Something strange is happening on the cutting edge of Artificial Intelligence.

A central premise of the AI revolution was that AI would reduce costs. Either by helping companies do more with the same number of people or by replacing people entirely.

But something is happening that is up-ending that narrative. AI is no longer looking like cheap labour. In fact, for certain tasks, it may now be costing more than the humans it was meant to replace.

Companies are pulling back on AI spend

AI spending is increasing across the economy. And for some companies it is getting too much.

A recent survey of American companies by CloudZero found that 45 per cent were now spending more than $100,000 a month on AI. That is up from 20 per cent last year.

The most recent stock market reporting season saw several tech companies – including Meta, Pinterest and Spotify – explicitly said that rising AI costs were eating into their profit margins.

As a result, many companies are pulling back on their AI spending.

Uber’s COO Andrew Macdonald admitted it is getting harder to justify AI costs internally, as the ride-sharing and meal delivery company burned through their entire year’s AI budget by April.

Commonwealth Bank CEO Matt Comyn recently acknowledged Australia’s largest bank increasingly views AI as a “scarce resource”.

And Microsoft, a company right on the forefront of AI, is cancelling Claude Code licenses across multiple divisions. The company that is spending $80 billion this year on AI data centres has come out and said that AI use by employees is getting too expensive.

The cost model has dramatically changed

A big reason for these rising costs has been the shift towards AI Agents. Agents use far more AI tokens than your traditional chatbot and as a result AI companies have changed the way they charge.

For large customers, the model has moved from flat-rate monthly subscriptions to usage-based billing based on token consumption. GitHub recently made such a move, and one developer reported their monthly costs rose from roughly €67 in April to around €966 under the new model.

The pace of cost increases doesn’t appear to be slowing down. Goldman Sachs analysts recently forecast a 22-fold increase in token consumption by 2030 and a 55-fold increase by 2040.

Scott Farquhar, the co-founder of Australian software giant Atlassian, recently explained the scale of the cost increases, “Two years ago, there was no line item for tokens or large language models in a CFO’s budget, and now there can be $10 million or $100 million a year being spent on something that didn’t exist previously.”

As these costs rise, companies are being forced to question their AI spending.

Spencer Rascoff, the CEO of Match Group – the company behind dating platforms like Tinder, Hinge and Match.com – said that his company was spending between $5 – $10 million a year on AI.

And yet, when he was asked about it, he was honest.

“I think we’re benefiting from it, but it’s hard to feel it.”

AI isn’t always cheaper than humans

As AI costs rise, the promise of saving money on wages may not be universal.

Recent work from Goldman Sachs looked at some of the most common uses for AI today – coding, customer support and basic data entry – and compared human wages with the cost of AI tokens required to do a comparable day’s work.

Turns out, in some instances, humans are cheaper.

According to Goldman’s analysis, a human customer service rep, for example, can be cheaper than the equivalent AI token spend to answer a full day of queries.

Don’t expect companies to give up on AI

The promise and potential of AI is so large, that we shouldn’t expect companies to give up on AI. Instead, the solution may be for companies to look overseas. We may see a shift towards Chinese AI models.

Models coming out of China are 10-30x cheaper than US models.

For example, Chinese AI company DeepSeek can charge as little as $0.14 per million input tokens.

Claude Opus, the latest model from US company Anthropic, charges $5 per million input tokens.

We’re already seeing the shift. One survey found that use of Chinese AI models by developers has increased from less than 1% in 2024 to more than 60% in May 2026.

And 80 per cent of AI start-ups based in the US already report using Chinese open-source AI models. Many of them use cheaper Chinese models alongside more robust American models.

This is the AI future we should expect to see. More choice at different price points will give companies more flexibility on what AI model is best suited to a particular task.

AI is not going away. But the free-for-all phase may be. The next stage of the AI boom will not be defined by who uses the most AI, but by who can work out where it is actually worth the cost.

DISCLAIMER: Information and opinions provided in this column are general in nature and have been prepared for educational purposes only. Always seek personal financial advice tailored to your specific needs before making financial and investment decisions.