We all know those people in the workplace who brag about using AI for every aspect of their job – replying to emails, creating slide decks, even writing their op-eds.
That enthusiasm is reflected in the numbers.
Google said recently that it now processes over 3.2 quadrillion tokens a month. These are the basic units of text that AI models read and generate. This is seven times more than the same period a year ago.
After a period of irrational exuberance, company executives are discovering that AI ‘tokenmaxxing’, where employees try to use as much AI as possible, is proving far more expensive than anticipated.
Every time we prompt a large language model, resources are consumed and costs are incurred.
Earlier this month, a United Nations report estimated that by 2030, AI could consume more electricity than Australia uses annually, more water than the 1.3 billion people of Sub-Saharan Africa, and produce more emissions than the United Kingdom.
But it is the financial costs, not the environmental ones, that are currently concentrating executive minds.
Several large technology firms, including Microsoft and Uber, have already begun restricting internal AI usage or warning that costs are becoming harder to justify as budgets are exhausted faster than expected.
We recently learned that employees at an unnamed company burned through $US500 million on Claude Code tokens in a single month.
In a textbook example of Goodhart’s law, Meta took down its internal AI usage leaderboards after discovering employees were using AI frivolously, even to check the weather, just to climb the rankings.
The majority of AI usage was never really about output; it was about signalling.
Executives boasted about adoption rates, employees rushed to demonstrate enthusiasm, and investors rewarded any company that could claim to be ‘AI first’. In that environment, few people had an incentive to question whether the productivity gains justified the cost.
Many employees are using AI to automate low-value tasks they dislike rather than work that meaningfully drives productivity.
But now sticker shock is forcing CEOs to evaluate AI the same way they evaluate any other operating expense.
Ironically, instead of cutting jobs because workers are too expensive, some firms may be cutting jobs to pay their AI bills.
Most of what the average user does with a large language model is basic querying. This doesn’t require the Rolls-Royce of AI models.
The differentiation that once justified premium pricing is quickly eroding, meaning companies will substitute away from the ‘brand names’ toward cheaper open-source alternatives.
This is the reason why AI is destined to become a commodity.
A key feature of a commodity is its fungibility, in other words, its capacity to be substituted. Given that all the major AI models are trained on broadly similar internet-scale datasets, the outputs are similar, or similar enough, that a cheaper generic model will do.
It echoes the trajectory of electric vehicles, where early technological differentiation gave way to scale-driven price competition and margin compression.
Going forward, enterprise AI adoption is likely to be constrained less by model capability than by price and input scarcity, as with any other commodity market.
The challenge for frontier AI labs is to differentiate themselves through proprietary datasets, unique capabilities or defensible product moats so that customers actively seek them out rather than simply opting for the cheapest model.
The commodification of AI will prompt a race to the bottom and erode the profit margins of companies that are, in many cases, already losing billions of dollars.
That’s before considering that many of those valuations assume a wave of new power and data centre projects that are facing delays and cancellations.
For investors, that raises an uncomfortable question: If AI becomes a commodity, driven by price, what if valuations may be based on subscription premiums that the market won’t pay?
