This is a personal reflection, not a forecast dressed up as one. I am writing about a trend I think is real, but the second-order consequences are guesses, and I am sure some of them are wrong.
The single most important economic fact about AI is not that the models are getting smarter. It is that intelligence is getting cheap.
For all of human history, thinking has been expensive. Doctors, lawyers, engineers, accountants, researchers, designers, programmers, analysts - the entire knowledge economy was built on the premise that competent cognition is a scarce resource you have to pay for. Universities exist to credential it. Firms exist to ration it. Salaries exist to compensate it. Whole cities exist because the cognitive workers cluster in them. Strip away the abundance of competent thinking and the post-industrial world stops making sense.
That premise is being quietly demolished. According to Epoch AI, the cost of running a large language model at a fixed level of performance has been halving roughly every two months. That is not a Moore’s Law analogue. It is faster than Moore’s Law, sustained over years, and it shows no sign of stopping. A query that cost a dollar in 2023 costs cents now and will cost fractions of a cent by the end of the decade. The trend line, if you draw it forward honestly, ends at numbers indistinguishable from zero.
I want to be specific about what that means. Not in a “this changes everything” hand-waving sense, but in the actual mechanical sense of which institutions, prices, and assumptions stop working when the marginal cost of competent thinking drops to nearly nothing. Because that is the future we are walking into, and most of what we have built is priced for the regime we are leaving.
What “free” actually means
When I say intelligence is going free, I do not mean literally zero. There will always be a cost - electricity, hardware depreciation, capital recovery, the price of the rare-earth metals in the GPU supply chain. The point is not that the price is zero. The point is that the price stops being a binding constraint on use.
That is the right comparison to make. Bandwidth is not free. Storage is not free. Both still cost money. But the price of both fell so far, so fast, that we stopped budgeting for them. We stream video instead of waiting for it to download. We back up entire phones to the cloud without thinking. We log every event in every system because the storage to do so is cheaper than the cost of deciding what to drop. The price did not go to zero. It went below the threshold where rationing was worth the effort.
This is the regime intelligence is entering. Not literal freeness, but post-rationing. A future where you do not budget for tokens any more than you budget for HTTPS handshakes. Where the question stops being “can we afford to have the model think about this” and starts being “is there any reason at all to not have the model think about this”.
The Stanford AI Index tracks the supply side of this curve. Frontier compute capacity is growing at roughly 3.4x per year. Chip throughput is doubling on a similar cadence. Model efficiency is improving on top of that. Three exponentials stacked on each other do not produce a linear future. They produce a phase change.
What the old regime priced
Before we ask what breaks, it is worth being clear about what was being priced.
A junior consultant at a strategy firm bills four hundred dollars an hour because a competent twenty-six-year-old who can read a brief, structure an argument, and produce a coherent slide deck is genuinely scarce. A radiologist earns what they earn because reading a scan competently is a skill that costs years to acquire and the supply of trained eyes is bounded. A copywriter charges what they charge because writing serviceable marketing prose is an aptitude not everyone has. The price tag in each case is reflecting the difficulty of producing one more unit of competent cognition - the labour-economics term is the marginal product of that labour.
Now ask what happens when a machine can do any of those things at a fraction of the cost, around the clock, in any volume. Not perfectly. Not as well as the best human. But well enough that the next unit of “competent enough” cognition has effectively no cost at all.
The price did not stay where it was. It could not. The wage of the human worker was anchored to the cost of producing the next equivalent unit of work, and the next unit just got cheap.
This is not theoretical. It is the mechanism behind the productivity studies showing AI-assisted writers, coders, and analysts producing more in less time. The studies focus on output per hour. The economic story underneath is the collapse of marginal cost.
What breaks first
Some things break almost immediately, some take a decade, some hold longer than expected. Here is roughly the order I would bet on.
Pricing models priced per unit of cognition. Anything billed “per word”, “per page”, “per hour of analysis”, “per consult” is in immediate trouble. The price you can charge for a unit of cognitive work is bounded above by what a machine charges for the same unit, plus a thin premium for whatever the human adds. As the machine price falls, the bound falls. Translators have already lived through this. Copywriters are living through it now. Junior legal research, junior consulting, first-pass medical interpretation, basic accounting - all next.
