Author：Wall Street CN
Two of Silicon Valley's hottest AI unicorns are racing to IPO this year, but a rare disclosure of financial data reveals the same problem: the astronomical computing costs of training AI models are eroding the profit margins of both companies.
According to financial documents recently obtained by the Wall Street JournalOpenAI projects its computing power spending to reach $121 billion by 2028. Even if revenue nearly doubles by then, its losses will still reach a staggering $85 billion that year.This figure will surpass the loss records of almost all listed companies in history.
Meanwhile, Anthropic's projected spending is far less than OpenAI's, but even its most optimistic forecast reflects the ever-increasing cost of computing power. Furthermore, according to a Bloomberg report on Tuesday, Anthropic's latest annualized revenue (Run Rate) has surpassed $30 billion, a significant jump from $9 billion at the end of 2025.
The financial data of the two companies together outline the real logic of the AI arms race: revenue is growing rapidly, but training costs are also ballooning at an alarming rate, and the path to profitability remains long. For investors interested in participating in their IPOs, this financial picture demonstrates both tremendous growth potential and clearly identifies the risks.
Revenue soared, but losses were equally alarming.
Both OpenAI and Anthropic expect their revenue to more than double this year, with the accelerated adoption of AI tools by enterprise customers being the main driver.
OpenAI's revenue streams include consumer subscriptions, enterprise services, and new products (including hardware); Anthropic, on the other hand, relies almost entirely on enterprise customers and includes sales through cloud partners in its revenue.
However, behind the impressive revenue growth lies an equally alarming scale of losses. OpenAI projects that even with significant revenue growth by 2028, it will still incur a loss of $85 billion that year. The company expects to achieve overall break-even by 2030, while Anthropic anticipates reaching this milestone much earlier.
It is worth noting that both companies use a dual-caliber profit disclosure method: after excluding expenditures on "research computing power", OpenAI is expected to achieve a small pre-tax operating profit this year, as is Anthropic in the most optimistic scenario; however, once training costs are included in the calculation, both are deeply in the red.
Computing power arms race: Cost is the biggest variable
The out-of-control training costs are the core source of pressure on the financial structures of both companies. Each generation of model intelligence requires far more computing power to improve than the previous generation, and both companies are currently releasing new models at an unprecedented frequency.
OpenAI projects that computing power spending for AI research will reach $121 billion by 2028. In comparison, Anthropic's training expenditures are smaller, but its financial projections also show a trend of continuously rising computing power costs.
Inference costs (i.e., the expenses incurred in processing user queries) also constitute a significant burden. Currently, inference costs account for over 50% of the revenue of both companies, although this percentage is expected to gradually decrease as technological efficiency improves. ChatGPT's paying users represent a very small percentage, meaning that a large portion of inference costs cannot be covered by revenue.
In terms of cash flow, both companies will continue to consume large amounts of cash in the coming years, and IPO proceeds are seen as a key source of funding to maintain business operations.
Anthropic's annualized revenue surpasses 30 billion, securing new allies in computing power.
According to Bloomberg, Anthropic's annualized revenue has surpassed $30 billion, more than doubling from $9 billion by the end of 2025. Currently, there are over 1,000 enterprise customers spending more than $1 million annually on the Anthropic platform, a number that has doubled since February of this year.
To support this growth momentum, Anthropic has signed a major computing power cooperation agreement with Broadcom and Google. According to a filing by Broadcom on Monday, the three parties will expand their strategic partnership, giving Anthropic approximately 3.5 gigawatts of computing power starting in 2027. Broadcom is developing chips based on Google's Tensor Processing Unit (TPU) to provide an alternative to Nvidia, and the two companies have signed a long-term supply guarantee agreement extending to 2031.
Anthropic CFO Krishna Rao stated that the partnership with Broadcom and Google will help the company build the computing power foundation necessary for a significant increase in its customer base. Following the announcement, Broadcom's stock price rose as much as 3.6% in after-hours trading.
Furthermore, to accommodate the potentially record-breaking IPO fundraising needs of the two companies, Wall Street is seeking to circumvent existing rules. Bankers are lobbying major index providers to relax inclusion criteria; Nasdaq recently announced it will allow newly listed companies to join its indices more quickly, thereby gaining access to a broader pool of capital. OpenAI stated that the company is currently prioritizing growth over profits, and while it can cut training spending, it expects the related investments to yield substantial returns.
