Rails Moved Goods, the Internet Moved Information, AI Will Move Judgment
Some of history’s most important technological revolutions were also among its most dangerous investments if entered at the wrong time.
Railways changed civilization. The internet did too. Yet both went through periods when capital ran ahead of economically usable demand. Two claims can be true at once: a technology can be real, necessary, and transformative, while the infrastructure built for it can still be overfunded.
That distinction matters to how I think about AI.
Infrastructure investment has a unique characteristic. It is a bet with a time lag. Capital is committed first, and it takes years before the actual capacity begins operating. A data center powered on today is the result of a decision made long ago. When many companies see the same shortage at the same time and make the same judgment, they are not simply meeting today’s demand. They are simultaneously creating supply for several years from now.
That is why scarcity can become excess much faster than people expect.
Telecommunications infrastructure during the internet era is an uncomfortably good precedent. Around 2000, the industry was genuinely worried about a severe shortage of fiber capacity. Demand for bandwidth was not imaginary either. Internet traffic continued to grow after the bubble burst.
But something else happened at the same time. Technologies such as wavelength-division multiplexing allowed far more information to travel through a single fiber already in the ground. The physical infrastructure stayed in place while its effective capacity kept rising.
Demand grew enormously.
Supply capacity grew faster.
AI may follow a similar pattern.
The world is now committing immense amounts of capital to data centers, accelerators, memory, networking equipment, cooling systems, power generation, and electrical grids. A data center can take years to become operational, and the surrounding grid infrastructure can take much longer.
On the other side, the unit cost of producing intelligence is falling at extraordinary speed.
This is not only a matter of better hardware. Quantization, routing, caching, memory management, model architecture, and serving software are all improving. The same installed equipment can produce far more intelligence than before.
My own view is that the risk of meaningful excess supply could rise around 2027, when much of the large-scale infrastructure now under construction begins operating. I would not be surprised by a substantial correction at some point before 2030.
But I do not believe I know the exact timing.
That may be the more important realization.
Trading this cycle accurately would require near-real-time tracking of hyperscaler capital expenditure, GPU deliveries, HBM contracts, optical networking, fab utilization, data-center grid connections, cloud utilization, model efficiency, financing conditions, order cancellations, and customers’ actual usage.
For an individual investor who has another job and receives much of this information late through the press, entering this timing game feels to me like going to war with a horse and a spear against an opponent armed with guns and tanks.
The battlefield has to change.
Rather than guessing who will sell the most equipment next quarter, I find it more interesting to ask how the world changes if all this investment succeeds.
What happens when intelligence itself becomes cheap?
As AI agents move deep into companies, financial systems, legal work, medicine, research, and public administration, the cost of many forms of execution that are expensive today will fall. People are basically not good at law or accounting. AI is much better. There are already signs of people using AI for contracts instead of lawyers and using AI instead of accountants.
Finance is already offering an early view of this transition.
Stablecoins, tokenized assets, programmable payments, and blockchain settlement are moving into both experiments and production systems at financial institutions. If machines are to become economic actors, they need machine-readable ownership, programmable rules, and a way to transfer value without waiting for a human to approve every transaction.
At first, the result may look little different from the present. A dollar will still be a dollar, merely represented as a stablecoin. Banks will continue to exist, but may operate on programmable ledgers.
Once the entire financial infrastructure becomes digitally native, however, a more uncomfortable question can arise.
Must the money that exists on top of it remain dependent on the state?
I can imagine a future in which Bitcoin increasingly plays the role of decentralized gold, while Ethereum—or a system based on similar principles—becomes part of a decentralized economic and monetary substrate.
I do not think that outcome is inevitable.
There is a large leap between blockchain becoming financial infrastructure and cryptoassets themselves becoming dominant money. States also have strong incentives to defend monetary sovereignty.
But the permanence of the present system is not, by itself, a reason to dismiss the possibility.
American political traditions in particular have repeatedly expressed suspicion of concentrated power and a desire to preserve a sphere of individual choice. Public blockchains fit those tendencies surprisingly well: they allow certain forms of social coordination without giving a single institution ownership of the entire ledger.
The state may lose control in some domains while gaining far stronger powers of observation and intervention in others.
The future, then, is unlikely to be simply decentralized or centralized. Both movements may proceed at once.
There is another change I find even more interesting.
Even today, almost no one understands exactly how a modern computer works from transistor to operating system to application. Very few people understand, end to end, the financial system through which their salary, mortgage, pension, brokerage account, deposits, credit cards, and international transfers pass.
We use these systems every day anyway.
In the future, this condition may become far more extreme.
We may live in a world whose systems are extraordinarily efficient and deeply interconnected, yet which no human being fully understands. People will interact with those systems through interfaces and AI, using them through rough intuition rather than a grasp of the whole structure. We may rely less on individual understanding than on the belief that machines, protocols, audit systems, institutions, and other machines are checking one another.
In some domains, we may come to trust an algorithmic system more than an explanation from a person.
That no longer feels like science fiction to me.
Human nature has changed much less than the institutions surrounding it.
The desires for status, safety, belonging, freedom, power, family, recognition, and meaning are old. What technology repeatedly changes is the social structure through which those desires are negotiated. Political systems, family structures, corporations, markets, religions, states, and social conventions keep changing clothes.
The human being inside them changes much more slowly.
AI and blockchain may be the next clothes we put on that human being.
For that reason, I do not want to call this future progress automatically.
A society in which machines handle more judgment and execution will be extraordinarily convenient. It may be safer, wealthier, more productive, and capable of far more than the society we know.
At the same time, people may directly control less.
Convenience has a peculiar irreversibility. Once society reorganizes around a much more efficient system, returning to the old way becomes prohibitively expensive. We can acquire capabilities no one wants to surrender while gradually giving up forms of direct control that are difficult to recover.
Whether that counts as progress depends on what one values.
If execution becomes abundant, the scarcest resource may ultimately be judgment: the ability to decide what should be optimized in the first place.
AI may become far better than humans at finding the best means to reach an objective we have chosen. It may also warn us when our choices contradict one another, are destructive, or are likely to destabilize society. A future in which an intelligent assistant quietly steers humanity away from collectively destructive decisions, as in science fiction, no longer sounds particularly strange.
I would welcome much of that future.
But I hope one boundary remains with human beings.
We can hand AI the choice of better means.
But I do not want to hand over the choice of what we should pursue.
Not everyone values autonomy equally. Many people willingly exchange some control for safety, convenience, belonging, and predictability.
But I do not want a life in which I cannot meaningfully choose the direction of my time and my life.
If I must spend most of my time following someone else’s instructions—whether they come from a company, a state, or one day an algorithm—and cannot choose my own direction, then however comfortable that life may be, it feels to me like a form of bondage.
The more powerful machines become, the more important this distinction will be.
The last human bottleneck may not be intelligence.
It may be the insight and judgment to decide what that intelligence should be used for.