On September 28 (local time), foreign media got hold of an undisclosed IPO prospectus from Anthropic. This document represents a preliminary draft of the S-1 registration statement that the company had confidentially submitted earlier, with pertinent data subject to refinement following the U.S. Securities and Exchange Commission's scrutiny. Anthropic aims for a listing valuation surpassing $2 trillion, more than twice its self-assessed valuation of $1 trillion in May 2025. As the first avant-garde large model lab in nearly four years to openly share its comprehensive financial records with public investors, the prospectus unveils several crucial insights:
Firstly, of the approximately $42 billion net loss reported for 2025, $34 billion arises from accounting costs associated with the escalating valuation of financing instruments, rather than tangible operational outlays. The company's revenue skyrocketed over tenfold year-on-year that year, with operating losses hovering around $8.06 billion, hinting at the emergence of economies of scale.
Secondly, full-year revenue for 2025 hit $4.59 billion, with explosive growth witnessed in 2026. By late July, annualized revenue surpassed $65 billion, although this projection is based on extrapolating a single month's earnings. Nearly a quarter of the revenue originated from two undisclosed major clients, and the majority of customers lacked long-term contracts, casting doubt on the sustainability of growth.
Thirdly, Anthropic's expenditure on computing infrastructure reached $7.33 billion in 2025. Investors Amazon and Google also double as its primary cloud service providers, with investment funds cycling back to these tech behemoths in the form of cloud service fees. This capital loop, prevalent in the AI sector, has turned the proportion of genuine end-user demand into a focal point for market examination.
Fourthly, the company has instituted a unique governance framework, with seven co-founders wielding control over 50.1% of total voting rights through Class F shares. The founders have committed to donating 80% of their personal stakes to charity and have explicitly notified investors that corporate decisions might prioritize objectives like safety over financial gains, potentially diluting the rights of ordinary investors.
Fifthly, the company has opted against developing commercial products for image or video generation, choosing instead to channel computing power into research and safety priorities. This decision mirrors both a business strategy amidst computing limitations and a presentation of values.
Sixthly, the 80-page risk factors section of the prospectus explicitly cautions that AI models could pose existential threats. This underscores the company's safety stance but also serves as standard legal risk disclosure, highlighting the industry's broader predicament of reconciling calls to "slow down" with demands to "accelerate iteration."
Seventhly, the reported $518 billion in computing expenditure represents the cumulative value of multi-year procurement contracts. Currently, Anthropic possesses only $20.28 billion in cash reserves, rendering the IPO largely a strategy to finance pre-signed computing contracts. Anthropic's listing is now anticipated to be postponed until after the November U.S. midterm elections, aligning with rival OpenAI's confidential IPO filing and subsequent halt. Their divergent interpretations of the "safety" concept have transformed this technical ethics issue into an element of capital narratives. Anthropic's IPO will set Wall Street's valuation benchmark for leading AI firms, fundamentally testing whether capital markets can embrace a company that places "what not to do" on par with "what to do."