What happens when the world’s most expensive input becomes its cheapest
On September 4, 1882, Thomas Edison flipped a switch at the Pearl Street Station in lower Manhattan and began selling electricity to eighty-five customers. Before that moment, if you wanted electric power, you built your own generator. After it, you could buy power by the hour from a wire in the wall.
The technology was revolutionary. The economic transformation was not — at least, not yet.
For nearly thirty years, factory owners connected electric motors to the same systems they had used with steam. They kept the same layouts, the same shafts and belts, the same centralized power transmission. Productivity barely moved. It was not until a new generation of managers realized they could redesign the entire factory around distributed electric power — small motors at individual machines, rearranged floor plans, rethought workflows — that the productivity miracle arrived. The general-purpose technology was available for decades before the complementary reorganization unlocked its value.
We are standing at the same threshold, this time with intelligence. Reasoning, writing, planning, coding, analysis, persuasion — all of it increasingly available on demand, at a predictable price, from a wire in the wall. The question is not whether this is happening. It is. The question is what happens next, and history suggests the answer is: less than you expect in the short run, more than you can imagine in the long run, and almost none of it where you were looking.
I. The Basic Logic: When One Thing Gets Cheap, Other Things Get Expensive
For most of economic history, skilled thinking was the bottleneck. Could you find enough smart, trained people to analyze the problem, draft the plan, write the code, close the deal? The answer was almost always: not quite enough, not fast enough, not cheap enough. Organizations were built around this scarcity. Hierarchies existed to ration access to judgment. Salaries reflected the cost of renting a mind.
What happens when that cost collapses?
Value does not disappear. It migrates — away from the newly cheap input and toward whatever it cannot replace. This is the oldest pattern in economics, and it has held for every major technology before this one. Cheap steel made architecture valuable. Cheap bandwidth made content valuable. Cheap compute made data valuable. Cheap intelligence will make something else valuable. The interesting question was never about intelligence. It was always about everything else.
II. What Stays Expensive
When thinking itself becomes cheap, the constraints shift to everything thinking alone cannot provide. It is worth walking through these carefully, because this map is also a guide to where money, power, and competitive advantage are headed.
Good inputs. A brilliant reasoning engine operating on bad data produces confident, well-formatted garbage at unprecedented scale. Clean data, clear context, reliable ground truth — these were always important, but they were overshadowed by the cost of the reasoning itself. When reasoning is cheap, data quality becomes the binding constraint. Garbage in, garbage out does not soften when the processor in the middle gets smarter. If anything, it sharpens.
Accountability. Someone still has to carry the risk. The surgeon who lets an AI plan the procedure still holds the scalpel. The executive who follows the AI’s recommendation still faces the board. Decision rights — the authority to commit, to accept liability, to bear consequences — cannot be delegated to an API call. That is the whole point of accountability: a specific person is on the hook. As thinking commoditizes, the premium on being that person increases.
Coordination. When many AI agents can execute tasks in parallel, systems fail not within individual steps but at the handoffs between them. Getting one agent to draft a document is easy. Getting five agents to collaborate across a workflow without contradicting each other is an unsolved engineering problem. The bottleneck was never the individual worker. It was always the seams.
Attention and trust. When content is infinite, the ability to notice, evaluate, and care about anything specific becomes the scarcest resource in the system. Provenance — knowing that what you are reading was produced by a process you have reason to trust — becomes a load-bearing economic asset. A world drowning in content will pay dearly for reliable curation.
Distribution. When ideas are abundant, the ability to reach the right audience with sufficient credibility separates the ideas that matter from those that evaporate on contact with the world. Channels, brand, relationships, earned audience — these were always valuable. They become decisive when the supply side of insight approaches infinity.
And beneath all of this sits the physical layer. Intelligence runs on electricity. “Tokens per dollar per watt” is becoming a strategic metric as surely as “barrels per day” once was.
The pattern is simple, even if the consequences are not: automate the thinking, and you discover that thinking was never the whole job.
III. How Businesses Change (Slowly)
The first visible change is that thinking turns into a service you buy. Companies purchase reasoning the way they already purchase cloud computing: usage-based, standardized, multi-vendor. You do not hire a server rack. Increasingly, you will not hire a reasoning engine.
This creates something like cognitive supply chains — sequences of specialized AI systems where one gathers information, another drafts, a third checks, a fourth negotiates — all coordinated by software that replaces what middle management used to handle through emails and meetings.
But here is where the Pearl Street parallel bites. The firms that capture the most value will not be the ones that bolt a chatbot onto their existing processes. They will be the ones that redesign the process from scratch — the equivalent of rearranging the factory floor around distributed motors. And most firms will not do this quickly, because reorganization is expensive, risky, and politically messy in ways that purchasing a new subscription is not.
This means the competitive landscape will be deeply uneven for years. Some firms — typically smaller, newer, less burdened by legacy — will achieve dramatic productivity advantages in specific workflows. Most incumbents will see modest gains from shallow integration and will blame the gap on “culture” or “talent” when the real gap is in reorganization.
The old moat — “we hire the smartest people” — weakens when everyone can rent equivalent intelligence by the token. The new moats are distribution, proprietary data that improves with use, compliance infrastructure, and operational discipline. In a world of cheap thinking, the advantage belongs to whoever figured out everything around the thinking.
