A forecast is a wager dressed in argument, and the honest forecaster owes his reader two things: the date on which the wager can be settled, and an account of why he is willing to make it. What follows is a set of predictions about where artificial intelligence and the industry around it will stand in roughly twelve months — by about May of 2027. They are deliberately aggressive. A cautious observer, surveying the same evidence, would call several of them too bold, and I want to be plain at the outset that I have arrived at them on purpose. For each of the threads below I first worked out the consensus case — the base-rate forecast, the thing a sensible person would expect — and then asked a single further question: where, granting the consensus its due, could this run faster or further or stranger than the sensible person allows, without crossing into the merely silly? The published forecast is the answer to that question. It is the aggressive case, and it is offered as a wager, with the settling date attached.
It is worth being equally plain about the present from which the wager is made, because a forecast is only as good as its grasp of the ground it leaves. The ground, as of this spring, is roughly this. Four laboratories — call them by the families of their models, the GPT line, the Claude line, the Gemini line, and the Grok line — now sit within a narrow band of one another on the public benchmarks, with the leading systems clearing the high eighties on the standard software-engineering test, while the hardest of the contamination-resistant reasoning benchmarks — the ones designed precisely so that they cannot be memorised — remain conspicuously unsaturated, the frontier still well short of the human-expert ceiling on them. The agentic frontier, which is the more interesting number, has been advancing at a pace that the people who measure it describe as a doubling, every seven months or so, of the length of task a model can carry out on its own; the present figure, the length of task a frontier agent completes unassisted at even odds, is now measured not in minutes but in hours — the strong models of the middle frontier at something like four or five hours, and the very leading systems somewhere past the point at which the measurement itself can be trusted, with the doubling showing no sign of slackening. The money is on a scale the financial press strains to render legible: the largest infrastructure builders have committed something near seven hundred billion dollars of capital expenditure for this year alone, sharply up on the year before, and a leaked set of internal revenue figures this April was sufficient to send the whole sector of related stocks downward in an afternoon. The labour effect has begun, narrowly and unevenly, in the hiring of the young. The grid is straining. The public, in the United States at least, has turned decidedly cool. That is the ground. The wager begins here.
The Agentic Frontier Crosses the Working Day
The first and most consequential thread is the one the laboratories themselves watch most closely, and it is not raw capability on examinations — that contest has narrowed to the point of tedium — but the length of task a system can carry through without a human hand on it. The consensus expectation is straightforward and follows the trend line: if the autonomous task horizon has been doubling roughly every seven months, then a frontier agent that today manages a task of some hours at even odds will, in a year, manage one several times longer at the same odds — a full working day, and then some, of the even-odds figure. That is the cautious case, and it is itself remarkable enough.
The aggressive case holds that by May of 2027 the frontier agent will reliably — and I mean by reliably the eighty-per-cent figure, not the fifty — complete a full working day’s worth of software engineering, the kind of task a competent professional would be assigned on a Monday and expect to finish by the end of it, unsupervised, in a sandbox, from a one-paragraph brief. The trend line alone gets one most of the way there for the even-odds number; the aggressive move is to claim the reliable number too, and to claim it because the laboratories have, this past year, visibly redirected their effort from capability toward exactly this. The distinction between a model that can do a thing once and a model that consistently does it is the distinction between a demonstration and a product, and it is the gap every serious deployment now lives or dies in. A field that has noticed a gap and pointed its best people at it tends to close that gap faster than a smooth extrapolation predicts, because the extrapolation is drawn from a period when nobody was trying. The thing I would least like to defend in this paragraph is the word reliably; it is the load-bearing word, and it is the one the trend line does not strictly license.
Software Work Is Reorganised Around the Agent
If the agent crosses the working day, the practice of building software is reorganised around that fact, and the reorganisation will by next May be visibly under way rather than merely discussed. The consensus already concedes a great deal here — that the coding agent reads a codebase, plans across many files, runs its own tests, opens its own pull requests, and that this compresses the manual labour of programming by something like a third to a half. That is the present, not a prediction.
The aggressive case is about the shape of the working day rather than its length. By May of 2027 the median professional engineer at a competent firm will not, as a rule, write code as the first move of the day; the first move will be to specify, to review, and to arbitrate among the work of several agents running in parallel — and the job title will not yet have changed, which is the part that makes it hard to see. The change will be inside the title. A programmer asked at the time whether the systems have altered his occupation will very often say no, and the same programmer asked the same question a year apart will be describing two different jobs. I expect, further, that at least one well-known software company will state publicly, and with figures, that a majority of the code it ships is now written by agents under human supervision — not as a boast on a stage, which has already happened and means little, but in the dry register of an earnings call, where the statement carries liability. The thing I would least like to defend here is the word majority; the honest version of the claim may be that the figure is large and rising and that no one can audit it, which is a weaker and truer thing.
