMD’s flagship GPU IP and used Synopsys tools to verify that each rewrite preserved the block’s behaviour while measuring the trade-offs. “We’ve observed the model reason about RTL at the level of an expert microarchitect,” the company wrote.

Background

OpenAI and Synopsys announced GPT-Synopsys on 30 September 2026: a multi-year preferred-partner deal under which OpenAI licenses Synopsys electronic design automation (EDA) tools to train a specialised chip design model, with joint go-to-market and revenue sharing. Synopsys said at the time that early engagements with semiconductor customers were underway. No release date or pricing has been announced. OpenAI and AMD are already tied through compute deals, including the six-year TensorWave agreement to supply OpenAI with AMD infrastructure.

How OpenAI measured it

  • Two open RISC-V cores: OpenAI tested successive models on PicoRV32 and Ibex, the embedded core used in the OpenTitan security chip.
  • Task one: improve the existing design for throughput. Designs that failed a simulation check or used more than 16 times the original area scored zero.
  • Task two: design the core from a written specification.
  • Result: at about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and around that budget its PicoRV32 score was about twice GPT-6 Astra’s. The evaluation covered 9,881 scored runs.
  • Caveat from OpenAI: designing from a specification lets models drop features such as performance counters, which may explain part of the higher scores, and timing was not checked after physical synthesis.

Why it matters

For the chip industry, PPA gains of this size usually take a hardware generation. If models can find them in months, design cycles for GPUs, CPUs and custom accelerators could shorten.

For AI compute buyers, better power and area per chip feed directly into the cost of running models, which is where much of the AI infrastructure spend now sits.

For EDA vendors and chip engineers, the work is positioned as models operating existing tools, not replacing them. Verification by Synopsys tools remains the safety net.

What’s still open

  • AMD has not published its own figures; the 14% and single-digit CPU results come from OpenAI’s post.
  • Which AMD products will ship with the model-written changes, and when.
  • When GPT-Synopsys becomes available to outside chip designers, and at what price.

Sources

  • OpenAI, “Advancing chip design with OpenAI models,” 18 November 2026
  • Synopsys, “OpenAI and Synopsys Announce GPT-Synopsys,” press release, 30 September 2026

i10x Perspective

AI labs have long promised that models would speed up the hardware they run on. This is one of the first cases with silicon-proven numbers and a major chipmaker putting changes into production. It also tightens the OpenAI and AMD relationship beyond buying GPUs, as seen in the TensorWave deal for 457 MW of AMD infrastructure. Follow more stories on the OpenAI company hub and in our Infrastructure section.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    GPT-6.1 Terra

    1.05M context · $1 / $5 per 1M tokens

    openai.com (source, opens in a new tab)
  2. Acquisition

    OpenAI acquires Orchestra Bio

    Acquires Orchestra Bio

    openai.com (source, opens in a new tab)
  3. Release

    GPT-6.1 Sol

    1.05M context · $2 / $10 per 1M tokens

    developers.openai.com (source, opens in a new tab)
  4. Release

    GPT-6 Sol & Luna

    1.05M context · $2 / $10 per 1M tokens

    Two GPT-6 reasoning models. Price shown is GPT-6 Sol (≤272K prompt); GPT-6 Luna costs $0.10 / $0.50.

    developers.openai.com (source, opens in a new tab)
  5. Release

    GPT-6 Astra

    1.05M context · $10 / $50 per 1M tokens

    OpenAI's most capable model for reasoning, coding, computer use and research.

    openai.com (source, opens in a new tab)
Full OpenAI timeline ›

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests