ENVIZNWhen Thinking Gets Cheap
The cost of a unit of machine cognition fell more than 280-fold in two years, while the transformers, substations, and concrete behind it barely moved. Much of what happens next depends on that gap.
A model can draft a distribution design in seconds. The switchgear in that design still has a lead time measured in quarters, and the price has moved the wrong way.
Every economic transition so far has worked by making physical effort cheaper. The Industrial Revolution cut the labor in a manufactured good. Computers cut the labor in a calculation. Automation let factories raise output without raising headcount in proportion.
Artificial intelligence is doing something structurally similar to cognitive work, and the price signal is already unmistakable. Stanford’s 2025 AI Index tracked the cost of running a model at roughly GPT-3.5’s level on a standard benchmark. In November 2022 it cost about $20.00 per million tokens. By October 2024 it cost about $0.07. That is a fall of more than 280 times. It describes one capability level rather than every AI task, which makes it useful as a measure of price change at fixed output.
Discussion of this number usually turns to whether AI takes people’s jobs, a framing I think is too narrow for what the number implies. If the cost of a unit of cognition is collapsing while the cost of a megawatt, a transformer, and a poured foundation is not, then the interesting consequences reach past the labor market, into the widening distance between what is now easy to think up and what remains hard to build.
This essay is a set of open questions rather than a forecast, and it relies on three distinctions: exposure to AI is different from displacement, wage compression is different from wage decline, and an economy that produces abundance can still leave people unable to afford it.
Expertise is paid for being scarce
An underwriter prices risk. An engineer applies judgment to a constrained problem. A designer turns a brief into something a person will actually look at. Their pay reflects demand, responsibility, and the fact that relatively few people can do the work. Change the last term and the compensation changes with it, even if the work itself is no harder than it was.
There is already direct evidence that AI narrows the gap between workers. A study published in The Quarterly Journal of Economics in 2025 followed 5,179 customer support agents through the rollout of a generative AI assistant and found average productivity up roughly 14 percent. The gains were concentrated among the least experienced and lowest-skilled agents, while the most experienced saw little change. The assistant was effectively distributing the tacit knowledge that senior agents had spent years accumulating.
That is one workplace and one deployment, and it would be a mistake to generalize it into a law. But it describes a mechanism worth naming: when a capability that used to take years to acquire becomes available on demand, the premium attached to having acquired it compresses, though it does not fall to zero and it does not fall everywhere at once.
This is also where the argument most often goes wrong, so it is worth being careful. A 2025 IMF working paper, AI Adoption and Inequality, modeled exactly this and found something less tidy than the intuition suggests. High-income workers are more exposed to AI than low-income workers, which points toward narrowing wage gaps. But their tasks also turn out to be highly complementary with AI, which points the other way, and they are better positioned to capture returns on capital. The paper’s conclusion is that AI could compress wages while widening wealth, because wage effects follow exposure to AI while wealth effects follow returns on capital.
Those results come from a calibrated model, not from observed outcomes. What I take from them is narrower than a forecast: more accessible intelligence changes which capabilities still earn a premium, ownership conspicuously among them, and that change need not produce a more equal society.
The frontier keeps moving, and it does not wait for you
If capability becomes cheap to reproduce, the returns migrate toward whatever is not yet reproducible. Today that means AI systems, semiconductors, advanced packaging, and the power and construction capacity underneath them. Software has already run this cycle once: work that required a funded team and a data center became accessible to one person with a laptop and a cloud account, and the premium moved upstream to whoever was doing the genuinely new thing.
The complication is that the frontier does not hold still long enough to settle on. Once models get good enough at the work of building models, the scarce skill of this decade becomes an ordinary skill in the next one. Robotics looks like the next stop, and it has the same property: a sufficiently capable system eventually helps design and build its successor.
Advancing a technology and capturing its value are separate activities, and they pay differently. An engineer can contribute to a genuine breakthrough and be paid a salary for it. The firm that owns the resulting process, the customer relationships, and the installed capacity collects on that breakthrough for as long as it holds, and keeps collecting after the human effort required to maintain it has fallen.
None of this predicts that technical work stops being well paid. It suggests that advantages attached to human capability are becoming less durable, while advantages attached to owning what that capability produced are holding up.
Creative work splits into production and authorship
Creative labor is the clearest place to watch this happen, because the output is visible and the cost curve is steep. The ILO’s May 2025 refined index of occupational exposure found that roughly one in four workers globally are in an occupation with some generative AI exposure, with clerical work most exposed and media and web occupations rising as image, video, and voice generation improved.
The ILO is explicit that exposure does not mean displacement. Most occupations are bundles of tasks, only some of which are automatable, so transformation is the more likely outcome than replacement. Only about 3.3 percent of global employment sits in the highest exposure category.
Still, something real is separating, and one painting makes the split unusually visible. Jan Matejko finished Stańczyk in 1862, at twenty-four. A court jester sits alone in a dark room while a royal ball carries on in the next hall. On the table lies a dispatch reporting that Smolensk has fallen to Moscow. The court is celebrating a different victory and has either not read the letter or not understood it. His marotte, the fool’s sceptre, lies on the floor where he dropped it. Through the window: the darkened profile of Wawel Cathedral, a comet, and the three stars of Orion’s Belt.
