Artificial intelligence is now designing the very hardware it runs on — and the results are starting to show up in real silicon. In a research update published on November 18, 2026, OpenAI revealed that its frontier models rewrote circuits in AMD’s flagship GPU designs, delivering up to 14% improvement in power, performance and area (PPA) against baselines AMD had already heavily optimised. The company says the work compressed roughly a year of chip-design progress into less than five months.
The headline number is striking, but the detail behind it is more important: AMD has already included some of the AI-proposed changes in a chip design it recently sent for manufacturing, and added others to its production design files for future versions. That is the difference between a lab demo and a design win. A model that can improve circuits on paper is interesting; a model whose rewrites survive a chipmaker’s own verification and make it into a tape-out is something else entirely.
The story also matters for what it says about where AI is heading next. The same models that write code and answer questions are now working at the register-transfer level — the layer of chip design where engineers describe exactly how circuits behave. If AI can do expert-level work there, the economics of building faster, more efficient chips could change for every company that designs silicon.
The 14% AMD Result, Explained
PPA — power, performance and area — is the three-way trade-off every chip designer lives by. A circuit can be made faster, but usually at the cost of more power or more silicon area. Improving all three at once, against a design that human experts had already squeezed hard, is genuinely difficult. OpenAI says its models achieved up to 14% PPA improvement relative to heavily optimised baselines on AMD’s flagship GPU intellectual property.
The model worked by rewriting SystemVerilog RTL, the code that describes a chip’s circuits before they are laid out physically. It rewrote circuits that perform calculations, control operations or move data, reasoning about the design the way an expert microarchitect would. Each rewrite was checked with Synopsys tools to confirm the changed block still behaved exactly like the original, then scored on its power, performance and area trade-offs.
OpenAI is careful about what the number does and does not mean. The company does not name the GPU or CPU products involved, and it does not split the 14% figure into separate power, performance and area deltas. So the result should not be read as a 14% faster graphics card. It is the best reported improvement in the combined PPA measure on selected GPU blocks — and, crucially, changes good enough that AMD put some of them into a design headed for manufacturing.
The CPU side of the partnership is moving too. AMD verified single-digit percentage PPA gains across key blocks of an upcoming next-generation flagship CPU, using designs written by OpenAI’s models. OpenAI says it expects further gains on both GPU and CPU relative to the most optimised baselines. For more on the wider race to build AI hardware, see our coverage of Microsoft’s Maia 300 AI chip (https://inspireviraltimes.com/microsoft-maia-300-ai-chip-tokens-per-megawatt/), which is designed to squeeze more tokens from every megawatt, and Cerebras’ disaggregated inference approach (https://inspireviraltimes.com/cerebras-disaggregated-inference-gpus-cs-3/), which splits AI inference across different kinds of chips.
How the Model Rewrites a Circuit
The workflow behind the gains is as interesting as the gains themselves. The model does not simply suggest a tweak and hope for the best. It proposes a rewrite of the circuit description, then verifies the new version against the old one using Synopsys electronic design automation tools — the same class of tools human chip designers rely on. Only rewrites that preserve the block’s behaviour exactly, while improving the PPA trade-off, survive.
That verification loop is what makes the approach trustworthy enough for production use. A wrong circuit in a shipped chip cannot be patched the way a software bug can; once silicon is manufactured, a logic error is permanent and enormously expensive. By grounding every proposal in formal checks, the model works inside the same safety net human engineers use, rather than around it.
The partnership goes further. OpenAI says it is working with Synopsys to train GPT-Synopsys, a new model built to use Synopsys’ design tools the way expert chip designers do. The goal is to move from a general-purpose model that can reason about circuits to a specialised one that speaks the language of chip design fluently — and can be trusted with more of the design flow. The same verification-first thinking shows up in the AI coding world, where Microsoft’s MAI Code 2 models (https://inspireviraltimes.com/microsoft-mai-code-2-cyber-2-flash-coding-ai-models/) are built to write and check code in the same loop.

Jalapeño: Where the Loop Started
The AMD work grew out of OpenAI’s own chip. Jalapeño, the company’s first custom inference processor, was the test bed. OpenAI began using GPT-o3 mini on the design in early 2025, and even that early model found meaningful improvements. In August 2026, the company said AI assistance had helped take Jalapeño from initial design to tape-out in nine months.
The November update fills in the engineering behind those gains. Across the broader design effort, OpenAI estimates that model-proposed changes reduced standard-cell area by about 13% in the vector unit, which manages and executes parallel computing tasks, and by about 10% in the matrix unit, the arithmetic heart of the chip. These were not simulations: OpenAI says the optimisations were proven in silicon when Jalapeño was brought up in the lab.
Two concrete examples show the scale of the wins. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% while keeping its two-cycle latency. Small blocks, multiplied across a whole chip, add up to real silicon savings — and real power savings, since smaller circuits generally burn less energy.

The Models Are Getting Better at Chip Design
To measure how each model generation performs at chip design, OpenAI tested its models on two open-source 32-bit RISC-V cores: PicoRV32, a compact core for FPGAs, and Ibex, the embedded processor in OpenTitan’s security chip. Each model was given the original hardware description and asked to increase throughput — the number of instructions the core executes per second.
The results show a clear progression. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol on both. On the existing PicoRV32 core, Astra’s score at that budget was about twice GPT-6 Astra’s. The full evaluation covered 9,881 scored runs across both tasks, both cores and all token budgets. Astra is also the model behind OpenAI’s latest releases, which we covered in our report on GPT-6.2 Astra and Sol (https://inspireviraltimes.com/openai-gpt-6-2-astra-sol-launch/).
A second task asked the models to design a core from written requirements rather than improving an existing one. All five models scored higher from a specification than from editing existing RTL — but OpenAI flags a caveat. Starting from requirements gives models room to omit features in the original cores, such as performance counters and extra instructions, which may do little for throughput on the test workloads. Requiring those features, setting a tighter area limit and checking timing after physical synthesis would make the test more representative of real chip design.
What This Means for the Chip Industry
The obvious question is whether AI-designed circuits stay a research curiosity or become how chips are actually built. The AMD result points toward the second answer. When a chipmaker takes model-written changes into a manufacturing tape-out, the technology has crossed from demonstration to production — and other chipmakers will be watching closely.
The economics are compelling. Chip design is slow and expensive precisely because it demands rare expertise: engineers who can hold an entire system in their head while optimising circuits transistor by transistor. If frontier models can do a meaningful share of that work — verified, not guessed — design cycles shorten, smaller teams can attempt more ambitious chips, and the cost of a new design falls. That matters for AI hardware most of all: the NVIDIA RTX Spark desktops (https://inspireviraltimes.com/nvidia-rtx-spark-compact-desktops-128gb-local-ai/) now putting 128GB of local AI compute on a desk are only possible because chip design keeps advancing.
There are limits to keep in mind. OpenAI’s figures are self-reported, AMD has published no statement of its own, and no finished AMD silicon carrying the changes has been measured yet. The account rests on OpenAI’s telling until AMD ships a product with the changes or talks about them itself. Even so, the direction of travel is clear: the models that run on chips are now helping to design them, and the chips they design will run the next generation of models. It is a feedback loop — and it is already spinning faster than anyone expected.
Sources: OpenAI, “Advancing chip design with OpenAI models” (November 18, 2026) — https://openai.com/index/advancing-chip-design-with-our-models/ ; Unite.AI, “OpenAI Reports Model-Driven Gains in AMD Flagship Chip Designs” — https://www.unite.ai/openai-reports-model-driven-gains-in-amd-flagship-chip-designs/

