Tokenomics: Why making AI pay is tricky

Ask ChatGPT to plan your holiday and you're getting a genuinely good deal, subsidized by hundreds of billions of dollars that Microsoft, Google and Anthropic have poured into building these models. But somebody has to pay that money back eventually, and that's where things get messy. The bill comes down to tokens, the small chunks of text an LLM breaks a request into before spitting out an answer. The trouble is that the same question can produce wildly different numbers of tokens depending on phrasing, the model used, or just the roll of the dice, since these systems aren't deterministic in the way a calculator is.

That unpredictability turns into a real financial headache once businesses start stacking multiple AI agents together to handle tasks automatically. Goldman Sachs reckons token consumption by businesses will jump 24-fold between 2026 and 2030, hitting 120 quadrillion tokens a month. Individual tokens have gotten cheaper, sure, but volume is exploding fast enough to erase those savings. Uber reportedly blew through an entire year's coding-token budget in just a few months this year. Microsoft, of all companies, has had to pull back its own engineers' use of third-party coding tools.

Simon Gooch of Saviynt put it bluntly: locking anyone into a fixed price for the next one, two or three years makes no sense right now, because nobody actually knows what usage will look like by then. Some smaller firms are dodging the problem entirely by running flat-fee personal accounts instead of enterprise contracts, essentially flying under the radar, according to smartR AI founder Oliver King-Smith. He doesn't think that lasts. Once the big AI vendors face shareholder pressure to turn a profit, he expects a crackdown on exactly that kind of workaround.

And the math gets weirder the deeper you go. Companies budgeting for AI often only plan for the token costs tied directly to writing code, forgetting the extra spend needed for testing, security checks or safety guardrails, says LSE's Will Venters. Adding another AI agent takes one click; adding another human employee takes a hiring process. That asymmetry makes it dangerously easy for token costs to creep up unnoticed until a company gets its bill or simply runs dry.

Down the chain, firms selling AI-powered products are stuck figuring out how to pass these shifting costs to their own customers. Sumo Logic's Bill Peterson admits nobody has really cracked it yet, joking that internal pricing conversations have gotten pretty entertaining. Flat price hikes, pay-per-result models, bundled incident pricing, all are on the table. But whatever pricing scheme a company lands on could be wrecked overnight if the big LLM providers shift their own rates again, something Peterson says happens every couple of months. Customers hate that kind of instability. Nobody builds a budget around a number that keeps moving.