The Geniuses Were Four Percent

Gautam Mukunda wrote a sharp column for Bloomberg Opinion this week (it ran in the Taipei Times as "The AI folks do not seem to understand intelligence"). His target is the assumption hiding inside both the utopian and the apocalyptic AI forecasts: that intelligence is all you need. Get smart enough and you cure cancer, or you end the world. He doesn't buy it, and he lands on the right instruction at the end. Stop arguing about the curve and watch the bottlenecks.

He doesn't say which bottlenecks. That's the part worth finishing.

The ratio nobody quotes

The best number in the piece comes from Alex Wellerstein, the historian of nuclear weapons. Los Alamos, the place with Bohr, Feynman and Oppenheimer, accounted for about 4 percent of the Manhattan Project's roughly $2 billion cost. About 80 percent went to the production plants at Oak Ridge in Tennessee and Hanford in Washington, the sites that made the enriched uranium and the plutonium.

Sit with that ratio for a second. The most concentrated collection of scientific talent in history was the cheap part. The expensive part was industrial: separation plants, reactors, power, chemical processing, construction crews, supply chains. Oppenheimer told Congress as much in 1945. Without scientists there's no bomb, but "if there had been only scientists, there also would be no atomic bomb."

That's the frame I'd put on "a country of geniuses in a datacenter." Grant the geniuses. The question the ratio asks is who builds the Oak Ridge.

Intelligence compounds where the feedback is fast

Mukunda's sharpest mechanism is about feedback. AI has moved fastest in coding, math and weather forecasting because those domains have deep data, quick feedback and clear right answers. The model gets corrected constantly, so it stays tethered to reality. Where feedback is slow, like biology experiments that take years to respond, more intelligence improves your hit rate but can't shorten the wait.

That's right, and it has a corollary he doesn't draw out. Intelligence gets deployed first where feedback is fastest. What's left over are the problems where it's worth less. So the more intelligence you add, the more the remaining work is dominated by the slow-feedback parts.

That's a precise description of a bottleneck relocating.

I watched this curve bend from the inside

Mukunda cites one number from my old world. Nicholas Bloom and his coauthors found that doubling chip density now takes more than 18 times as many researchers as it did in the early 1970s.

That isn't an abstraction to anyone who spent a career in semiconductor equipment. The industry didn't keep Moore's Law alive by getting 18 times smarter per engineer. It kept it alive by throwing capital, tool complexity and process integration at a problem that had stopped yielding to cleverness alone. An EUV scanner is what a problem looks like after ideas can no longer solve it cheaply. Lead times on those tools run 12 to 24 months. A High-NA system costs in the neighborhood of $400 million.

When intelligence stopped being enough, the constraint moved into things you have to build, ship, install and qualify. That's the semiconductor version of the 4 percent and the 80 percent.

The constraint ladder is a feedback-speed ladder

Here's the connection the column is one step away from. The rungs of the AI buildout I've been writing about, from chips to packaging to memory to power to water to consent, aren't only ranked by how little money can move them. They're also ranked by how slowly they give feedback.

• Software: feedback in seconds. Anthropic says its model already writes most of the code its engineers merge.
• Chip design: feedback in weeks to months, simulation-heavy. AI helps a lot here.
• Fabs and packaging: feedback in quarters. A new process learns yield one lot at a time, and the tools arrive on a 12- to 24-month lead.
• Power: feedback in years. Transformer lead times run up to five years, and interconnect queues run longer.
• Consent: feedback on the political calendar. Zoning hearings, elections, moratoriums. No compute setting shortens a county commission meeting.

Each step down, feedback gets slower and intelligence gets less leverage. That's why the rung sequence isn't arbitrary. It's the order in which cheap intelligence runs out of things to compress.

What cheap intelligence actually does to the buildout

The instinct is to assume abundant intelligence eases the physical constraints. Smarter grid planning, better site selection, faster permit paperwork, higher fab yields. Some of that is real, and it deserves its due. AI in the fab speeds up defect detection. AI in design compresses cycles.

But watch what happens next. Make the fast-feedback rungs faster and demand for the slow ones goes up, not down. Every coding speedup, every model that ships sooner, pulls forward the need for more compute, which pulls forward the need for more power and more sites. Anthropic's own researchers made the Amdahl's Law point: speed up one part of a process and you move the bottleneck rather than delete it. The more intelligence succeeds, the harder it leans on the rungs it can't reach.

That's what both the utopian curve and the apocalyptic curve miss. A self-improving model in a datacenter still can't pour concrete, wind a transformer or win a zoning vote. Mukunda puts this year's combined capex for Alphabet, Amazon, Meta and Microsoft at roughly $760 billion. That money isn't going to the geniuses. It's going to the Oak Ridge.

Where the 80 percent sits now

If you want to track the AI buildout the way Wellerstein tracked the Manhattan Project, follow the 80 percent. The tell won't be a benchmark. It'll be the first quarter a major player's growth is capped by "couldn't power it" or "couldn't site it" instead of "couldn't train it."

The Manhattan Project had one advantage the AI buildout doesn't. A wartime government could command Oak Ridge and Hanford into existence on its own timeline. Today's version runs through interconnect queues, transformer backlogs, water boards and county commissions, each moving at its own speed and none of them answering to intelligence.

You don't remove a bottleneck, you relocate it. Cheap intelligence may be the most efficient relocation machine we've ever built, and it's pushing the constraint down the ladder to the rungs that move slowest.

So the next time someone tells you what the geniuses in the datacenter will do, try a different question. Who's building the other 96 percent, and how long will it take them?