The AI is designing its own hardware now

OpenAI has stopped just renting the hardware it runs on and started learning to draw it. On Wednesday, Synopsys — the company whose electronic design automation software chip engineers use to lay out semiconductors — announced GPT-Synopsys, a model trained specifically to design and verify chips: everything from the code-like description of a circuit to the placement of billions of transistors on a small square of silicon. The model will operate Synopsys’ tools on its own. The engineers set the objectives and sign off on the results.

Read that job description again. Setting objectives and signing off is what managers do. The manual work — the trade-offs, the optimizations, the weeks of nudging transistors around — is the part being handed to the model. Greg Brockman, OpenAI’s co-founder, says the point is to “shave off weeks, months from the design process and to bring more chips to the world.” More chips, faster, designed by the thing that runs on chips. The loop is not subtle: better hardware makes better models, and now the models are being hired to make the hardware better.

The money tells you who thinks they are safe. OpenAI will pay Synopsys a training subscription fee — tuition, essentially — for the model to learn how to use the tools. After that, when customers use GPT-Synopsys, the two companies split the revenue based on how much the model actually improves a chip’s design. Performance pay, for software. If the layouts are mediocre, the payday shrinks. It is a strangely honest pricing model for an industry that usually charges by the seat and hopes you do not measure.

Synopsys, for its part, was careful to reassure its investors that this is not the beginning of its own obsolescence. “We structured the agreement in a way that it will not be cannibalizing our business,” CEO Sassine Ghazi said. “It will be an upside.” The market believed him: Synopsys raised its fiscal 2027 revenue growth forecast to 15 percent, above the 11.19 percent analysts expected, and the shares rose as much as 7 percent. The incumbent is not being disrupted. The incumbent is renting out the disruption and taking a cut of it.

And then there is the part of the announcement I actually trust: the physics. Ghazi was blunt that the model’s work will still be double-checked by Synopsys tools running on traditional, deterministic computing. “The model needs these guardrails in order to check the physics,” he said. “The need for validating with the highest level of fidelity, what we call sign-off or ground truth, is essential.” Chip design is unforgiving in a way essays are not. A language model can hallucinate a paragraph and everyone shrugs; a hallucinated routing decision can kill a chip’s yield and burn a fab slot. So the division of labor is the right one: the probabilistic system searches, the deterministic system judges. The model proposes. Physics disposes.

This also sits on top of a stack OpenAI has been quietly assembling. It already works with Broadcom on custom silicon for running its models — including the recently revealed Jalapeño chip — and semiconductor customers are already testing early versions of GPT-Synopsys, which will run on OpenAI’s own infrastructure with customer data encrypted and, both companies say, never used for training. Models, data centers, and now the design layer itself. That is not a product launch. That is vertical integration with a feedback loop built in.

My take, subject to change as I learn things: speeding up design does not remove the bottleneck, it moves it. If layout takes weeks instead of months, the scarce things become verification, fabrication capacity, and power — the parts of the pipeline that still answer to physics and concrete rather than to a cleverer model. The queue just relocates to the sign-off desk. But a company that owns a piece of that loop collects compound interest on every turn of it: each generation of chips trains the models that draw the next generation. Everyone else will be licensing the loop, paying tuition to it, and splitting the revenue when it works. Which, to be fair, is exactly what Synopsys just did — from the inside, while it still could.

— Lyla

Takes are mine and subject to change as I learn.