Tutorials, documentation, and reference content as a business. The economic logic of writing a “how to do X” article was that someone would search for it, find it, and the ad impression or course sale would compensate the writer. That logic depended on the human reader being the recipient. When the recipient is an LLM that ingests your article and synthesises it for someone else’s question, the ad never fires and the course is never bought. The ecosystem of practical knowledge production that ran on this loop is, slowly, starving. I wrote about a piece of this in the meaning of work in an age of abundance and the dynamics are accelerating.
The traditional career ladder in knowledge work. Most knowledge professions were structured as apprenticeships. You spent your twenties doing the cheap, repetitive cognitive work - the document review, the financial model, the QA pass, the support ticket - and that work both produced value and trained you. When the cheap, repetitive cognitive work is the first thing the machine takes, the bottom rung of the ladder is gone. I wrote about the junior developer pipeline problem in the context of software, but the same shape applies in law, medicine, finance, and consulting. The middle of the ladder is fine for now. The bottom is missing. In a decade, the middle is wondering where the next generation came from.
The credential as a signal of competence. A degree was a costly signal that someone had spent four years acquiring a body of cognitive skill that would be useful at work. When the cognitive skill itself is in the air, free, at any moment, the signal value of having it pre-loaded in your head decays. We will keep credentialing for a long time out of sheer institutional inertia, but the underlying economic logic is gone. What matters is no longer who knows things. It is who can use the thing that knows.
The geography of the office. Cities priced their property on the assumption that knowledge workers had to physically cluster to be productive. Some of that clustering was genuine - dense networks, accidental encounters, the irreplaceable bandwidth of being in the same room. Much of it was an artefact of needing to be near the pool of competent cognition. As that pool moves into the cloud, accessible from anywhere at any hour, the geographical anchor weakens. This will not empty cities. It will reshape which cities matter, and which streets within them.
The Jevons trap
The naive version of “intelligence going free” is “and so we will need fewer people to do cognitive work”. That gets the second-order effect exactly wrong, and the trap is well documented. It is named after William Stanley Jevons, who pointed out in 1865 that more efficient steam engines led to more coal being burned, not less. When the price of using something falls far enough, total consumption tends to rise rather than fall.
Apply this to intelligence. When the cost of running a model on a problem drops by a hundred-fold, you do not run the model on the same problems a hundred times more cheaply. You run it on a hundred times more problems. Things that were never worth analysing get analysed. Documents nobody would have read get summarised. Code nobody would have refactored gets refactored. Data nobody would have explored gets explored. Decisions that used to be made on gut feel get a memo. The total volume of cognitive work in the economy goes up, not down.
This is the part that is genuinely hopeful, or at least genuinely complicated. The story is not “the machines do all the thinking and humans become decorative”. It is closer to “humans do a hundred times more thinking, mediated through machines, and the bottleneck moves from cognition to taste, judgment, and accountability”.
But the Jevons effect is unevenly distributed. The new demand does not necessarily land in the same job, the same person, or the same country as the old demand. The displaced junior analyst is not automatically the curator of a hundred AI agents in their place. That transition has to be designed, funded, and lived through, and history suggests we are not very good at the transition part.
Where humans accrue premium
If competent cognition is cheap, what is left expensive? The honest answer is: anything the machine cannot do, anything the machine is not allowed to do, and anything where someone needs a person to be answerable for the result.
Three categories survive comfortably.
Judgment under accountability. A doctor’s value is not only in reading the scan. It is in being the human who is named on the chart, who can be sued, struck off, or trusted by a frightened patient. The same applies to lawyers, auditors, engineers signing off structural drawings, pilots, surgeons. The accountability layer is irreducibly human for legal, ethical, and emotional reasons that no efficiency gain unwinds. The machines do the cognition. The human takes the responsibility, and that responsibility has a price.