The Money Does Not Burst, But It Cracks
The thread on which I am most reluctant to be aggressive is the money, because the money is where confident forecasts go to be embarrassed. The consensus has by now organised itself into two camps that talk past each other. One holds that the capital expenditure — the seven hundred billion this year, the talk of a trillion and more beyond it — is a sound bet on a real and growing demand. The other holds that it is a bubble, and points to the circular financing, the chip-maker investing in the model laboratory which spends the investment on the chip-maker’s chips, an arrangement of a scale and a shape that the soberer analysts have already compared, without much pleasure, to the vendor financing of the late dot-com period.
The aggressive case is not that the bubble bursts. It is more specific and more interesting than that. By May of 2027 there will have been a genuine and named correction — a sharp repricing of the publicly traded AI-infrastructure stocks, of a severity the financial press will agree to call a crash, occasioned by a particular disappointing disclosure rather than by any change in the underlying technology — and the building will continue through it and past it almost undisturbed. This is the part the bubble camp gets wrong and the boom camp also gets wrong: both of them believe the financial verdict and the physical verdict are the same verdict. They are not. A correction destroys fortunes and reputations and a certain number of the more leveraged firms; it does not, as a rule, destroy the thing already built. The halls will be raised and the chips installed and the power contracted whether or not the particular investors who paid for them are ever made whole, and a year from now one will be able to watch, in the same quarter, a column of red on the exchanges and a column of poured concrete in the desert, and the second column will be the longer of the two. The Victorians argued bitterly over the wisdom of their railways and built the railways; the railways remained when the argument was an antiquarian curiosity. I would least like to defend the timing — that the correction arrives inside the twelve months rather than just outside them. The direction I will defend without hedging.
The Labour Effect Stops Being Deniable
The labour thread is where the aggressive case is easiest to state and hardest to state responsibly, and I want to take some care with it. The present evidence is genuinely mixed. Broad measures of employment show, so far, little that can be cleanly attributed to artificial intelligence; the careful studies find their signal not in the aggregate but at the margin, and the margin is the young. Entry-level postings in the most exposed fields have fallen sharply since 2023 — in junior software work the decline is steep enough that the figure is disputed rather than denied — and employment among workers in their early twenties in the most exposed occupations has fallen by something on the order of a seventh relative to the period before the chat window. The displacement, in other words, has not arrived as a wave breaking over the whole workforce. It has arrived as a quiet closing of the door through which people used to enter.
The consensus expects this pattern to continue and broaden gently. The aggressive case is that within the year it stops being deniable — that by May of 2027 there will be at least one official statistic, of the kind a government agency publishes and a newspaper can put in a headline, that a reasonable person can no longer explain by appeal to interest rates or the ordinary slackening of a labour market. The recent-graduate unemployment rate, already elevated, is the most likely candidate; a sustained divergence between that rate and the rate for experienced workers in the same fields is the sort of thing that is arguable for one or two quarters and undeniable after four. I do not expect mass layoffs of the dramatic kind the early fears imagined, and I want to be exact about why: the firms do not need to dismiss the workers they have in order to capture the saving, because they can simply decline to hire the workers they have not yet hired, and a hiring freeze is invisible in a way a layoff is not. The aggressive prediction, then, is not a flood. It is that the tide, which is already moving quite as much water, finally rises high enough that the people measuring the water agree on what they are looking at. The claim I would least like to defend is the cleanness of the attribution — the economy is a single tangled thing, and a year is a short interval in which to ask it to confess.
The Grid Becomes the Binding Constraint
For most of the short history of this technology the binding constraint was held to be ideas, and then for a while it was held to be chips, and the prediction here is that by May of 2027 it will be neither — it will be electrical power, and this will be understood not by specialists but generally. The consensus already half-concedes the point: everyone who follows the matter knows that the datacentre’s appetite for power has outrun the grid’s capacity to deliver it, that transmission and substations and transformers take years to build and the buildings take months, and that this mismatch is becoming the thing that decides where a facility can go and how large it can be.
