AI Agents Devour Computing Power as Token Pricing Chaos Plagues Enterprises and Wall Street
Claire Weston
AI coding agents consume over 1,000 times the tokens a human does for the same task, corporate bills are spiraling, and Wall Street can't use token data to track AI monetization — cloud GPU rental prices are emerging as the real signal.
What is a token, and why is it suddenly a problem?
A token — the smallest unit AI models use to read and generate text — is not new. But two forces have made token consumption explode: reasoning models and AI agents.
Reasoning models no longer answer instantly. They run multi-step internal dialogues first. In a *Barron's* test, a single question triggered 32 seconds of internal deliberation, multiple web searches, and dozens of articles read — the visible output was roughly a thousand tokens, but the hidden consumption was far larger.
This means → the same prompt can now burn many times more tokens on a new model than on an older one, and the user never sees the hidden cost on screen.
Why do AI agents burn through tokens so fast?
AI agents — programs that take a simple instruction and autonomously execute a chain of complex operations — are the core driver of token consumption. A study published in April by Google, Microsoft, and several top universities found that coding agents use over 1,000 times the tokens a human needs for the same task.
More tokens do not guarantee better results. Failed tasks must be re-run, stacking costs further.
In plain terms = an agent is like a tireless but unrestrained intern — it researches, writes code, and retries on its own, burning tokens at every step, with no guarantee of getting it right.
How did corporate bills spiral out of control?
Consumer ChatGPT and Claude both offer $20-per-month unlimited plans. Enterprise use is billed per token, and top-tier ChatGPT models range from $0.50 to $30 per million tokens — a 60× spread.
Anthropic recently changed how some products count tokens. The same text now consumes 30% to 40% more tokens, while older models still use the original method. This means → switching models can inflate a bill by nearly 40% overnight, making budgeting virtually impossible.
Some companies built employee AI-usage dashboards early this year — then immediately received enormous bills and were forced to adopt leaner strategies. Uber president Andrew Macdonald said in a May podcast: "How many shelved projects have restarted because of productivity gains? That chain doesn't exist yet."
Why is Wall Street struggling too?
Wall Street hoped token pricing would serve as a proxy for tracking AI demand. But token consumption swings wildly by task and model — the April study found that neither humans nor AI models can reliably predict how many tokens a task will use.
This means → token data is neither stable nor comparable, making it nearly useless for answering the central question: "Is AI actually making money?"
Companies seeking to cut costs are shifting to cheaper models — including ultra-low-cost Chinese alternatives — putting some pressure on OpenAI's and Anthropic's premium pricing, though the market has not yet hit the tipping point that would force deep cuts.
So what metric should we watch instead?
According to *Barron's*, the more effective way to track AI supply-and-demand right now is cloud-provider pricing. AWS raised Nvidia GPU rental rates this week — its second increase this year. Microsoft Azure and Google Cloud pricing trends deserve equal attention.
In plain terms = token prices are too chaotic to reveal real demand. But GPU rents are rising, signaling that underlying compute remains undersupplied — where the money flows is a more reliable gauge than who talks up the AI boom.
Continued agent expansion and accelerating token demand form the core rationale behind the hundreds of billions of dollars still pouring into AI data-center construction — whether that rationale converts into quantifiable profit is the key test for sustaining AI valuations.
Content is for reference only, not financial advice.