Wall Street Needs OpenAI Token Crash Course Ahead of Human IPO

When OpenAI and Anthropic finally make it past the secretive IPO filing stage and publish their prospectuses, investors will be inundated with references to a term that is still unfamiliar to many on Wall Street: tokens.

These are the new currencies of artificial intelligence. That’s how major modeling companies get paid. Today’s developers are using tokens to build apps. However, converting tokens into dollars, a currency that investors understand, is complex and requires investors to act quickly.

“This is a work in progress for all of us as we navigate this new territory,” said Gil Luria, technology analyst at DA Davidson. He covers companies such as: Amazon, microsoft and alphabetreferences to tokens are sprinkled in statements made during financial results announcements over the past year.

OpenAI announced Monday that it confidentially filed a prospectus with the SEC, a week after Anthropic made a similar filing. Together with SpaceX’s IPO scheduled for this week, these could be the largest public offerings ever, making understanding tokens important. Just as the advent of cloud computing more than 20 years ago began the transition from software licenses to new business models of subscriptions, the AI ​​era is fundamentally changing the way companies pay and get compensated for what has quickly become a critical technology.

Every time a user of ChatGPT, Claude, or any other AI service builds a spreadsheet or generates an image or vibe code for your app through a text prompt, they need a certain number of tokens to complete the task. A token is about three-quarters of a word.

Model developers sell subscriptions that include token quotas and provide users with access to the API. This includes charging customers according to their token usage.

OpenAI makes some of its coding models available for free through its Codex app, which can handle more complex tasks, while Anthropic offers the Claude Code tool for paid subscribers. Both companies sell individual subscriptions for up to $200 per user per month. If people use up their token allocation, they can pay an additional fee for additional tokens.

OpenAI and Anthropic list prices for their models on their websites. For the most powerful model, GPT-5.5, OpenAI charges $5 for every 1 million input tokens that are user queries and $30 for every 1 million output tokens (model responses). Anthropic is priced similar to the Claude Opus 4.8 model, but costs $25 for an output of 1 million tokens.

It’s a complicated equation for beginners. Fortunately for Wall Street veterans looking to get a grip on the new digital economy, chipmakers are cerebrum and rocket operator space xThe owner of xAI published an extensive commentary on the token in his recent IPO filing. Cerebras’ prospectus mentions the token 23 times, while SpaceX’s filing has 62 references, including a glossary.

In defining the term, SpaceX said that a token “refers to the basic unit of text or image processed and produced by an AI model that is used for AI measurement.”

Later in the filing, the company said that tokens “represent the fundamental units of data consumed and produced by modern AI models, corresponding to, for example, words, images, sounds, and other modalities. Tokens serve as atomic units for models to read, reason with, and produce output.”

SpaceX, which plans to list on the Nasdaq market on Friday, has offered some discussion of the token, but the token is not that important to the company’s finances. About 70% of SpaceX’s first-quarter revenue came from its Starlink satellite internet business, and the company’s space division accounted for another 13%. While AI accounts for the remaining 17% of revenue, the AI ​​sector is a drain on capital and accounts for the majority of total capital expenditures.

SpaceX’s AI division includes OpenAI, Anthropic, and google. Grok 4.3 and Grok Build 0.1 are not the most widely used models at OpenRouter, a startup that provides developers with access to hundreds of models.

Google’s position in the market makes tokens an increasingly important metric for search companies. Consumers and employees use them in Gemini products, and developers use them through Google’s cloud infrastructure. Amazon Web services and microsoft Azure.

During the company’s quarterly earnings call in April, CEO Sundar Pichai said Google’s model “currently processes more than 16 billion tokens per minute through direct use of our APIs by our customers, up from 10 billion last quarter.” Pichai added that over the past year, 330 cloud clients have “processed over 1 trillion tokens” and “35 have reached the 10 trillion token milestone.”

Scott Breitenother, co-founder and CEO of AI startup Kilo Code, said what’s missing from the conversation around tokens is what kind of return on investment companies are getting from using them.

“Token volume is a useful directional indicator, but ultimately what companies care about is impact and ROI,” Breitenother said.

Although Google doesn’t directly tie token usage to revenue, its cloud business is in jeopardy. The segment’s sales rose 63% year over year to $20 billion in the first quarter, accelerating from 28% growth in the same quarter in 2025. Meanwhile, operating profit more than doubled to $6.6 billion.

This result is instructive as it shows that the demand for AI services is rapidly increasing. However, for OpenAI and Anthropic, it doesn’t make much sense because these companies don’t sell cloud infrastructure. Rather, they pay a lot of money to host the AI ​​models they provide to their customers.

Cerebras’ IPO filing could be helpful to investors. competing chip manufacturers Nvidia The company, which sits on the fringes of the AI ​​industry, went public in May, giving the market its first real taste of a pure AI company.

Cerebras is a hardware company that is in the token generation business. The company competes with other chipmakers to build cutting-edge systems for computing. The company says the current chip is 58 times larger than the Nvidia processor, called the B200, and has 19 times more transistors and 250 times more memory.

Cerebras explains the role of the token in its prospectus as follows:

Cerebras’ job is to build hardware so good that companies like OpenAI and Anthropic are willing to use it for unusual purposes. In January, Cerebras signed a deal to provide more than $10 billion in computing to OpenAI through 2028. Only users with high-end monthly subscriptions can try the Codex Spark research preview, which utilizes Cerebras chips.

For OpenAI and Anthropic’s business model to work, companies must be able to generate enough revenue from the use of their tokens to pay for all the hardware and systems provided by Nvidia, Cerebras, and cloud providers, and have enough money left over to pay for everything else. At the moment, the calculations still don’t work.

SpaceX said in its prospectus that its business success depends on how effectively it utilizes available computing power. The company says it currently enjoys “token cost advantages” thanks to rapid infrastructure deployment. Unlike OpenAI and Anthropic, SpaceX has its own large data centers, with two operating in and around Memphis, Tennessee.

SpaceX describes its vertically integrated approach as “shovel-to-token” and says it “allows you to train and iterate frontier models at low cost and speed, accelerating development cycles, eliminating external bottlenecks, and rapidly and continuously improving model performance.”

However, the company’s benefits from this strategy are sure to show up in the future. For now, SpaceX is choosing to lease data center capacity in addition to using it for its own models. In May, Anthropic committed to paying SpaceX $1.25 billion a month for three years for capacity at xAI’s Colossus 1 facility. And last week, SpaceX announced it would pay Google $920 million a month until mid-2029 for the use of 110,000 Nvidia GPUs.

SpaceX is renting out its capacity while offering Grok 4.3 at a much lower price (per million input/output tokens) than OpenAI and Anthropic’s flagship products. Regarding Grok, Kilo Code’s Breitenother said, “That doesn’t mean there isn’t demand.”

“However, it does suggest that there are still use cases where organizations prefer other frontier models, even if the costs are higher,” he said.

SpaceX did not respond to a request for comment.

—CNBC’s Lora Kolodny and Ashley Capoot contributed to this report.

clock: AI token or human? A new discussion to restructure corporate budgets