# The Constraint AI Can't Compress

*The semiconductor bottleneck has spent three years climbing a ladder. It has reached the rung money can't buy and machines can't build — and AI, the tool everyone expects to dissolve it, is making the problem steeper, not flatter.* --- > "Learning is local." — Morris Chang, founder of TSMC --- For three years we have been asking the same question, over and over, in slightly different words: what is the *real* bottleneck in semiconductors? And every answer we have reached for has been a noun you can buy. First it was **capital**. So the money came — tens of billions in subsidies, a global bidding war for fabs. Then it was **capacity**. So the shells went up across three continents. Then it was **advanced packaging** — CoWoS, the chokepoint behind every AI accelerator — so the lines multiplied. Then it was **memory**, so HBM booked out past 2027 before the wafers existed. Each time, we spent our way at the constraint. Each time, it did not dissolve. It *climbed*. That is the pattern worth naming, because it is the whole story. A bottleneck you can buy your way past is not really a bottleneck; it is a delay with a price tag. The constraint has been migrating, rung by rung, away from everything money moves quickly and toward the one thing money moves slowly — if at all. It has now reached the top of the ladder: **earned operational expertise.** And this is precisely the rung where AI — the universal solvent, the thing every pitch deck promises will dissolve every constraint — arrives and cannot pass. ## Two jobs inside a fab To see why, you have to separate two jobs that get collapsed into one word: "talent." The first job is **detection** — spotting the defect signature buried in a river of etch-tool sensor data, catching the drift before it becomes scrap. This work is *writable*. It is patternable. It can be reduced to features and thresholds and, increasingly, learned by a model. The proof is already on the floor: defect-detection models have been compressed small enough to run on an ordinary CPU right at the tool, no GPU required, watching the process in real time. On this half of the problem, AI is genuinely good — and getting better fast. The second job is **resolution** — knowing *why* the yield dropped, and what to change on the line to fix it. This work is not writable. It is earned, over years of standing in front of a specific tool when it misbehaves, accumulating a feel for its failure modes that never quite makes it into a document. A model cannot train on it, because it was never written down. A new graduate does not arrive with it, because it cannot be taught in a lecture hall. Half a century before the first neural network touched a wafer, the philosopher Michael Polanyi gave this distinction its name. *We can know more than we can tell,* he wrote. Some knowledge is **tacit** — real, valuable, decision-shaping, and fundamentally resistant to being written into a procedure. Detection is the knowledge you can tell. Resolution is the knowledge you cannot. So here is the sentence the entire AI-in-the-fab conversation keeps missing: **AI compresses the bottleneck you can write down. It is useless against the one you can only earn.** AI does not remove the talent shortage. It *relocates* it — from detection, where it was never really binding, onto resolution, where it always was. ## The scoreboard that counts the wrong thing The industry has a number it likes to quote. Deloitte projects the sector will need more than a million additional skilled workers worldwide by 2030; the SIA and Oxford Economics put the U.S. shortfall alone at roughly 67,000 by the end of the decade. Every AI-in-the-fab pitch points at that number and says, in effect: *this is the gap, and we will help close it.* That number is built to be closed. It counts **seats** — fill the seats, solve the problem. But seats are fungible, and the constraint is not. Hidden inside that million-worker gap is a much smaller subset — the resolution tier, the yield engineers whose judgment gates whether a fab full of the world's most expensive machines produces good die or scrap. That subset does not behave like the rest of the number. You cannot fill it with a model, and you cannot fill it with a fresh hire arriving in 2030, because the thing that makes those people valuable is the very thing that takes years to grow. The scoreboard measures the problem money can solve. The binding constraint is the one it can't. ## Learning is local Which brings us back to Morris Chang, standing at a podium at MIT, explaining why TSMC works where it works. The learning curve — the slow, compounding process by which a manufacturer drives yield up and cost down through sheer accumulated experience — only functions, Chang argued, when production is concentrated in one place, with workers side by side in common conditions, able to transfer what they learn to one another directly. Pull that experience apart across geographies and the curve flattens. Keep it dense and co-located and it compounds. *Learning is local.* Three words, from the man who built the modern foundry industry, and they are the quiet foundation under everything above. Earned expertise is not a fluid that flows to wherever the capital pools. It is a *stock*, accumulated in specific places, inside specific people, over specific decades. It concentrates. And once you see that, you can see the shape of the real map. ## The part everyone has backwards Here is where the consensus inverts. AI was supposed to **democratize** expertise. Level the field. Hand the thin-bench newcomer a copilot that closes the gap with the incumbents who spent thirty years learning the hard way. That is the promise in nearly every deck. In the fab, it does the opposite. Watch what AI-assist actually does at each end of the talent gradient. Where the bench is **deep**, AI is a force multiplier: it flags the anomaly, and the veteran two desks down resolves it in an afternoon — detection accelerated, resolution already in the building. Where the bench is **thin**, AI flags the same anomaly and then… the escalation path runs out. The tool surfaced a problem no one on site has the earned judgment to fix. It has raised the ceiling on detection and left the ceiling on resolution exactly where it was. So AI-assist delivers its *most* value where expertise is already densest, and hits its ceiling *fastest* where expertise is thinnest. It multiplies deep benches and merely patches thin ones. **AI does not flatten the talent gradient. It steepens it.** The tool sold as the great equalizer is, in the one place that matters most for the buildout era, a concentrator of advantage — a mechanism that widens the very gap it was supposed to close. ## The map nobody is drawing Which means the map of semiconductor power in the AI era is not the one we keep staring at. It is not drawn in dollars of capex, or square feet of cleanroom, or lines on an export-control chart. Those are the visible race, and everyone is scoring it. The map that will actually decide who converts capacity into yield is drawn in **talent density** — the accumulated, co-located, un-transferable judgment of a few thousand yield engineers. And AI is redrawing that map steeper, not flatter, in real time. The honest caveat matters here, because the thesis is not fatalism. Expertise is not *immovable*. You can poach it, relocate it, and — slowly — train it. Pipelines can be built. A dense bench can seed a thin one if you move enough of the right people and give them a decade. AI-accelerated training tools may genuinely bend the curve over time. But every one of those levers operates at the speed of a human career — *years* — while capital moves overnight and a fab shell goes up in a couple of build seasons. The gradient is sticky, not permanent. And the buildout boom is pouring the most new capacity, in the shortest time, into the thinnest benches. The mismatch is not a rounding error. It is the defining operational risk of the entire supercycle. ## The question worth sitting with So stop scoring the buildout in capex. Score it as a **ratio**: expertise growth over capacity growth. How fast are you growing — or importing, or concentrating — the people who can resolve what the new capacity produces, relative to how fast the capacity itself is coming online? Add capacity faster than you grow that resolution tier, and you have not built a fab. You have built a yield problem that AI can detect in exquisite detail and cannot debug. We spent three years watching the constraint climb a ladder — capital, capacity, packaging, memory — and at every rung we assumed the next check, the next machine, the next model would dissolve it. It never did. It only moved up. It has finally reached the top. Money couldn't buy this rung. AI can't compress it. You don't remove a bottleneck. You relocate it. This one has nowhere left to climb. --- *Steve Williams — babnews.org*