ReadReal vs Hype04 Aug 20262:25Semiconductors: chips, fabs & yield

AI Just Started Designing Chips By Itself — Then They Locked It In a Cage

Three Companies Made AI Design Chips Alone. Nobody Covered the Catch.

A dark glass panel showing an abstract chip floorplan as colour blocks, one block outlined in warm gold light
The object this week · generated illustration, no people, no brands
Video

This one has not been published to the channel yet. Subscribe on YouTube and it will turn up there.

The number
50×

SYNOPSYS · DAC 2026

The 60-second version
  • In one week, three of the companies that build the software behind every chip on Earth — Synopsys, Cadence and Siemens — shipped AI that designs by itself, and almost nobody outside the industry noticed.
  • Synopsys claims a verification agent up to 50× faster to a validated design; Cadence shipped an agent stack; Siemens wired its agents into physics engines.
  • Chip design runs a zero-tolerance error regime, verification already eats ~70% of a project, and no autonomous agent has taped out a real chip without humans yet.
  • Meanwhile Gartner expects AI agents to run just 10% of production operations by 2030, up from 2% today.

Why this matters

But the part everyone skipped is the one that matters: Siemens' design agents can't proceed on a decision until a deterministic, physics-based tool confirms it. The AI proposes — a boring, non-AI checker disposes. That's not a limitation, it's the whole reason this is allowed to ship. Design sprints, the factory walks — and the reason is a rule anyone can use: speed follows reversibility.

What to do Monday

Microsoft made it a few clicks to publish your own agent to your whole company — so "what should an agent do without me?" stopped being theoretical. Paste your team's recurring tasks into Copilot, Claude or ChatGPT and sort them by two things only: what does being wrong cost, and can you undo it in 5 minutes. Three buckets — let it run / draft-and-I-approve / human only. When I ran it, "summarize a 60-minute recording" landed in let-it-run and "approve a $5k invoice" in human-only. Difficulty wasn't the axis. Reversibility was.

In the video
  1. 0:00Three companies, one week, almost no coverage
Over to you

What's the one task on your team you'd never hand an agent — even if it got 50× better tomorrow?

Argue with me on LinkedIn
Sources

References for this piece are in the pinned comment on the video. Nothing is cited here that cannot be linked.

Full transcript, 332 spoken words
Three companies that build the software behind every chip on Earth just let AI do the job by itself. Same week, same conference. Did you hear about it? Almost nobody did. Synopsys claims a verification agent up to FIFTY times faster to a validated design. Cadence shipped an agent stack. Siemens wired its agents into physics engines. So why isn't this the story of the year? Because the headline everyone wrote is wrong. Here's what got buried. Siemens' agents can't move forward until a deterministic, physics-based tool confirms the decision. The AI proposes. A boring, non-AI checker disposes. Why would you build the cage before you build the freedom? Because chip design is a zero-tolerance error regime. Verification already eats SEVENTY percent of a project. And no agent has taped out a real chip without humans. Not one. And if you work anywhere near a factory, you already know this pattern. It has a name: advanced process control. The controller proposes the next setpoint — a guard band decides if it's allowed. Twenty-five years old. So is this AI taking your job? No. It's AI finally being trusted with a slice of it — because someone built the fence first. Now — your FabSpeak Tip of the Week. Microsoft just made it a few clicks to publish your own agent to your entire company. So the question stopped being theoretical: what should an agent be allowed to do without you? Two-minute version. Paste your team's recurring tasks into Copilot, Claude, or ChatGPT, and sort them by two things only — what does being wrong cost, and can you undo it in five minutes. Three buckets: let it run, draft-and-I-approve, human only. I ran it. It put "summarize a sixty-minute recording" in let-it-run, and "approve a five-thousand-dollar invoice" in human-only. Difficulty wasn't the axis. Reversibility was. Save this, run it on your team this week. This is FabSpeak. Every week, one real thing pulled out of the AI noise, for the people who actually build. Come along.