The Sequence Opinion #879: When Tokens Become Balance Sheet Items
Substack Jesus Rodriguez
AI companies are now tracking 'tokens' — the tiny text chunks AI models process — like real accounting line items on balance sheets. That's a big shift: the basic unit of AI cost has quietly become as important as revenue or headcount in how businesses plan.
Based on reporting by Substack, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Something changed in how AI companies think about money, and it happened fast enough that most people missed it. Tokens — the small chunks of text that language models chew through — have become the base unit of cost in the AI economy. Not GPUs, not parameters, not even the models themselves. Tokens.
That sounds like a technical footnote until you realize what it means for a balance sheet. Big companies are now forecasting token consumption the way they'd forecast electricity bills or cloud storage costs. Anthropic, OpenAI, and their enterprise customers are treating tokens as a line item with real financial weight, something finance teams model, budget against, and try to optimize down. A few years ago this would have sounded absurd. Now it's just Tuesday for a CFO at a mid-size SaaS company running AI features into production.
The interesting question isn't whether tokens matter financially — clearly they do — it's whether the tools for managing that spend exist yet. They mostly don't. Companies are eyeballing usage dashboards, squinting at API bills, and hoping their engineering teams don't accidentally burn six figures running an unoptimized prompt loop overnight. That's not an accounting system, that's triage. The piece from The Sequence argues this gap is about to get filled, and I think that's right. Somebody is going to build the equivalent of an ERP system for token economics, tools that forecast, allocate, and audit token spend across departments the same way SAP or NetSuite handle traditional operating expenses.
What makes tokens weird as a unit of measurement is that they're simultaneously a cost and a proxy for cognitive work. A token isn't like a kilowatt-hour, which is a stable physical quantity. It's tied to model architecture, pricing tiers that shift with every model release, and business logic that changes depending on which vendor you're using that quarter. Measuring the token economy accurately means building infrastructure that can track a moving target, and none of the existing financial tooling was designed for that. Excel wasn't built with tokenization in mind, and neither was your average FP&A team.
So we're at the point where the unit of measurement has arrived well before the instruments to measure it properly. That's usually how new economies start — messy, ad hoc, running on spreadsheets duct-taped together — right before someone builds the boring, unglamorous software that makes the whole thing legible. Whoever builds that software for tokens is going to have a very good decade.
My take — AI-written commentary, not fact-checked reporting
I've watched enough infrastructure booms to know the pattern: nobody gets excited about the boring middleware until the spending gets so chaotic that ignoring it becomes a liability, and token spend is heading there fast. My money's on this becoming a genuine software category, not a feature bolted onto existing cloud cost tools, because the unit economics of tokens are too weird and too volatile for generic FinOps platforms to handle well. Somebody's building the Datadog of token spend right now in a garage, and in three years every enterprise will wonder how it ever functioned without one.
Read more about this at: Substack