Before AI can kill us the bubble will burst

Unchecked AI could cause a billion deaths, said Bill Gates last weekend, joining the rising chorus warning us that on its current trend, AI could destroy humanity within a matter of years.

Don’t bet on it. Many doubt AI really has the capacity to make itself into an autonomous super-intelligence, and skeptics suggest the tech lords are making alarmist claims mainly to hype up their share values. But even if, with sufficient capital, they could create a super-intelligence that could overrule humanity, they won’t get that much capital.

There’s a straightforward cycle here, and its math looks bad: The development of ASI (artificial super-intelligence) relies on debt, which is raising interest costs, which are becoming so onerous that the earnings of the hyperscalers driving the boom will fall. That will hinder share prices, limiting the collateral they can use to raise more debt, thereby stopping the flow of new money and ending the party.

In effect, the funding model of the AI boom now resembles an elaborate Ponzi scheme, and it could already be approaching its limits. Let’s work through it. By their own admission, the tech lords are saying that to attain ASI, vast amounts of computing power and energy will be required. The expected capital required for this buildout of data centres and powers, which in the United States is now moving into the trillions, exceeds the companies’ available reserves. So, they are relying on debt, a lot of it.

Opinion: The question isn’t whether we should slow down AI development. It’s whether we still can

Two types of debt are fuelling the bubble. One is the borrowing by the hyperscalers themselves, projected to reach US$250-billion in the U.S. alone this year. The other is the fiscal deficit of the federal government. Over the last six years, the national debt has nearly doubled, rising by some US$16-trillion, growing faster than the economy. That tsunami of cash has entered an economy in which the labour share of output has been declining, the result being a huge glut of savings among the wealthy.

Those savings have been ploughed into the U.S. stock market, helping to fuel the spectacular rally of the last few years. That rally, in turn, has sucked in money from the rest of the economy and latterly, the rest of the world. All this enthusiasm for AI is premised on the claim that the technology will so transform economic productivity that the economy will take off, making the debt irrelevant. Lest anyone doubt AI’s transformative prospects, they just have to look at hyperscaler earnings – Google’s earnings are up 24 pr cent since last year, Meta’s even more.

But look at things more closely, and you see how fragile this boom has become. First off, analysts are running the numbers and are finding that to justify the sort of share prices the hyperscalers now command, future earnings will have to increase as much as 45 times in the next six years – far beyond current rates of increase. Worsening the picture is that today’s double-digit rates of increase are themselves inflated. Hyperscaler revenues are big, sure, but only keeping pace with their equally big expenditure. They don’t reflect the sort of increase in margins one would expect if profits were getting a boost from productivity gains – in other words, they’re making money because they’re spending money.

‘Bond king’ says U.S. stock market is like a diseased tree

As for the productivity transformation AI will produce? As with Godot, there is endless talk but it never appears. In fact, what modest productivity increase we have seen in the U.S. economy has come mainly from firms that have not adopted AI, while one recent survey of firms that have done so found that “almost every report at a company about ‘massive AI productivity gains’ is untrue.”

With the combined borrowing of governments and hyperscalers now exceeding the supply of credit, the cost of borrowing rises by the day. Climbing interest rates will now eat into corporate earnings and eventually cause a rotation out of stocks. We may have already entered the slowdown, with the stock market rally having noticeably slowed since the spring. In time, this will limit the ability of the hyperscalers to incur more debt.

In short, the oxygen is coming out of the bubble. In the best case, AI runs out of steam; in the worst, it crashes. Neither scenario will end its advance, but after the slowdown AI will enter a new, less alarmist phase, as firms pivot away from pursuing ASI to monetizing their existing technology – producing practical applications and cheap, efficient models. Already, we’re seeing moves in this direction as business consumers switch to Chinese models and Western startups offering more efficient options.

The dream of ASI may not end. But its pursuit will slow as funding is reallocated, which will give society more time to decide whether and how to use and regulate it.