Who pays for the race?

For five years, I worked the floor of London hospitality. More than six hundred shifts, around eighty venues, alongside more than a thousand people who carried plates, washed pans and set banquet rounds. I learned every venue from someone standing next to me who was never paid a penny to teach me. So when I read about the money being raised for artificial intelligence this year, I keep asking a question that comes from the floor: who is actually paying for all this?

In Part 1, I argued that the returns on AI were never going to come from the tool. They come from the layer nobody funded: the person. This part follows the money. The answer, it turns out, is almost everyone — and the person on the floor more than once.

The tool no longer pays for itself

The five biggest cloud builders — Alphabet, Amazon, Meta, Microsoft and Oracle — will spend more than $690 billion this year on data centres and chips. For most of them, that is more than their businesses bring in as cash, so the difference is borrowed: new debt now covers about a third of their spending, and they issued around $220 billion in bonds before mid-August.^1,2

Last November, OpenAI’s finance chief went a step further. What the industry needed from government, she said, was “the backstop, the guarantee that allows the financing to happen.” She walked it back the next day.³ But the instinct was revealing. When the tool needs money, the first reflex is to look for someone else’s.

Your pension is in the race

Picture a server at a restaurant in Midtown Manhattan, and a mother of two. Call her Ana. She has never typed a prompt in her life. Until this year, she probably had no retirement savings either: only 17% of American hospitality workers take part in a workplace retirement plan.⁴ In 2026, New York’s Secure Choice programme began enrolling workers like her automatically, at 3% of pay.⁵ It is real progress. It also means the first dollars she ever invests will most likely go into the race.

Britain, where I work, took this step first. From October 2012, employers began enrolling workers into a pension automatically, and by 2018 the duty covered every employer in the country. Nest, the state-backed default scheme, alone now has around 13.8 million members. Their savings are spread across global markets — and the main global stock index, MSCI World, is now about 72% American, with its ten largest companies all from the US.⁶ A kitchen porter in Manchester is in the race too. Nobody asked her either.

The ten largest companies now make up about 40% of the S&P 500 — the most concentrated the US market has been since the 1960s. At the height of the dot-com bubble, the figure was 27%.⁷ And the debt is travelling the same road: as banks reach their lending limits, the loans behind data centres are being sold on to insurers and pension funds. Meta’s Hyperion campus in Louisiana alone carries more than $27 billion of it. Asked where the risk ends up, Deanna Malatesta told The Bond Buyer: “It’s going to end up being taxpayers who end up paying, or ratepayers.”⁸

The canary

This week, the lenders started to flinch. CoreWeave rents out Nvidia chips to Microsoft, OpenAI and others. It has never made a profit and carries about $50 billion of debt. On 14 October its shares fell 15.5% as the interest demanded on its bonds approached 15%.⁹ The cost of insuring its debt against default hit a record.¹⁰ Two days later, the Nasdaq fell 2.1%.¹¹ One analyst called CoreWeave “the AI trade’s canary in the coal mine.”¹²

Whether this is a bubble, nobody knows yet. The structure underneath is easier to read. Nvidia sells CoreWeave its chips, owns about a tenth of the company, and has agreed to buy up to $6.3 billion of the capacity CoreWeave can’t sell to anyone else. Since August, it has offered other operators the same kind of safety net. “These are revenue-generating assets now,” Jensen Huang said.¹⁰

The chip seller guarantees demand for the chips. The lender lends against the chips. The pension fund buys the loan. Every guarantee in the chain is written for the tool.

The bill comes home

A family in Ohio doesn’t use a frontier model to heat its home. It may pay for one anyway. PJM, the grid that serves 65 million Americans, has just cleared a record $16.4 billion capacity auction, driven largely by data-centre demand.¹³ Once the current price cap expires in 2028, the average household there could pay around $70 a month more.¹⁴

Taxpayers are in too. Ohio’s sales-tax break for data centres cost the state more than $1.5 billion in 2025 — eleven times the official forecast — until the governor paused new approvals this summer.¹⁵ Virginia, the world’s largest data-centre market, gave up about $1.6 billion in a single year, and every state that has studied the return has found it “sharply negative.”¹⁶ Britain is spending on a smaller scale, but in the same direction: public AI contracts for compute and infrastructure total about £1.7 billion, against £730 million for skills and the workforce.¹⁷