IV. What Happens to Work
The effect on labor is best understood not as jobs disappearing but as jobs being disassembled. Every job is a bundle of tasks, and cheap thinking pulls those bundles apart. The tasks with clear inputs and evaluable outputs go first. What remains is everything the machine cannot do or the society will not let it.
These residual human roles cluster around four things. Accountability: someone must sign off and face consequences. Ambiguity: when the question is not “how do we optimize X” but “what should we optimize for,” human judgment — messy, value-laden, politically negotiated — remains essential. Relationships: sales, leadership, care, negotiation, every context where trust depends on a human being present. And taste: brand sensibility, product instinct, the ability to say “this feels right” in a way no loss function yet captures.
Two outcomes coexist here, and this is what makes prediction hard. The optimistic one is augmentation: many workers, especially junior ones, become dramatically more productive. The gap between a novice and a veteran analyst narrows when both have the same reasoning engine. This raises the floor. The pessimistic one is polarization: the returns to owning platforms, audiences, and decision rights increase while the returns to raw cognitive skill decline, concentrating income around those who control the complementary assets. Everyone has the same AI brain, but not everyone owns the distribution.
Both forces operate simultaneously. Which one dominates is not a technology question. It is a political one. The technology decides what is possible. The institutions decide who benefits.
V. What Gets Created
Nearly every discussion of cheap intelligence focuses on the supply side: what happens to the existing economy when production costs drop. This is natural but incomplete. Cheap intelligence also creates demand that does not currently exist.
Personalized tutoring at scale. Continuous health monitoring with real-time clinical reasoning. Legal help for people who currently cannot afford a lawyer. Scientific hypothesis generation. Creative tools that let one person produce what used to require a studio. These are not speculative — they are the straightforward consequence of removing the cost constraint from cognitive labor.
Counting only the jobs that disappear while ignoring the ones created by new demand is the recurring error of automation anxiety. It was wrong about ATMs and bank tellers — branches actually increased for decades after ATMs arrived, because cheaper branch operations expanded demand for banking services. It was wrong about spreadsheets and accountants. It will likely be wrong again, though the transition costs for displaced workers are real and should not be waved away with historical averages.
The deeper point is that cheap thinking does not just reshuffle existing tasks. It expands the frontier of what is economically feasible. We are not automating the old economy. We are fertilizing a new one. The question is who gets to harvest it.
VI. The Trust Problem
When persuasion and content generation become cheap, lying scales faster than checking. The cost of producing convincing deception drops below the cost of detecting it. This is not a forecast. It is the present.
The counter-move is verification infrastructure: cryptographic content provenance, watermarking, digital signatures, identity layers, and “authentic by default” systems for high-trust contexts like government, banking, and healthcare. These are not nice-to-haves. They are the load-bearing pillars of a functional information environment.
But verification is not a technology problem alone. It is an institutional design problem. The societies that navigate this well will be the ones that build trust architecture before the damage becomes entrenched. You do not install a smoke detector after the house burns down.
VII. Where This Analogy Breaks
I have been treating intelligence like electricity — a general-purpose input that gets commoditized, with value migrating to its complements. The framework is useful. But honesty requires asking where it fails.
Electricity does not hallucinate. Cloud compute does not have agency problems. Bandwidth does not persuade you of things that are not true. Intelligence is different from prior commoditized inputs in ways that matter.
First, it produces outputs that are hard to verify without applying more intelligence. You can test whether the lights turn on. You cannot always test whether the reasoning is sound without reasoning yourself — which partly defeats the purpose of making reasoning cheap in the first place.
Second, it can be used not just to produce things but to shape what people want. Cheap electricity powers factories. Cheap intelligence powers persuasion. Every prior commodity changed what we could make. This one changes what we choose to buy.
Third, it may not commoditize cleanly. If the best models retain significant quality advantages over cheaper ones, and if those advantages compound in multi-step workflows where small differences at each step multiply, the market may look less like the electricity grid and more like the semiconductor industry — a few dominant players with wide moats. In that case, much of this essay still applies, but the gains concentrate far more than the word “commodity” suggests.
These are not fatal objections. They are boundary conditions. The electricity analogy is a powerful first approximation. But electricity never had opinions, and it never made you change your mind. Hold the analogy loosely, and watch for the moments when the commodity starts acting like a colleague.
VIII. The Honest Reckoning
I began at the Pearl Street Station. Edison sold electricity to eighty-five customers in lower Manhattan, and then the world waited thirty years for factory owners to figure out what cheap power actually meant. The technology was available. The reorganization was not. The productivity miracle arrived a generation late because the real cost was never the electricity. It was the redesign.
We are in the same gap now. The intelligence is available. The complementary reorganization — of firms, of labor markets, of institutions, of our own habits of thought — is not. If history is any guide, the gap will be longer, messier, and more politically contested than the technologists expect.
But I want to end with genuine uncertainty rather than false resolution. Electricity was inert. You plugged it in and it powered whatever you pointed it at. Intelligence is not inert. It touches cognition itself — the faculty we use to evaluate everything else, including whether the transformation is going well. We are commoditizing the instrument of judgment at the same moment we need judgment most. There is no precedent for this, and anyone who claims to know how it ends is selling something.
The factory owners of the 1890s, staring at electric motors bolted onto steam-era shafts, could not see the factory of the 1920s. We are those factory owners. The question is not whether the redesign is coming. It is whether we are the generation that does it — or the one that waits thirty years because the old layout felt safer.
The electricity is flowing. The redesign is the work.