When AI is more expensive than people, why replace the people?

Andrew Orlowski is a technology journalist who writes a weekly Telegraph column every Monday. He founded the research network Think of X and previously worked for The Register. You can find him on Twitter @AndrewOrlowski

The free lunch is over. Google says demand has risen sevenfold over the past year. But now the real cost of AI is finally beginning to emerge, and the consequences are cataclysmic.

“This is the first time ever that I can remember that technology costs the same as people,” says Arvind Jain, one of Google’s first and most distinguished engineers-turned-AI entrepreneur.

Sevenfold means nothing, because the true demand signal has been hidden by massive cross-subsidies and the biggest capital splurge in history. Now prices are rising alongside demand suppression measures like feature restrictions and rate limits.

“The industry is simply not organised around real market signals,” Matt Stoller, the competition crusader, said last week.

It has been a surreal interlude in our economic history that to someone in a coma since 2022, the following would read like a satirical work of science fiction:

Driven by a fear of missing out, management introduced a new key performance indicator.

Rather than judge the AI projects by the quality of their output, tech firms would reward AI consumption instead.

Burning through tokens, the building blocks of AI became a competitive metric for employees. This was dubbed “tokenmaxxing”, and company leader boards showed off who had burned through the most tokens, in a strange parody of the Soviets’ Stakhanovite contests for workers.

But this created perverse incentives. Employees were rewarded for getting AI to perform pointless tasks, such as reading their email. Want to know the time? Don’t look at the clock, but ask the chatbot. More tokens burned each time, more on your personal score.

After every binge comes a hangover, and this is arriving in the form of eye-watering bills. Amazon reportedly burned through half a billion dollars worth of tokens in one month. Now it pleads with employees to stop.

“Please don’t use AI just for the sake of using AI,” David Treadwell, a senior executive at the company, begged last week. Uber, Salesforce, Meta and Microsoft have all done similar.

As the true demand emerges, two apparent certainties of the AI revolution are cast into doubt. The first is the belief that the AI we have today – which cannot count to 10 – is about to cause a massive dislocation of the labour market.

On BBC’s Question Time last week, Fiona Bruce seemed almost possessed, aggressively forcing every panellist to respond to this prospect, as if it were a foregone conclusion.

But human replacement will only be attractive once AI starts showing some real value, and real productivity gains. As The Telegraph reported last week, Goldman Sachs found that AI spending associated with engineering roles is approaching 10pc of what it costs to pay a human, but at current growth rates, the two could be on par within months.

“That ‘tokens got burned for millions of dollars without any real significant return on investment to show for it’ might well turn out to be the epitaph for an era,” says Professor Gary Marcus.

Admittedly, Rachel Reeves has made hiring humans much more expensive. But perhaps still not expensive enough to make a fundamentally unreliable and increasingly expensive automation tool seem attractive.

Secondly, if business isn’t seeing value, then the case for all that expensive AI infrastructure also collapses with it. Especially as those investments depreciate so rapidly.

Let’s do the maths. Will Sommer, an economist forecaster, explains that when capital looks at an Amazon, a Microsoft or a Google, it sees a 25pc return on capital invested (ROIC).

Below 12pc ROIC, capital drifts off elsewhere. Below 7pc, it’s “an unmitigated disaster for all of the investors in this technology”, Sommer told The Verge.

To hit even that 7pc return, Sommer estimates that at current token costs, the big AI hyperscalers need to bank $7tn in revenue over the next three years to meet their spending commitments. And token consumption would need to increase by 50,000 to 100,000 times over what it is today.

Recommended

So out of the great cloud of AI hype, the outline of the crash can now be clearly discerned. Humans are turning out to be quite useful after all.

Stoller raises something else disquieting. It shouldn’t be a shock, he says, because much of this was quite predictable 18 months ago, when we could see what China was doing.

It was building very competitive but highly tuned AI models and making them free for anyone to use. The AI labs did not need to monetise them directly – some other industrial sector could choose to use it, or not.

China’s courts now essentially prohibit making human employees redundant for an AI replacement: they reckon that what little productivity gain the AI may yield is not worth the damage to social cohesion. The China issue has aired many times in this column. We have been building the wrong AI.

“It turns out that the Chinese approach is better for producing everything except market capitalisation,” snarked Stoller.

You could almost say that China has hacked the West’s economic system by weaponising the greed and stupidity of our venture capitalists and our pundit class, and turning it against us.