Seeing "net losses in the first nine months exceeded US$104 billion", the first reaction may be that this company is burning money like crazy.
It's not that exaggerated when you take it apart. Of the 104 billion, 94.3 billion is the change in the fair value of liabilities, which is a book loss in accounting terms, and the money was not actually spent. Operating loss was $1.26 billion. For a company whose ARR reached $81 billion as of the end of October, this loss is considered restrained.
What really weighs on the company is another set of numbers: $900.9 billion in long-term computing power commitments, of which $70.7 billion expires within a year.
Comparing ARR, the computing power commitment is more than 11 times the annualized revenue. The 70.7 billion to be paid in one year alone is close to 90% of the annual revenue.
Amodei called computing power "the deepest business problem" in the prospectus. The strategy he gave was to buy less at first, and then add more when demand really increases. But GPUs and data centers are not available just because you want to. Micron just said that memory will be tighter in 2027 than this year. If you buy too little, you won’t be able to meet the demand. If you buy too much, you’ll end up with 900.9 billion.
This also explains why the private placement subscription list grows like that. 26 billion US dollars, contributed by Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC. Google and Nvidia each contributed 10 billion, all of which sell chips or cloud. A considerable part of the money they invest will be returned to their accounts in the form of computing power purchases.
What's even more awkward is that there are also opponents on the list. On the eve of its launch, Google lowered the input price of Gemini 4 Carbon to US$2 per million tokens, and the Claude Fable 5.2 in the same range was US$8; Microsoft said that MAI Code 2 achieves 96% of the coding capabilities of Opus 5.5 at a price of 5%.
Money bags and price wars entangled on the same list are the most unique situations Anthropic is in.
The other side is the employees. The prospectus showed that as of September 30, option holders had unrealized paper gains of approximately $119 billion, of which approximately $73 billion had already vested. The third-party platform estimates based on the issue price of US$83 that the shareholding value of about 600 non-founding employees exceeds US$100 million, about 2,800 people exceed US$10 million, and most of the more than 5,000 employees exceed one million US dollars. These are model estimates, not individual shareholdings disclosed by the company, and exercise costs and taxes will be deducted from the final results.
Once the lock-in period is over, how many of these people will leave to start their own businesses is Anthropic’s next real management issue. In the past, we competed with OpenAI and Google to retain people through options, but now the options have been realized. Claude's product rhythm will not change in the short term, and we will have to wait until next year to see whether the team is stable.