The missing layer: what $2.6 trillion didn’t buy

This weekend, the leaders of the world’s three most powerful AI companies agreed, publicly, that the industry should slow down. Anthropic’s CEO published an essay calling on the frontier to pace itself; OpenAI’s chief endorsed it within hours; xAI’s owner replied, “Dario is right.”

Whatever one makes of the motives — genuine caution, competitive positioning, regulatory pre-emption; reasonable people are asking all three questions — the debate is about speed. And the most interesting data of the year suggests speed was never the problem.

The balance sheet

The world will spend roughly $2.6 trillion on AI in 2026 alone — up 47% on last year, with infrastructure taking almost half of it.¹

In August, McKinsey published its State of AI survey of 1,719 executives, measuring what that spending has bought so far.² The results deserve to be read slowly: 80% of respondents say AI has improved their individual productivity. Yet only 37% can attribute any impact at all to their organisation’s earnings — unchanged from the year before. Just 6% qualify as “high performers.” Put the other way round: for 94% of enterprises, record spending has not moved the needle. And beneath the aggregate, a detail rarely quoted: 47% of mid-level managers report AI-related strain. The tool is not only failing to reach the bottom line — it is tiring the people who use it.

Individual productivity up. Organisational results flat. Trillions in; faster emails out.

A pattern with a history

Economists have seen this exact shape in every general-purpose technology since steam. When factories first electrified, productivity barely moved for decades. The reason, documented in Paul David’s classic study of the dynamo,³ was that factory owners bolted electric motors onto buildings designed for steam: one giant power source, machines arranged around a central shaft. The gains only arrived — forty years later — when factories were redesigned around the technology: distributed motors, reorganised floors, retrained workers, new ways of managing.

The lesson held through computing (“You can see the computer age everywhere but in the productivity statistics” — Robert Solow, 1987) and it is holding through AI: the returns never come from the tool. They come from the complementary investment — redesigned processes, trained people, and systems built around the humans doing the work.

We built a Formula 1 car and handed the keys to everyone who “knows how to drive.” The lap times should not surprise anyone. Nor should the current debate about whether the car is too fast.

The person got a rounding error

How lopsided is the split between tool and human? The numbers are almost embarrassing to place side by side.

US employers spent $102.8 billion on all corporate training in 2025 — about $874 per learner per year.⁴ The world will spend twenty-five times that on AI in 2026. Within AI budgets themselves, training and change management run at 8–12% of spend — consistently the most underbudgeted line, and the first one cut. Grant Thornton’s 2026 survey found training to be the single most underfunded AI investment area — and that only 6% of executives consider workforce enablement a top skill for an AI-driven organisation.⁵ Economist Impact put the same paradox in one sequence: 88% of executives call AI a competitive advantage; 38% budget adequately for AI training; 4% achieve repeatable, scalable value.⁶ That progression is not a footnote to the ROI crisis. It is the explanation of it.

And on the shop floor, the void is total: ManpowerGroup’s 2026 Global Talent Barometer found that 56% of the global workforce received no recent training at all, and 57% have no access to mentorship — while AI use among workers jumped to 45%.⁷ Adoption without preparation. A Formula 1 grid where nobody funded the driving lessons.

Governments caught the same fever

This is not only a corporate reflex. Look at where the new public money goes. The EU’s InvestAI programme mobilises €200 billion — for infrastructure. France pledged €109 billion — for compute. The UK announced a £500 million sovereign AI fund this very week, aimed at firms and supercomputing access — while the training sector publicly warned that the money “will flow into a workforce that is not ready to absorb it.”⁸ Set against it: a £1.5 billion skills package covering every sector of the economy.

At every level — the company, the government, the race itself — the pattern repeats: the tool has a lobby. The user doesn’t.

Manufactured insecurity — and what it costs

There is a second bill attached, and it is paid in people.

The same ManpowerGroup study recorded something unprecedented in its series: worker confidence fell for the first time in three years — with confidence in using technology dropping 18% even as usage surged. 43% of workers now fear automation may replace their job within two years, up five points in a year. Researchers have a name for the behavioural result: “job hugging” — people clinging to roles they have disengaged from, out of fear of what is outside.

Some of that fear is manufactured. Workers have absorbed a years-long stream of contradictory messaging — mass-displacement predictions one week, trillion-dollar utopias the next, and now the industry’s own leaders debating whether their creation is moving too fast to control. Whatever the merits of each headline, the cumulative signal a worker receives is: your future is being decided elsewhere, and nobody is preparing you for it.

Insecurity of that kind is not free. It converts into disengagement — which Gallup prices at roughly $10 trillion a year⁹ — and into turnover, where the economics are brutal: replacing a single employee costs one-half to two times their annual salary, and voluntary turnover alone costs US businesses close to $1 trillion a year.¹⁰ In hospitality, where turnover runs at roughly double the all-industry average, the sums compound shift by shift.

Follow the chain: underinvest in people → people feel disposable → engagement falls → turnover rises → the returns on the tool never materialise → conclude the tool is the problem. The industry has reached the last step — and proposed slowing the tool down.

The wrong lever

That is why this weekend’s debate, for all its gravity, is aimed at the wrong lever.

The problem is not decelerating the development of AI models. It is accelerating a mechanism that meets the real needs of the people who work — security they can bank, training they can access, progression they can see, and a share of the value their hours create. Pause the frontier for a year and none of those appear on their own. Build them, and the returns the trillions were chasing finally have somewhere to land — at whatever speed the models advance.

Most of the world’s work is physical, present and human — the shift served, the room turned, the guest looked after. AI will not carry three plates across a dining room at any pace the frontier chooses. That work needs architecture, not automation — and it needs it whether the models accelerate, pause or plateau.

Let the frontier pace itself. Accelerate the floor.