Analog Chip Design Was Tech's Last 'Dark Art.' AI Just Found a Way In

There’s an old joke in chip design: digital is engineering, analog is black magic. For decades, digital chips have been built by software that places and wires hundreds of millions of transistors automatically. Analog and RF (radio frequency) chips stayed stubbornly different — a domain where a small priesthood of veteran engineers shaped circuits by intuition and feel. Now AI is finally pushing into that last sanctuary. Here’s why that’s a bigger deal than it sounds.

Why analog and RF resisted automation for so long

Start with the basics. Most chips you use are digital — a world of ones and zeros, on and off. Here, electronic design automation (EDA) tools are extraordinarily mature. An engineer describes the desired behavior almost like writing code, and the tool figures out how to lay down and connect all those transistors.

Analog is the hard part. The clearest example is the RFIC — the radio-frequency integrated circuit that handles your phone’s Wi-Fi, 5G modem, and Bluetooth. Here, signals aren’t clean ones and zeros. They’re continuous waveforms, and those waveforms are absurdly sensitive to their surroundings.

Make one wire a single micrometer longer and the signal shifts. Electromagnetic interference leaking from an adjacent circuit can wreck performance. The same blueprint comes out slightly different every time a fab manufactures it. With hundreds of these variables tangled together, there’s no clean formula to solve. So designers leaned on something else: instinct earned through experience.

Why “dark art” isn’t just a flex

When the industry calls RF design a dark art, it’s not posturing.

People say it takes nearly a decade to grow a single veteran RF engineer. Two designers handed identical specs can produce chips with wildly different power draw and performance. Worse, that know-how rarely makes it into documentation. Ask why a particular transistor sits where it does, and the honest answer is often “because that’s what worked when I tried it.”

The result is a brutal talent bottleneck. Skilled analog and RF designers are chronically scarce worldwide. Demand for wireless silicon is exploding — past 5G into 6G, satellite communications, and self-driving radar — but the pool of people who can design it is fixed. The pressure to automate is enormous.

How AI attacks the problem differently

The key is that AI doesn’t think like a human.

A human engineer uses intuition to land on a design — “this should be close enough” — then validates it in simulation. Each simulation run is slow, so the number of attempts is limited. AI, especially reinforcement learning approaches, throws tens or hundreds of thousands of candidate designs at the problem and learns from the results, including while everyone’s asleep.

A useful analogy: a human grandmaster plays the established joseki; AI is AlphaGo finding moves humans never considered. In practice, AI has produced circuit layouts that look wrong to veteran eyes — “you can actually do it like that?” — yet measure better on the bench.

Layer on the recent wave of large language models and it gets more interesting. New tools can parse a design spec in natural language, generate circuit-design code, interpret simulation results, and suggest the next round of edits. The tacit knowledge that used to live only inside one person’s head is starting to get absorbed as data.

It’s too early to say “AI does it all”

Time to cool the hype, though.

Today’s AI shines at relatively well-defined problems: optimizing component values inside a fixed circuit topology, or refining a layout. The genuinely creative work — inventing an entirely new RF circuit structure from a blank page — is still human territory. And anything AI produces still needs an expert’s sign-off before it goes into mass production.

There’s also a data problem. High-quality training data is rare in RF, because no company is going to hand its hard-won design secrets to outsiders. So for now, the realistic picture isn’t AI replacing the veteran. It’s one veteran using AI as a tool to do the work of ten. That doesn’t erase the talent bottleneck, but it eases it considerably.

The last front of automation just opened

Back when digital design was being automated, plenty of engineers swore a machine could never do their job. The fact that analog and RF — the final holdout — is now cracking means more than just making chips a little faster. It’s a signal that a domain long treated as pure human intuition and tacit knowledge is starting to yield to data and learning.

Is there a part of your own work you assume can’t be automated because it’s all experience and feel? Watching a 30-year black art finally get cracked open, it might be time to question that assumption.