Using AI seems fairly straightforward: you pay a flat fee and prompt to your heart's content. This creates the impression that AI costs next to nothing. The logical conclusion is that AI could replace an arbitrarily large number of human work hours or workers, because it must surely be cheaper.
That illusion is now bursting. Data centers consume land and enormous amounts of energy. The investment bank Morgan Stanley estimates that the global construction of data centers alone could cost around $2.9tn by 2028. This is driven by demand for compute – that is, demand for processing power, which manifests in concrete, tangible form as chips, servers, electricity and data centers that train AI models and handle requests. Here, demand currently far exceeds supply.
It is a gold rush market, but the prices for the tools are rising just as fast as the need to generate profits. The capital required can only be raised on stock markets. Shareholders want to see dividends and share price gains. For this reason alone, AI providers cannot permanently hide their costs behind flat-rate subscriptions.
The Hidden Price Tag
The technology is expensive – the training, IT security, quality control and much else besides. Behind the scenes lie enormous fixed costs. Graphics processing units (GPUs), servers, data centers, cooling, electricity, networks, storage and financing all need to be paid for. Morgan Stanley therefore describes the AI expansion less as normal software scaling and more as an industrial infrastructure boom. Amid the founder frenzy, AI is experiencing the same effect that the early internet once went through: it gives the impression of costing almost nothing, yet proves, when all is said and done, quite expensive.
The internet taught us that nothing is truly free: what costs nothing in fees costs us our data, which is worth real money to those who collect it. The dot-com bubble burst in the early 2000s for the same reason – viable monetization models were nowhere to be found.
Whether the AI boom ends in a bubble burst remains an open question. Internal cost calculations from providers, now coming to light, reveal that the actual costs of AI diverge sharply from what pricing models suggest. Large chat models convert every request into tokens – words, syllables, commands, code or numbers – and providers like OpenAI and Anthropic use these to track their costs with great precision.
The Flat Rate Was Always an Illusion
The core problem lies with the flat rate. To the user, a brief chat and a multi-hour coding session look much the same – both covered by the same subscription. For the provider, they are entirely different propositions. Simple chat questions and autonomous coding sessions, in which the AI independently writes entire program complexes, generate vastly different costs, yet have until now been priced similarly. That cannot go on indefinitely.
From 1 June 2026, Copilot switched to a system of "AI Credits" billed according to token consumption, model and usage type. The pricing is granular: a single token on GitHub Copilot costs $0.01 per AI Credit, with rates varying further depending on the model used.
Billing is calculated per one million tokens rather than per individual token. The price list illustrates the stakes: GPT-5.5 costs around 25 times as much as GPT-5.4nano. OpenAI, the provider of ChatGPT, puts the broader picture in sharp relief – media reports point to an adjusted operating margin of minus 122%, meaning the company loses $1.22 for every dollar of revenue. High usage, high infrastructure costs and low end-user prices are a combination the stock market will not tolerate indefinitely.
A Really Stupid Way
The shift in provider pricing is already being felt in offices. Where companies once pursued tokenmaxxing – maximizing token consumption as a proxy for productive AI adoption – a correction is setting in. High consumption, it turns out, is not the same as high output.
Numerous managers are now openly criticizing tokenmaxxing as a "really stupid way" to measure AI value. Fast Company reports that one unnamed company spent $500m on Claude usage in a single month, according to an AI consultant. Anecdotal as that figure may be, it points to a reckoning that providers and users alike can no longer avoid.
The industry is moving from all-you-can-prompt to pay-for-what-you-compute, and the economics of AI applications are shifting with it. The question that now matters is not how many tokens a workflow consumes, but what a usable result actually costs.
That calculation is more involved than it once appeared. Tokens, model choice, agent runs, failed attempts, human review and rework all enter the equation. For simple tasks, AI remains extremely affordable. For complex or poorly managed workflows, an experienced human can ultimately be cheaper. Businesses are beginning to grasp this.
The cooling of AI cost-saving hype may yet spare some companies a painful reckoning - provided they have not already dismantled the human expertise they will need to fall back on. The pipe dreams of certain cost accountants about frictionless, human-free workflows have met the oldest of obstacles: the numbers do not add up.