Replace staff with AI before it becomes too expensive

Will AI take our jobs? Yes, maybe. If the calculations work out.

This debate has been waging since generative AI broke through, but it’s gained new momentum of late now that Claude Code and agentic AI have proved capable of meaningful advances in programming and large companies have referenced AI when reducing headcount (regardless of whether AI is really taking over those roles). 

From a macro perspective, however, AI’s impact on jobs is barely visible, although trends suggest fewer junior developers are being hired. 

That AI is having a significant impact on the developer profession is clear. The big assumption in headline these days is that AI will have the same impact on all other white-collar professions. Whether the technology is that capable is debatable (I doubt it), but if we buy for the moment that generative AI is good enough to replace humans in their jobs, then assumptions of gigantic proportions remain. 

The AI economy’s honeymoon phase

If we look at the current state of the “AI economy,” many still seem to ignore the fact that AI use is heavily subsidized by its providers. Just as Uber was cheap in the beginning thanks to billions in venture capital, AI companies sell their services for a fraction of what it costs to train and run the models. Even if revenues are growing, mainly on the corporate side, margins are being eaten up by huge investments in chips and data centers. Sooner or later, this will have to change.

In a way, we are in a honeymoon phase. Many companies are counting on AI profits based on today’s prices, without taking into account that we are in a technology rally where prices are being kept down artificially. It can be compared to the period of zero interest rates a few years ago, which created a false profitability in projects that were not really self-sustaining and where companies were able to hire lots of people they would not being able to afford in the long term — the reason behind many of the layoffs we are seeing now.

This does not prevent companies, especially those driven by the quarterly economy, from calculating what they can do right now. And companies that act today may well take advantage of subsidized AI prices before the market corrects, to reduce their costs and increase their margins. But it also becomes a kind of first mover advantage, because if everyone does the same thing, the advantage disappears. Assuming everyone can even do it.

Because at the societal level, the calculation starts looking very shaky. AI companies are struggling with capacity for today’s relatively limited use. Anthropic has introduced limits on its most loss-making Claude Code subscriptions to give capacity to corporate customers who pay the most. All while waiting for more capacity, more chips, more data centers. A situation, again, that is mainly about a single profession, in a single use case.

Computing power reality check

Replacing a significant portion of all white-collar jobs with AI — hundreds of millions of jobs — will require computing power we don’t have anywhere near today. Of course, that’s why AI companies and cloud giants are promising to invest astronomical sums in new data centers that haven’t been built yet.

But even with all the capacity and data centers being built, prices need to be kept down despite a sharp increase in demand — and a sharp increase in energy costs. All these data centers require electricity, which is not an unlimited resource. Just last week, it was announced that Open AI is pausing its “Stargate” data center in the UK because electricity costs are becoming too high. 

AI vendors are thus faced with the task of making AI cheap enough to make it profitable to replace human labor, and expensive enough to pay for the largest infrastructure investments, and likely operating costs, in history. Otherwise, nothing will happen.

As stockbroker Citadel Securities laconically states: “If the marginal cost of compute rises above the marginal cost of human labor for certain tasks, substitution will not occur.”

Above all this hangs the fact that AI companies’ primary interest is not in making your accounting more cost-effective. Rather, it is a necessary evil in the fight to be the first to create a digital god, AGI.

When the bill comes due

A counterargument is that AI chips are becoming increasingly resource-efficient and therefore cheaper to operate. And that’s true; Gartner, for example, predicts the cost per token for AI inference will drop by 90% in the coming years. However, that doesn’t mean lower prices, as chatty AI agents’ token usage is increasing even more. “As token consumption rises faster than token costs fall, overall inference costs are expected to increase,” the research firm writes.

Another counterargument is that you can’t just look at the AI ​​lab’s frontier models. The existence of open models, and specialized smaller models, changes the calculation. And that’s true. The problem is that the best open models today are mostly developed in China, which has its drawbacks. Development needs to pick up speed elsewhere for open models to make a real difference. But the inference cost doesn’t disappear just because the model is open.

Smaller models are definitely going to be left behind. They could change the game going forward, because today we probably use more advanced AI than is needed in many cases. This is due both to everyone wanting to be first with the latest and to AI companies aggressively promoting their best models.

So the technology we use for simple office tasks is like using Artemis II to go out to buy milk. It works only as long as someone else pays for the fuel. The day the bill comes due for users, the milk becomes extremely expensive.

But by all means, replace staff with AI. Just remember that it’s not necessarily a sign of strategic triumph, but an attempt to reap the benefits of a subsidized golden age before the bill comes.

––

Translated from the original Swedish, this column is taken from CS Veckobrev, a personal newsletter featuring reading recommendations, link tips, and analysis sent directly from Editor-in-Chief Marcus Jerräng’s desk. Would you also like to receive the newsletter on Fridays? Sign up for a free subscription here.