Pre-IPO Investors Press Anthropic on Core Metrics Including AI Token Economics
nashnova research
Anthropic plans to file its IPO prospectus after Labor Day, but large investors are already demanding disclosures far beyond regulatory minimums — token economics, revenue per gigawatt, and net revenue retention — metrics that will shape how the market prices the entire AI sector.
What numbers are investors actually asking for?
Three metric clusters top the list: AI token economics (revenue and cost per token processed), revenue per gigawatt of compute, and net revenue retention (whether existing customers spend more over time).
This means → investors want granularity down to how much each dollar of compute earns back, not just top-line revenue and gross margin.
In plain terms = traditional financials say "how much the company made"; investors now want to see unit cost versus unit return for the AI business itself.
Yet there is no guarantee Anthropic will disclose any of these — legally, it has no obligation to do so.
Growth is surging — so why the anxiety?
Anthropic's revenue growth is exceptionally strong, and by some measures it has reached operating profitability. But investors worry about two things: enterprise clients may shift to open-source models to cut costs, and the profit-margin impact of total compute spending plus partnerships with SpaceX, Google and others remains opaque.
This means → high growth alone is not enough; investors need proof the growth is not subsidized by cash burn.
They also want to see the revenue split between new and older models — whether massive R&D spending is translating into real returns, the key test for whether "ever-more-expensive models" are worth it.
Why does token economics matter more than gross margin?
Investors want to know Anthropic's total token volume, net revenue per million tokens after discounts, and the cost to serve those tokens.
In plain terms = gross margin tells you "does running the model make money"; token economics breaks open the profit-and-loss behind every single AI response.
This reflects a fundamental gap between AI companies and traditional software — traditional software sells licenses at near-fixed cost; an AI model burns compute every time it answers, making unit economics the lifeline.
Why is net revenue retention so hard to calculate?
Anthropic sells Claude subscriptions, but a large share of revenue comes from API calls — usage that generates no traditional renewal record.
Developers switch models for different tasks, so API volume fluctuates month to month, making a stable retention baseline elusive.
This means → even if Anthropic wanted to disclose retention, the math itself is the problem — it is not hiding the number; the number is genuinely hard to compute.
Why does Anthropic's choice matter for the whole industry?
According to a person familiar with the matter, OpenAI's IPO advisory team is already debating how many similar metrics to disclose; OpenAI's listing is expected next year.
This means → what Anthropic chooses to reveal — or withhold — will set the transparency benchmark for AI-sector IPOs.
In plain terms = the first company to hand in the exam paper sets the difficulty level — the more Anthropic discloses, the harder it becomes for those that follow to hold back.
市场有风险，内容仅供研究参考，不构成投资建议。