The aggressive case adds two things to the consensus. The first is that the constraint becomes acute enough, within the year, to visibly bend the behaviour of the largest builders — that the dominant story in the build-out by next spring is not the signing of compute deals but the securing of generation, and that at least one frontier-scale facility will be brought online running substantially on power it generates for itself, behind the meter, because it could not wait the years the public grid demanded. The second, and the more aggressive, is political: that the local resistance to datacentre construction, which polling already finds to be the majority position by a wide margin, hardens within the year into a genuine and organised obstacle — that the cancellation of projects in the face of community opposition, already recorded at a record rate this past quarter, becomes common enough that the industry begins to treat the winning of local consent as a line item rather than a formality. The thing I would least like to defend is the word generally — the claim that the public, and not merely the trade press, comes to understand power as the constraint. Publics are slow to relocate their attention, and a year may not be long enough.
The Temperature Drops Further
The last thread is the social and cultural one, and it is the thread on which the aggressive case is, perhaps surprisingly, the gloomy one. The consensus, insofar as the industry holds one, expects that familiarity will do its usual work — that as the systems become more useful and more woven into the ordinary furniture of working life, the public’s unease will soften into the same unbothered acceptance that every previous technology eventually secured.
The aggressive case is that within the year it does the opposite. American sentiment toward these systems has been moving in one direction for some time, and moving faster among the young, who use the tools most and have grown most wary of them; the share who report themselves more concerned than excited has climbed to about half, and the enthusiasm of the youngest cohort has fallen sharply in a single year. The aggressive prediction is that this does not bottom out and reverse but continues, and that by May of 2027 a discernible anti-AI position — a settled, articulate, and socially respectable scepticism, the kind of stance a person can hold at dinner without having to defend it as a quirk — has consolidated into something a politician finds it worthwhile to address directly. The mechanism is not mysterious. A technology that is experienced primarily as something done to one — as the closing of the entry-level door, as the datacentre proposed for one’s own county, as the steady degradation of the open web into fluent and confident sludge — does not win affection by becoming more capable. It wins resentment. The industry has been spending its moral energy on the distant prophecies, the catastrophe and the deliverance, and attending hardly at all to the nearer and smaller thing, which is that a great many ordinary people are forming a low opinion of it for reasons that are neither irrational nor likely to be argued away. The claim I would least like to defend is socially respectable — that the scepticism becomes not merely widespread but high-status. Status is hard to forecast, and I may be a year early.
What I Would Least Like to Be Wrong About
A forecast of this kind owes its reader one final honesty, which is a plain statement of where the forecaster’s own confidence is thinnest — not the predictions he secretly disbelieves, for one should not publish those, but the ones whose failure would least surprise him.
The first is the reliability claim in the agentic thread. That the autonomous task horizon will keep lengthening I hold with real confidence; the trend has survived too many years and too many changes of method to be dismissed. But the leap from can to consistently does — from the even-odds figure to the eighty-per-cent figure across a full working day — is a leap the trend line does not strictly underwrite, and reliability has a way of being the last and hardest mile. If one prediction here is wrong, I should expect it to be that one, and I should expect the failure to take the modest form of the frontier landing at, say, a reliable half-day rather than a reliable full one.
The second is the timing of the financial correction. That a sharp repricing is coming I do not really doubt; the circular financing alone is enough to make me sure of the direction. But whether it arrives inside this particular twelve-month window or just beyond it is very nearly a coin toss, and markets have humbled steadier forecasters than I am by remaining buoyant for years past the point at which the arithmetic stopped working. If the correction has not come by May of 2027, I will not regard the underlying judgement as refuted. I will regard it as merely early, which is the forecaster’s most familiar and least dignified condition.
The third is the consolidation of a respectable scepticism. The direction of public feeling I am confident about; the speed and the social standing of the result I am not. Publics can stay diffusely uneasy for a long while without that unease ever hardening into a position one can name, and it is entirely possible that May of 2027 finds the mood exactly as sour as I expect and exactly as shapeless as it is now.
What I notice, setting these three aside, is that the predictions I hold most firmly are the physical ones — that the building continues through whatever the markets do, that power becomes the constraint, that the entry-level door stays closed. The ones I hold most loosely are the predictions about minds: about how reliable a system will feel, about how investors will price their hopes, about how a public will decide to feel concerning a thing it has not chosen. This is, I suspect, the correct distribution of confidence, and it is worth saying why. The physical world is slow and obeys arithmetic; concrete poured is concrete poured, and a transformer ordered today arrives, or fails to arrive, on a schedule indifferent to anyone’s opinion. Sentiment, capital, and the felt reliability of a tool are all, in the end, made of judgement, and judgement is the thing this whole technology has taught us, repeatedly and at some expense, is the hardest of all things to predict. A year from now the wager can be settled. I have given the date; I have given the reasons; I have marked the three places where I would not be astonished to lose. That is the whole of what a forecast can honestly offer, and I would rather offer that, and be checkable, than offer the safer thing and be merely vague.