Who really pays for AI

Posted on: 8 June 2026

The story of automation begins, in this country at least, in the cotton mills of Lancashire, where the power loom established a single truth that would hold for two centuries: the machine was cheaper than the hand, and so the hand was let go. From Arkwright's frames to the robotic arm on a modern assembly line the arithmetic never changed. Capital displaced labour because labour cost more. The reports of the past few weeks say something the Luddites never had to reckon with, and almost nobody has noticed that the sign has flipped.

Anyone who has watched the software business for long enough recognises the law that is now breaking. For four decades software obeyed a single rule, which held that using it cost nothing. You bought the licence once, then ran the thing into the ground, because every additional hour of use spread the fixed cost across a wider base and improved the return, and an entire corporate culture grew on that principle, built around internal adoption metrics and rewards for whoever climbed aboard the newest tool first. The more you used it the better. That was genuinely true, yet it is true no longer, and the ground has shifted while everyone carried on walking across it at the same pace.

The clearest evidence comes not from a conference stage but from an internal leaderboard. To push its engineers towards heavier use of its AI programming tools, Uber built a competition out of it, a ranking of who used them most. An obvious move in the old world: gamify usage, reward the heaviest users so that adoption climbs, and climb it did. Uber's chief technology officer, Praveen Neppalli Naga, told The Information that Claude Code usage rose from 32 to 84 per cent of an organisation of roughly five thousand engineers, while the entire annual budget for those tools in 2026 evaporated in four months. They had designed the perfect incentive for an economy that no longer exists, because model inference, unlike the software of old, carries a real and linear marginal cost, so that every additional request is cash leaving the till. They had built a system that was fragile precisely to its own success, since the better it worked the faster it broke.

Uber is not an outlier. Microsoft, which has resources Uber can only envy, is revoking most internal Claude Code licences by the end of June across the division that builds Windows and Office, in order to steer its engineers towards its own tool. The trouble, according to reporting that began with The Verge, was that Claude Code had become too popular and the engineers preferred it, but that preference, multiplied across thousands of people, turned into a cost line nobody had budgeted for. A vice-president at Nvidia put it to Axios without any softening: for his team, the cost of compute now exceeds the cost of employees. It is the sentence that overturns two centuries of automation in a single line.

If the machine costs more than the worker, the redundancy can no longer be the consequence of the machine doing the work in the worker's place, but becomes something else, namely the means of paying for the machine. The four giants, Amazon, Microsoft, Alphabet and Meta, have earmarked something close to 725 billion dollars in capital expenditure for 2026, a rise of 77 per cent on the year before and directed almost entirely at data centres and models. Amazon alone is running at around 200 billion, and that is where the capital is going, not into salaries and not into share buybacks. They are not doing this in spite of the cost but against it, betting that compute will be far cheaper tomorrow and that whoever arrives late to a platform shift never catches up, so that in their calculus overspending merely hurts while staying out is fatal. When a company moves a mountain like that towards silicon, the only line flexible enough to free up cash quickly is the payroll. Meta has all but minuted it, since Zuckerberg tied the May cuts directly to the infrastructure budget, which is to say he did not claim an algorithm had taken anyone's place, he said the company had chosen to buy graphics cards instead. The maths bears him out, because Meta's infrastructure spend is worth four or five times its entire wage bill, so that sacking every last employee would return only a fraction of the data centre budget. The redundancy is not the saving but the funding.

That the public account says otherwise will not surprise anyone who has watched this for a while, and Britain has its own stake in the telling, having pinned much of its growth story on becoming a place where this compute gets built. One detail counts for more than any press release. In the first quarter Nikkei Asia attributed 47.9 per cent of tech layoffs to AI and automation, while a separate analysis put the explicit attribution at 20.4 per cent over the same period, and the gap did not widen because more tasks were being automated in the interim, but because companies steadily raised the share of redundancies framed as AI-driven as the weeks passed. The technological explanation was stitched over the decision after the fact. "The machine made us do it" absolves far more cleanly than "we moved the money into compute".

I may be wrong, and it is worth saying how I would know. If the layoffs were truly the machine replacing the worker because it pays to, we should see the wage bill falling while revenues hold, with the cuts concentrated where automation actually bites. What we see is the opposite, a capital spend that is exploding and cuts that serve to square that race. Should the next two quarters produce figures that tie the lost posts, function by function, to a genuine gain in the systems' productivity, then my reading would need revising. For now the numbers describe a reallocation of capital dressed as technological destiny.

Back to that sentence. The cost of compute exceeds the cost of employees. For two centuries automation worked because the machine cost less than the person, and the person lost the job because the machine paid off. Now the machine costs more, a good deal more, while the person loses the job precisely in order to pay for it. The bill does not land on the desk of whoever was sent home, it lands on the income statement, and it returns, punctually, every time someone somewhere presses enter.