And the person pays more than once

Go back to Ana. She is now paying as a saver, through her new retirement account. Had she lived in Ohio, she would be paying twice more: as a taxpayer, and as a ratepayer on the bill that keeps her children’s home warm. Had she worked in tech, she might have paid with her job: Meta began cutting around 8,000 roles in May, in the same year its spending on infrastructure nearly doubled,¹⁸ and by the end of May about 143,000 tech jobs had gone in 2026.¹⁹

What Ana is unlikely to get is training. American training budgets grew last year — yet hours of training per employee fell from 47 to 40 in a single year. Per hour delivered, training became roughly a third more expensive.²⁰ Companies are paying more to give people less time to learn. And for a mother working doubles, any hour of training that isn’t paid comes out of the only time she has left: her children’s. In Britain, where I work, it is starker still: employers invested £53 billion in training in 2024, 19% less in real terms than in 2011 and the lowest since records began — about half the EU average per employee.²¹

On the floor, I saw what that looks like. Training in hospitality was never a classroom. It was the person next to you on the shift: how to carry three plates, how to set a banquet round, how to read a room when service goes sideways. That is the industry’s real training system. It runs on goodwill, it appears on no budget, and it leaves the building every time a good worker does.

The fifth line is paid off the clock. After a double shift and her children’s bedtime, Ana reaches for her phone, like most of us. Meta reports that 3.56 billion people use its apps on an average day, and almost all of its revenue comes from selling their attention to advertisers. Gallup’s new State of the World’s Emotional Health report finds that people who spend two hours or more a day on social media report somewhat more stress, worry, loneliness and sadness than moderate users — small differences, and not proof of cause. The broader picture is clearer: in 2025, 38% of adults worldwide spent much of the previous day worried and 36% stressed, both far above where they were twenty years ago.²² The hours people use to recover from work are being sold to help pay for the race.

Why every request is for the tool

I don’t think the people running these companies are villains. Most of them are following four ordinary rules, and all four point the same way.

The accounts. A server is an asset. It goes on the balance sheet, and its cost is spread over years. Training is an expense, recognised the moment it happens.²³ Buy a server and your balance sheet grows. Train a person and your profit falls this quarter. Let a person go and your profit rises. Jensen Huang calls GPUs “revenue-generating assets.” Nobody uses that phrase for a trained waiter.

The tax code. In the US, equipment can now be written off in full in the year it is bought. In the UK, machinery gets full expensing. The tool gets accelerated relief. The person gets an ordinary deduction — and in Britain, since April 2025, a higher National Insurance bill for the employer who hires her.

The collateral. You can borrow against a data centre. You cannot borrow against the colleague who taught you to carry three plates, because she can leave tomorrow — and on the floor, the best people are always the first to go. They have options.

The lobby. In October last year, OpenAI wrote to the White House asking for a 35% tax credit to be extended to data centres and AI servers.²⁴ The tool has letters, proposals and a seat at every summit. Nobody has written a letter for Ana.

Put the four together and the result is predictable. The tool is the only thing in the system that behaves like an investment. The person behaves like a cost — and costs get cut.

Capital funds what stays. Under today’s system, the tool stays and the person leaves.

The design problem

The usual answer is moral: companies should invest in people. They have been told so for decades, and the numbers above show how far that has got us. Conscience doesn’t move capital. Structure does.

None of this is an argument against the tool. I have spent more than four decades building software, and I use these models every day. It is a reminder of where the tool comes from. Every model in this race was designed by people, built by people, and trained on what people made: their books, their code, their conversations, their craft — the accumulated intelligence and creativity of millions of workers who were never asked and never paid. The person is not only financing the race. The person is the reason there is anything to race.

The person isn’t just paying for the car. The person built it.

So the real question is not whether to invest in people. It is how to make that investment behave the way investment in tools already does: something that stays, grows, and can be checked. If training stops walking out of the door, it becomes an asset. Once it is an asset, the money currently borrowed for the tool finally has a second place to land.

That is a design problem, not a charity problem. It is the problem GOE was built around: a work system where the people doing the work have their own reasons to stay — so that the person who teaches the next one is recognised and paid for it, and what they pass on keeps building, hour by verified hour.