Does Intelligence Need The Railroad?

Direct Source Verification: This story is aggregated from Forbes (forbes.com). Full reporting rights and copyright belong to the primary publisher.
We have accepted it as settled that AI needs all of this infrastructure. We should be asking whether that is a fact about intelligence or a fact about one design.

Charles Yeomans is the chairman and founder of Atombeam.

getty​In Fortune’s recent Term Sheet (paywall), Allie Garfinkle asked investor Hans Tung to take the Industrial Revolution analogy for AI seriously, which he did. His argument runs from Britain’s steel steamships to the railroads that helped decide the American Civil War, and he casts data centers as the railroads of this century. The models that make a difference, he argues, are the ones with the infrastructure behind them: “The models that win are the ones that are available.”

As history, I find the argument persuasive. As a prediction, it rests on an assumption almost nobody examines: that intelligence has mass. We have accepted it as settled that AI needs all of this infrastructure. We should be asking whether that is a fact about intelligence or a fact about one design.

Railroads were important because cargo was heavy. Steel ships became critical because trade goods had to cross oceans in hulls. Data centers matter today because the dominant AI architecture makes every act of machine intelligence heavy. A large language model answers a question by driving it through billions of calculations on specialized processors, and it does so again for the next question, and the next, forever. Usage rises, and the compute bill rises with it. Under that architecture, Tung is simply right. Intelligence must ship to the railhead, and the nation with more track wins. One estimate he cites puts the American buildout at an average of 3.6% of GDP every year into the next decade, which tells you how heavy we have decided intelligence has to be.

But heaviness is a property of the architecture of AI. The way our brains work suggests there is a different way to build machine intelligence: Hold knowledge as durable structure, and compute each answer directly from it. The engineered version can go one step further than biology: The structure can be explicit and inspectable, so a machine can show the basis for what it concludes. In a system built this way, learning is the expensive step, and it happens once. After a subject is learned, producing an answer would be computationally trivial, small enough to run on an ordinary processor, with no round trip to anyone’s railhead.

We do not need much compute to identify a car when we see it, or a person we know. Our brains run on 20 watts, not megawatts, so we know this kind of efficiency is possible. The economics would be the reverse of the token meter: Cost falls as the machine learns, because knowledge would persist instead of being recomputed. Something else follows from structure: A machine that computes from what it actually knows can decline to answer when it does not, instead of generating a plausible guess. The reliability problems keeping AI out of high-consequence work are, at root, the same architecture question.

If this idea for a new kind of AI proves possible, and prototype work by my own team and others gives me reason to believe it is, the strategic picture inverts in two ways. The first concerns Tung’s own Civil War detail. He points out that most battles clustered near the railroad because the railroad was the dependency, and dependencies are what adversaries target. A force, a hospital or a factory whose machine judgment lives locally needs no railhead; everything stays behind the firewall, and the whole thing runs on the power budget of a laptop, or even less, rather than a substation. Local AI is not just more secure and more available; it steps almost entirely out of the energy problem, because computing an answer from learned structure draws watts where generating one from scratch draws megawatts.

The second is the geopolitical race itself. We need to face the fact that we are betting on out-pouring concrete against China, perhaps the most formidable builder of physical infrastructure on earth. The asymmetric answer to that is not more gigawatts. It is more intelligence per watt, which is an architecture question before it is an energy question.

The railroads will be built, and much of that buildout will be valuable; training, research and the broad models that benefit from scale will use every mile of track. But the assumption that all intelligence must ride those rails deserves the scrutiny we give any assumption worth trillions. Tung says the winning models are the available ones. I agree. The most available intelligence is the kind that never has to travel.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Original Source
https://www.forbes.com/councils/forbestechcouncil/2026/11/18/does-intelligence-need-the-railroad/
Visit Forbes ↗
SHARE STORY:
𝕏 f in

Related Coverage in Technology