On 16 July a Beijing company called Moonshot AI released the most powerful open-weight model ever built. Kimi K3, trained on the wrong side of US export controls, ranks third on the Artificial Analysis Intelligence Index, behind only the best American labs. By the end of this month, Moonshot says, anyone on earth will be able to run it for nothing.

Investors deep in AI-induced delirium are beginning to grasp the obvious: frontier models may not be commercially viable. Intelligence looks more and more like a commodity, closer to electricity than to the winner-take-most economics of Google Search.

OpenAI looks cooked. CEO Sam Altman recently offered Washington a 5 percent stake in the company, dressed up as a scheme for sharing AI’s bounty with the American people. A firm certain of its prospects does not hand equity to the bureaucracy that regulates it. The aim is to entrench OpenAI as critical to American national security, too important to fail.

Leaked financials, verified by the Financial Times, explain why. The business booked $13.1 billion in revenue in 2025 and spent $34 billion, of which $5.7 billion went on marketing: enough to buy every advertising slot in seven consecutive Super Bowls. Its operating loss of $20.9 billion ranks among the largest corporate losses ever recorded. Apologists prefer a figure nearer $8 billion, reached by stripping out the one-off cost of the for-profit conversion and the non-cash charges beneath it. But those are real costs. Someone is confused about the difference between a loss and a cash flow, or counting on you to be.

A business selling below cost has two options: raise prices or subside into the corporate muck. The first option has effectively closed for OpenAI. Chinese open models like K3 have grown from under 5 percent of business usage to nearly half in a single year. Meta has priced its newest model 75 percent below OpenAI’s, and Altman’s firm has been reduced to undercutting its own flagship by 80 percent to hold market share. Anthropic hesitated for weeks before shifting its cheaper subscribers toward the true per-token cost of its flagship model, Fable, while keeping it bundled for those who pay more. Charging full price for intelligence has become something you apologise for.

The American AI giants have misread China. Beijing will subsidise vast losses to dominate a new industry, as it has with solar panels and electric cars. When Washington cut China off from high-end chips, Beijing’s labs answered by giving the models away. The Party has long-term domination, not short-term profit, in mind. No for-profit company can win a price war against an opponent that charges nothing. The price of intelligence will stay pinned below the cost of making it for as long as someone gives it away.

AI is not free, you might object; it burns real energy. If the customer never pays the true cost of intelligence, somebody else must, and here a Victorian economist had the measure of it. In 1865 William Stanley Jevons studied a century of British engineers wringing more work from every tonne of coal and predicted the opposite of what his contemporaries assumed: efficiency would raise coal consumption, not lower it. Cheaper steam meant more engines, more factories, more uses for the fuel than anyone had imagined while it was dear.

Cheaper intelligence means more of it, used everywhere, and an energy bill to match. Coal’s appetite was at least bounded by the number of engines Britain could build. A near-free query has no such limit: it attaches itself to every line of code on earth, and each one draws power from a data centre. The billions the labs lose subsidising intelligence boomerang as demand for chips, buildings and power.

That demand is already reaching households. In June 2025, families in Washington, DC found their power bills $21 a month higher. The rise could be traced upstream with unusual precision: PJM, the grid operator serving 67 million Americans, had seen its price for future generating capacity rise more than tenfold in two years. The grid’s own independent monitor put 63 percent of that increase, some $9.3 billion, on data centres.

A sector worth 4 percent of GDP is now underwriting the American economy. Harvard economist Jason Furman calculated that investment in data centres and information-processing equipment accounted for 92 percent of growth during the first half of 2025. Strip that out and the economy grew 0.1 percent, which is to say not at all. Nine-tenths of the expansion of the world’s largest economy comes from sheds of blinking light.

The comforting precedent suggests this ends well for consumers. Every previous investment mania has finished by making its product cheap: the railway bubble of the 1840s ruined its investors but left Britain covered in railways, and the fibre-optic bubble of 2000 made bandwidth almost free for the generation that followed.

AI will not follow the script. A Victorian rail outlived its investors by a century. An AI chip has an economic life of two to three years, though the hyperscalers spread the cost over five or six. Michael Burry, who called the 2008 housing crash, estimates that gap will overstate industry profits by $176 billion between 2026 and 2028. Satya Nadella, Microsoft’s chief executive, is blunt about why he refuses to stockpile: “I didn’t want to go get stuck with four or five years of depreciation on one generation.”

Electricity is only the first channel through which AI’s real cost hits ordinary people. The second is interest rates. An investment surge on this scale bids up power, equipment and construction at a time when inflation already sits above central bank targets.

The standard answer is that AI’s productivity gains will offset the pressure. But productivity gains are a one-time hit. You cannot automate the same task twice: you bank the saving once, while the spending to build the next round of capacity comes due every year. Bond investor Christopher Joye warns clients to prepare for rate rises across the developed world. His bear case for Australia is a cash rate grinding into the 5 to 6 percent range. Whether AI lifts productivity fast enough to spare us, he concedes, “remains an empirical question.” I am betting it does not.

The politics of this will be uglier than the economics. I argued in April that people losing jobs to AI will not quietly subsidise the machines replacing them. Maine has proposed a freeze on new data-centre approvals, and ten states have followed. Some in the industry dismiss this as Luddism. But a householder whose power bill and mortgage are both rising to fund the machine trained to replace her is not confused about her interests. She can count.

The geniuses on Wall Street, last seen floating SpaceX at 94 times sales to ride-or-die retail investors, are pricing none of it. The bond market still treats the chance of Federal Reserve rate rises as close to zero, resting on the lazy assumption that technology always pushes prices down. There is a great deal riding on that serenity: AI-linked companies now account for a record 45 percent of the value of the S&P 500, and without them and the energy firms they feed on, the index would have gone backwards this year. “This AI thing better work out,” says Torsten Slok, chief economist at Apollo, a credit investor.

By the end of the month, anyone will be able to download Kimi K3 and run frontier intelligence for nothing. The electricity to run it is another matter. PJM has already auctioned its generating capacity out to 2028, at prices set by the data centres. Power is the great exception of the AI age: the one product in this story that will never be discounted.

About the Author

Luke Heilbuth is CEO of strategy consultancy BWD Strategic, and a former Australian diplomat.

On Substack, Luke writes about the systems we’re breaking and the blindness that lets us — from climate and geopolitics to AI and the future of work. Read & Subscribe on Substack here.