A model could generate that scene now, and generate a fluent reading of it too, about the sad clown, the isolation of the performer, and public function set against private state. Both outputs would be competent, and both would miss what actually holds the painting together.
Matejko gave the jester his own face. He was a Pole painting in 1862, under partition, a year before the January Uprising. He put himself in the room as the one person who reads the dispatch correctly while everyone else dances. A generator cannot be prompted into that symbolism, because it sits outside the image, in the relationship between the image, the year, and the man, and a viewer who knows all three is looking at a different painting than one who does not.
The argument depends only on people caring whether a thing came from a life, whatever one concludes about machine experience, and the evidence that they do is strong. As production gets cheaper, provenance, authorship, and the story behind the work carry more of the value, which is why commercial creative labor can get cheaper at the same time that specific human artistry gets more valuable.
The same logic applies beyond art: once making something is routine, getting people to pay attention to it becomes the constraint. Attention, trust, and reputation start doing the work that technical scarcity used to do. That is not obviously good news: those are precisely the assets that existing wealth, established networks, and prior fame are best at accumulating.
Cheap cognition, expensive physics
The abundance argument runs into a physical limit here: falling inference costs make designing things faster but do very little to the cost of building them.
A model can produce a plausible electrical distribution architecture in minutes. The project still needs engineering validation, switchgear with a lead time measured in quarters, a construction crew, capital, an interconnection agreement, and a power supply that physically exists.
The infrastructure serving AI is the cleanest illustration available, because it is the one place where the two curves are visibly diverging. In Key Questions on Energy and AI, the IEA projects global data centre electricity consumption roughly doubling from 485 TWh in 2025 to around 950 TWh in 2030, reaching about 3 percent of global electricity demand. The same analysis names the binding constraints: electricity infrastructure, advanced semiconductor manufacturing, equipment supply chains, and access to capital.
None of the items on that list gets cheaper because tokens got cheaper, so the pace of physical buildout tends to be set by them rather than by model prices.
Robotics and automation will eventually erode some of these physical costs too. But the two cost curves fall on different schedules, and people have to live through the interval between them. What happens to someone whose profession reprices before the cost of living does? Falling design costs arrive years ahead of falling construction costs, and a price decline that has not yet reached the goods a person buys does nothing for their budget.
This is also why I would resist reading productivity gains as imminent job losses. A March 2026 NBER working paper surveying nearly 750 corporate executives, Artificial Intelligence, Productivity, and the Workforce, found wide variation in adoption, positive but uneven labor productivity gains concentrated in high-skill services and finance, and no evidence of a near-term aggregate employment decline. A companion survey of roughly 6,000 executives found nine in ten reporting no employment effect over the prior three years, while predicting modest effects ahead. The technology adds capacity, and who benefits from it depends on factors these surveys do not measure.
What money would still be for
Push the trend far enough and a genuinely strange question appears. If AI and robotics eventually take most of the labor out of food, transport, healthcare, and manufactured goods, then an ordinary income buys a life that currently requires real wealth. What is money for at that point?
Money would still matter, though its job would change. Today it is primarily a claim on survival, and that pressure shapes nearly every large decision a person makes: what to study, what job to take, whether to move, when to have children. Take most of that pressure out and the gap between an ordinary income and a somewhat larger one stops being the difference between security and precarity.
Scarcity would shift toward things that cannot be manufactured. A robot might build a house cheaply, but it cannot make more waterfront, another copy of a particular neighborhood, a specific school district, or a view. Construction costs could fall while location, access, and proximity stay exactly as contested as they are now, and more of what money buys would be position.
All of which depends on a condition that is doing enormous work in that paragraph: that lower production costs actually reach people as lower prices, which they might not. Cheaper to produce and cheaper to buy are different claims, connected by competition, distribution, and policy rather than by physics. If that link fails, money stays exactly as load-bearing as it is today no matter how automated production becomes.
That possibility deserves more attention than it gets, because it breaks the usual argument in both directions. Inequality and material prosperity are separate measurements. A society could see wealth concentrate further while the median person lives materially better than the wealthy did two generations earlier. Whether that trade is acceptable is a political question that economic measurement alone cannot settle.
The hard part is the transition
None of this is settled. AI may keep improving quickly, hit a wall, or develop somewhere nobody is currently looking. New occupations may appear faster than old ones compress. The IMF paper is explicit that AI could raise pay for workers whose skills complement it, which is the opposite of the compression story and rests on the same evidence.
The argument I find hardest to dismiss concerns the transition. Modern economies route production to people through wages: people are paid for their work and use that pay to buy what others produced. If cognitive labor keeps repricing, the benefits of higher productivity still have to reach people somehow, through lower prices, employment, ownership, public provision, or something not yet designed. Productive capacity alone guarantees nothing about who can afford the output.
Human labor may lose economic value before people stop depending on wages for their income.
The endpoint might be abundance, and the path there still runs through a period where the price of what people sell falls faster than the price of what they need. Beyond whether AI makes us richer in aggregate, the questions worth asking are what wealth means once intelligence is ordinary, how long the gap between cheap thinking and expensive building stays open, and which workers bear the cost while it stays open.
Inference got about 280 times cheaper in two years, and transformers, substations, and poured concrete did not follow it down.
All insights