Only 14% of chips work on first silicon today — the lowest rate in over two decades of industry tracking. And our collective answer is to hand more of the verification work to AI.
Electronics For You (EFY Group) interviewed me about this recently, and I called the trap the “sign-off paradox”: AI can write the tests, trace coverage, rewrite a codebase in minutes — but when a $100M chip fails in the field, no AI vendor signs the tape-out release. That accountability is still mine. Still human.
I stand by every word. But there’s a piece I didn’t get to say in the interview, and it’s the part that actually keeps me up at night.
The judgment to catch a killer bug isn’t downloaded — it’s grown. You build it over years in regression logs, chasing coverage holes, staring at waveforms at 2am on a bug that turns out to be your own testbench. That junior grind is exactly the work AI is now absorbing. We’re quietly automating the apprenticeship that produces the one thing AI can’t replace.
A Blackwell-class GPU carries hundreds of billions of transistors. The state space stopped fitting in one human head long ago — yet someone still has to own the call that it’s safe to ship. Where does that engineer come from if they never did the reps?
Verification was never the risk in the schedule. It’s what buys the risk down. AI makes a brilliant junior engineer. It doesn’t yet make a senior one — and I’m not sure who’s training the next batch.
If AI takes the entry-level verification work, where does the next generation of sign-off judgment come from? Genuinely curious how your teams are thinking about it.