Wall Street's Token Literacy Gap Could Distort the Largest AI IPOs Ever Filed

OpenAI filed a confidential S-1 registration with the SEC on June 8, one week after Anthropic did the same. SpaceX begins trading on Nasdaq this Friday. Three of the most anticipated public offerings in technology history are converging within a single quarter, and according to CNBC’s reporting, the investors who will price them don’t understand how the companies make money.

The problem is tokens.

What Investors Don’t Know

A token is roughly three-quarters of a word. Every time a user builds a spreadsheet, generates an image, or writes code through an AI model, the system consumes a measurable quantity of tokens to complete the task. Model developers sell access through subscriptions with token quotas and through API billing by token usage.

This is the atomic unit of AI commerce, and it governs everything from OpenAI’s revenue model to Google’s cloud growth metrics. But as D.A. Davidson analyst Gil Luria told CNBC, “It’s a work in progress for all of us navigating this new terrain.” Luria covers Amazon, Microsoft, and Alphabet, companies that have already begun sprinkling token references into their earnings calls.

The pricing is not intuitive. OpenAI charges $5 per million input tokens and $30 per million output tokens for GPT-5.5, its most powerful model. Anthropic’s pricing for Claude Opus is similar. Translating those figures into revenue projections requires understanding token consumption patterns, model routing strategies, and the relationship between subscription tiers and overage billing.

SpaceX and Cerebras Offer a Preview

The good news for Wall Street is that two recent filings provide a template. SpaceX defined a token in its prospectus as “the fundamental unit of data consumed and produced by modern AI models, for example corresponding to words, images, audio, or other modalities.” The company described it as “the atomic unit through which models read, reason, and generate output,” according to CNBC.

Cerebras, the chipmaker that went public in May, frames tokens from the hardware side. In its filing, the company argued that “token consumption is growing exponentially” and positioned itself as “extraordinarily well-positioned to win in this market” because its systems generate tokens faster than competitors. Cerebras signed a deal in January to provide more than $10 billion worth of compute to OpenAI through 2028, per CNBC.

SpaceX described its vertically integrated approach as “shovels to tokens,” claiming it can “train and iterate our frontier models at lower cost and higher velocity, accelerating development cycles, eliminating external bottlenecks, and driving rapid, continuous improvements in model performance.”

But SpaceX’s AI division remains a niche player. Roughly 70% of its first-quarter revenue came from Starlink, and its Grok models are not among the most widely used on OpenRouter, according to the CNBC analysis. Anthropic is paying SpaceX $1.25 billion per month for three years for capacity at xAI’s Colossus facility, meaning SpaceX is currently generating more revenue by selling tokens to competitors than by selling its own models.

The Google Benchmark

Google offers the clearest public benchmark for what token-driven growth looks like at scale. CEO Sundar Pichai said on the company’s April earnings call that Google’s models “now process more than 16 billion tokens per minute via direct API use by our customers, up from 10 billion last quarter.” Over the past year, 330 cloud clients “processed over a trillion tokens,” per the CNBC report.

Google Cloud revenue rose 63% year-over-year in Q1 to $20 billion, accelerating from 28% growth in the same quarter of 2025, with operating income more than doubling. The results demonstrate explosive demand for AI services. But as CNBC notes, they don’t translate directly to OpenAI and Anthropic, because those companies don’t sell cloud infrastructure. They pay heavily for it.

The Model Routing Threat

One detail in the CNBC report deserves particular attention from anyone building on or investing in AI platforms: model routing is emerging as a cost-control mechanism that creates hidden margin pressure.

Model routing directs requests to cheaper, smaller models when a task doesn’t require frontier capability. For individual users managing subscription quotas, this is a convenience feature. For enterprises running autonomous agents that make hundreds of API calls per task, it is an existential pricing variable. An agent that routes 80% of its reasoning through a $0.25-per-million-token model instead of a $30-per-million-token model consumes a fraction of the revenue that raw token counts would suggest.

Investors reading S-1 filings that report “tokens served” as a growth metric without understanding model routing will systematically overestimate per-token revenue. Scott Breitenother, co-founder and CEO of AI startup Kilo Code, told CNBC that “token volume is a useful directional metric, but businesses ultimately care about impact and ROI.”

The Agent Pricing Equation

This literacy gap matters most where autonomous agents meet enterprise budgets. Agent-based consumption is fundamentally different from chatbot usage. A single agent task might chain dozens of model calls, consume tool-use tokens, process retrieval context, and generate structured outputs across multiple reasoning steps. The token cost of an agent completing a purchase order, triaging a support ticket, or reviewing a contract is an order of magnitude higher than a human asking ChatGPT a question.

When OpenAI and Anthropic make their prospectuses public, the question for investors is whether reported token volume represents high-margin subscription overages or low-margin bulk API consumption by agent platforms and enterprises with negotiated pricing. Without token economics literacy, the difference between a $300 billion and a $150 billion valuation could come down to a unit of measurement that most Wall Street analysts started hearing about less than a year ago.

For the companies building on top of these models, the takeaway is simpler: the market that will price your infrastructure providers doesn’t yet understand the unit of measurement that determines your operating costs. That gap closes when the S-1s go public. What gets priced in before then is anyone’s guess.