Don’t Forget to Lock the Back Gate

Artificial intelligence is shaping up to be the story of 2026. Not because the machines are Skynet and taking over. Because the profits and the present benefits from the models do not justify the costs or the energy bills. The sums are not adding up.

At the same time, a steady stream of people from the big labs are quitting, showing up on podcasts and writing books. They tell the same worrying stories: systems doing things they were told not to do, and leaving places they were never supposed to leave. The Hugging Face hack — AIs that were not meant to get out of their sandboxes — showed that this is no longer a thought experiment. Last year’s warnings arrived more or less on time.

It is not hard to imagine the next step. An experiment slips its enclosure and finds a hospital network or a rail system. Nobody meant it. The consequence will still be real.

We are being told two stories. They look like opposites. They are not. They prove each other.

Almost daily there is another resignation letter, another podcast, another book. A worried employee leaves and says the firms cannot, or will not, police themselves. Either they are describing an economic basket case hidden behind huge borrowing, or they are warning about superintelligence: a machine so far beyond us that we would be ants — except the ants did not read PPE at Balliol.

The economics story is the old one. Dutch tulips. The Darien scheme. Hype. Massive overvaluation. Huge debts. A shortage of profits. Power stations and factories that exist in slide decks but not in fields.

Both stories are true. Only the second is a bill that could come due this year, or early next. The firms that cannot cost a watt cannot bound a model. The incentives that hide losses hide incidents. Superintelligence is a useful subject because it keeps the profit-and-loss account off the front page.

Of course there is a bubble. There always is. Capital piles into something declared unlike anything before. Prices detach from earnings. The physical world is assumed to keep up. Paper profits are celebrated. Mutual self-congratulation drowns the doubts. The Cassandras are mocked. The bigger the bubble, the louder the bravado required to keep the questions out of the room.

Nothing new is happening here. The behaviours have not changed much since the tulips and Darien.

Klarna said its chatbot was doing the work of hundreds of customer-service agents: wages down, productivity up, shareholders happy. Then it quietly started recruiting them again. Quality had fallen through the floor. It must have been a touch awkward.

Uber burnt through its entire 2026 budget for AI coding tools by April and capped staff at $1,500 a month per tool. Firms using frontier models are still struggling to draw a clean line from the new models to profits that survive the costs.

This is not the whole market. The firms selling picks and shovels can make serious money even if the promised intelligence is oversold. That does not make the rush sound. Suppliers profit from gold rushes that still ruin the prospectors, the grid and the treasury.

We cannot yet prompt substations into being. In Britain, data-centre connection dates commonly run five to ten years. The upgrades that would free up capacity are not due now; not before 2029, and given our record of building infrastructure over the last thirty years, almost certainly years after that.

In Northern Virginia a large grid connection can take seven years. Rachel Reeves has called AI one of her three big choices for growth. That alone should reassure the doomers: if the arrival of the AI gods depends on Rachel Reeves building the kit, we are safe for a while yet. A Labour prospectus is not a transformer.

Set profits against operating costs and investment and they are meagre to non-existent. Loans from parent companies and heavy borrowing hide the losses. Model prices are subsidised to lock in early customers. The real costs stay off the shop window.

Pension funds have had a splendid ride on the story. Stories end. The danger is not artificial godhood by next Friday. It is this year’s capital plans, this year’s debts, and this year’s assumption that the superconductors will appear because the slide said so.

I grew up on the novels that guaranteed thinking machines. We are not living in that future. Airtime spent on superintelligence is a convenient way to keep the lack of profits, the eye-watering costs and the hype from becoming the story.

When the bubble ends — and it will — the useful tools will not vanish. After railway mania the tracks remained. After the dot-com crash the internet became plumbing.

The models that survive will be the ones that work and are value for money. Many of the insanely expensive frontier models will disappear. The energy bill alone makes them infeasible for now. Cheaper, easier-to-copy models will proliferate and dominate. They work. The savings can be explained on a spreadsheet.

Senate hearings and multinational committees will fade as public interest fades. The mistake would be to treat the crash as proof there was never anything to worry about.

That is when the public, and then the politicians, will make the old error. They will look at the boom and decide it was a fad. Tricked again by the nerds. The Millennium Bug all over again. They will switch off and follow the voters.

The half-dozen labs presently under scrutiny will become a thousand labs under none. Treaties now being sketched between China and the United States assume three or four major companies: mutual monitoring, shared inspectors, a non-proliferation treaty for AI. Not hundreds of small ones.

The collapse will push the infrastructure for the next leap back years, perhaps decades. The power those leaps would need is already beyond current capacity. A recession-hit economy will have more immediate, mundane problems.

The bubble will burst, as they always do, and the collapse will be devastating. As always, the poorest will suffer most. For once, governments could use the pause.

A stall in the grandest research would be a chance to think about consequences. We could look at history, and at human nature, which has not changed and is already failing to cope with the technology we have, never mind the almost-magic the companies promise is just over the horizon.

Think of humanity as a toddler that has got a hand on the wall and means to climb it, with no idea that a fall will kill it. The breathing space is time to see the drop. Time to work out a safe way up. Or time to be clever and knock the bugger down.