Sixty-two times. That is how often the word "token" turns up in the prospectus SpaceX filed ahead of its planned Nasdaq debut this Friday, glossary entry and all. Cerebras, the chipmaker, managed twenty-three references in its own filing. The count itself is beside the point. What matters is that a term most portfolio managers couldn't have defined two years ago is now load-bearing in the documents that will decide how some of the largest stock offerings in history get priced.
OpenAI confirmed on Monday that it had filed its prospectus with the Securities and Exchange Commission on a confidential basis, about a week behind Anthropic, which did the same. Both companies remain in the quiet phase, so the public hasn't seen the figures. When those documents do surface, investors will hit the word again and again, and many of them will be forced to work out, more or less on the fly, what it actually means for the bottom line.
What a token actually is
Strip away the jargon and a token is simply the unit of measurement for how much work an AI model does. One token equals roughly three-quarters of a word. Ask ChatGPT to draft a spreadsheet, generate an image, or stitch together a small app from plain-English instructions, and the system burns through a measurable quantity of tokens to produce the result. SpaceX, in its filing, called tokens the basic building blocks of text or images a model processes, the smallest pieces through which it reads, reasons and produces output. Rival companies have offered near-identical definitions, mostly because there isn't a competing one to offer.
The model developers get paid by those units. A subscription comes bundled with a token quota; blow past it and you pay for more. Developers building on top of these systems are billed through the companies' programming interfaces, again metered by token. OpenAI and Anthropic both publish their rates in the open. For OpenAI's most capable system, GPT-5.5, the company lists $5 per million tokens of input, meaning the user's request, and $30 per million tokens of output, the model's answer. Anthropic prices its Claude Opus 4.8 in the same neighborhood, at $25 per million tokens on the output side. Consumer plans from both firms top out around $200 a month per user.
The complication, and it is a genuine one, is turning all of that into the metric Wall Street actually trades on: dollars of recurring revenue with predictable margins. Gil Luria, a technology analyst at D.A. Davidson who covers Amazon, Microsoft and Alphabet, called the whole exercise a work in progress for everyone trying to make sense of the new terrain. He would know. The companies he follows have been slipping token references into earnings calls for the better part of a year, and the analyst community is still calibrating how seriously to take them.
Why the comparison to the cloud holds up
There is a useful historical rhyme here, and it rewards a moment's pause. Two decades ago, the rise of cloud computing dragged the software business away from one-time licenses and toward subscriptions billed by the month. That shift took years for investors to price correctly. Recurring revenue eventually became the gold standard, but in the early going plenty of analysts undervalued the companies that pioneered it, because the old yardsticks for measuring software simply didn't fit. The token economy looks like a comparable inflection. Usage-based billing, metered by consumption rather than by seat or license, is fast becoming the default way the most important technology of the moment gets bought and sold.
The catch is that consumption swings in a way subscriptions do not. A customer who burns through a token allotment one month might barely touch it the next. Heavy coding users can run up enormous bills; casual chatbot users barely register. For a company trying to forecast revenue, and for an analyst trying to model it, that variability is far harder to underwrite than a flat monthly fee. The metric that captures growth, total tokens processed, doesn't convert cleanly into profit, because the cost of generating those tokens, mostly the price of running vast fleets of chips, can swallow the revenue whole.
SpaceX is a preview, not the main event
It would be a mistake to read too much into SpaceX's filing as a guide to the pure AI businesses, and the company's own numbers explain why. About 70 percent of its first-quarter revenue came from Starlink, the satellite internet operation. The launch and space division added roughly 13 percent. The artificial intelligence unit, home to xAI and its Grok models, accounted for the remaining 17 percent, and that slice is losing money while consuming the lion's share of the company's capital spending. So when SpaceX writes about tokens across sixty-two pages, it is describing a business line that is, financially, the tail rather than the dog.
It is also a relatively small player in the market it is writing about. On OpenRouter, the startup that routes developers to hundreds of available systems, neither of xAI's flagship releases of the moment cracks the list of most-used models. OpenAI, Anthropic and Google dominate that traffic. Which is exactly why the SpaceX document is instructive rather than definitive: it hands investors the vocabulary they will need, the way a phrasebook prepares you for a country you haven't visited, without telling you much about the destination itself.
Google, by contrast, has more riding on the metric than almost anyone. Tokens flow through its consumer Gemini products as well as its cloud arm, where its principal rivals for enterprise spending are Amazon's and Microsoft's hosting platforms. On the company's April earnings call, chief executive Sundar Pichai pointed to the volume of tokens its models were processing as a marker of momentum. When the largest companies in technology start quoting a usage figure to prove they are growing, the analysts covering them have little choice but to learn the language.
What to watch when the filings go public
The real test arrives once OpenAI's and Anthropic's prospectuses emerge from behind the confidential curtain. Together with SpaceX, these could rank among the biggest offerings ever recorded, which means the token economy is about to face its first proper stress test in front of public-market investors, the sort who price things by the quarter and punish surprises. The hard questions are no longer about what a token is. They are about whether token revenue carries durable margins, whether usage growth is sticky or fickle, and whether the cost of producing all those tokens leaves anything behind.
Nobody has fully convincing answers yet, the companies included. And the timing, with three landmark listings stacking up in the same stretch, is hard to ignore. Investors will get their crash course whether they feel ready or not. The grade comes later, in the price.