Normally, if someone told you a company had reported operating profit up 557% year on year, you would assume somebody had made a mistake in the maths.

Normally, if someone told you a company had reported operating profit up 557% year on year, you would assume somebody had made a mistake in the maths.

SK hynix did exactly that. Its shares then fell nearly 10%, because apparently a sixfold increase in profit is no longer enough when the market has priced in an AI boom without limits.

Chipmakers have benefited enormously from the pursuit of ever more powerful frontier models. They have also brought about enormous price increases affecting buyers of almost every modern electronic device, from phones and computers to servers and cars. Gartner expects the combined price of DRAM and SSD storage to rise by around 130% this year, with the cost eventually flowing through to the products we buy.

The connection to AI is fairly direct. Manufacturers are prioritising high-bandwidth memory and other components needed for AI servers, leaving less production capacity for conventional memory and storage. There is little sign of the pressure easing soon, possibly not until 2028, by which point Chinese manufacturers may have added enough capacity to change the balance.

What interests me is the capital expenditure behind all this. Hundreds of billions are being committed to accelerators, memory, data centres and power infrastructure, much of it built around hardware whose economic usefulness can deteriorate within a few model generations. The machines may remain perfectly functional, but that does not mean they remain competitive.

In almost any other industry, proposing this level of spending on assets that depreciate so quickly would lead to some uncomfortable questions about cash flow and returns. In AI, it is treated as the unavoidable cost of remaining in the race.

My question is what part of this is sustainable, because it increasingly looks like an arms race that nobody can afford to lose and nobody knows how to win.

By Anthropic’s own account, Claude now writes more than 80% of the code merged into its internal codebase, under human direction and supervision. That does not mean Claude is doing 80% of the engineering, because people still decide what should be built, judge the output and take responsibility for whether it works. It does, however, suggest that current models are already extraordinarily useful.

So what justifies the next tens of billions? What frontier are the labs still trying to cross that would produce enough additional value to repay the chips, power stations, data centres and depreciation required to reach it? Greater reliability and lower operating costs would be commercially valuable. Longer autonomous work, scientific discovery and robotics may be genuinely transformative. But it is also possible that the race continues partly because no company wants to discover what happens when it is the first to stop.

I believe AI is already useful. What remains unclear is whether the next leap in capability will be large enough to pay for the infrastructure race required to reach it.