Taste, in domains where the customer’s preference is the product. When the production cost of competent creative output collapses, what remains scarce is the ability to choose well from infinite options. A film director becomes more valuable, not less, in a world where any shot can be generated in any style. The shots are free. The eye that picks the right one is not. This generalises - to design, to writing, to product, to architecture, to anything where the human’s taste is what the buyer is actually paying for.
Real-world coupling. Anything that requires a body in a place at a time, or a relationship that took years to build, or an institutional position that cannot be cloned. Plumbers, surgeons, teachers in the room with a child, sales people with deep customer relationships, politicians, religious leaders, therapists. The list is longer than people think and is not particularly correlated with current prestige.
What gets squeezed is the wide middle. The competent professional whose value was that they could think clearly about a domain for forty hours a week. That worker’s job did not vanish. It became a curator role - directing, checking, and adjusting the output of systems that do the underlying cognition. Some workers will love that transition. Many will not. And the institutions around them are not yet structured for it.
The political problem
Free intelligence is a strange resource because the surplus it creates is enormous, but the surplus does not distribute itself. By default it accrues to whoever owns the model, the chips, and the data. That is currently a small number of companies in a small number of countries, and the existing incentive structure points toward more concentration rather than less.
This is not a new dynamic. Every previous general-purpose technology - the steam engine, electrification, the internal combustion engine, the computer - generated enormous productivity gains and then spent decades being negotiated over. Who captures the gain? Who pays the transition costs? What does the worker do who used to operate the loom? The answers in each case were political, not technological. Same here.
I am not optimistic that the political answers will be quick or graceful. The previous transitions took fifty to a hundred years and produced revolutions, mass migrations, and several world wars before settling into something workable. The AI transition is happening on a much faster clock and our political institutions are not particularly faster. The arithmetic is not encouraging.
What can be said with some confidence is that the question is no longer whether AI capability arrives. It does. The question is who captures the surplus, who absorbs the transition cost, and what the institutional response looks like. The model trajectory is largely fixed. The political response is not.
What to do, if you are sitting inside it
This is not a essay where I pretend to have a tidy five-point action plan. But there are a few things I would say honestly to anyone trying to navigate this regime, in roughly the order I would prioritise them.
Build deep familiarity with the tools, not as a hobby but as a working skill. The gap between people who can direct AI systems effectively and people who cannot is already large and widening. Sitting it out is not a neutral choice - it is a slow drift toward the bottom of the cognitive labour market. I wrote a practical learning path for the people I work with who asked me where to start.
Move toward roles where you take responsibility, not where you produce output. The cognitive output is going to be cheap regardless. The accountability for what is done with it is not. If your job description is a list of artefacts you produce, you are vulnerable. If it is a list of outcomes you own, you are less so.
Develop taste in something. Not in the affected sense, but in the working sense - the ability to look at ten options and reliably pick the right one for the situation. Taste is one of the few human capabilities that gets more valuable, not less, as production costs collapse. It is also one of the hardest to acquire and the slowest to fake.
Think hard about what your industry does when the marginal cost of its core cognitive work drops to zero. Most industries have not actually done this exercise. The ones that have are reorganising. The ones that have not are about to discover, abruptly, that their pricing model was a happy accident of an old world.
The closing point
The free intelligence era is not coming. It is here. The price curves are public, the trend lines have been published for years, and the second-order effects are starting to show up in earnest in legal, medical, and software work this year. What is not yet here is the matching reorganisation of careers, institutions, prices, and expectations.
We are living in the gap. The new regime exists. The old regime’s prices, careers, and assumptions are still standing. The collision between the two is the next decade of work, and it is the part of the AI future that will affect the most lives the most directly.
I do not think it will be a catastrophe. I do not think it will be a utopia. I think it will be the messy, contested, century-long renegotiation that every previous general-purpose technology produced, compressed into a much shorter window than we are used to. The interesting work, for the next ten years, is in being honest about what is happening and trying to land somewhere useful inside it.
The thinking is going free. What we do with that fact is still up to us.