Author: Lin Wanwan

On March 31, 2026, OpenAI announced the completion of US$122 billion in financing, with a valuation of US$852 billion, the largest private equity financing in human business history.

Amazon invests $50 billion in OpenAI. Among them, 15 billion will be received immediately, and the remaining 35 billion will not be transferred until a condition is met.

This condition is that OpenAI completes its IPO or implements AGI.

One is to go public, and the other is to create general intelligence that surpasses humans. The largest e-commerce company on earth bet an amount of money that is higher than the annual military expenditure of most countries on an "or".

Let's break down all OpenAI's financing and look at the structure.

Nvidia spent 30 billion, and OpenAI happens to be one of Nvidia's largest GPU customers.

OpenAI CFO Sarah Friar herself has said that most of the money will go back to Nvidia.

Amazon's 50 billion investment, OpenAI puts the model on AWS to run inference, AWS revenue has increased, and Amazon's financial report is good. Microsoft has invested more than $13 billion in total, and OpenAI has committed to purchasing $250 billion in cloud services on Azure.

Money goes around in a closed loop and comes back again. Wall Street calls this circular financing.

Bernstein analyst Stacy Rasgon said each such deal is deepening concerns about circular financing. Statistics from the CFA Institute are even more disturbing. The total number of mutual investment and mutual procurement commitments in the field of AI is close to $1 trillion.

But the topic of revolving financing has been discussed for a whole year, and everything that needs to be said has been said.

What is really worthy of attention in this 122 billion financing is not how the funds circulate. It lies in a more direct question: What exactly is the money buying?

What did 852 billion buy?

The answer is, buy time. To be precise, buy the time before the IPO.

OpenAI currently has monthly revenue of 2 billion, which is approximately $24 billion annually. 8520 亿估值对应大约 35 倍市销率。 This multiple means the market is paying for OpenAI three or four years from now.

Find a few frames of reference to get a feel for it. NVIDIA PS about 20x while making crazy money. Snowflake peaked at 100x but quickly fell back below 30x. It was about 10x when Salesforce went public.

35 times is already very aggressive for a company that is still losing money.

OpenAI's own plan is to achieve revenue of 100 billion and profit of 14 billion in 2029. From 24 billion to 100 billion, the compound annual growth rate will exceed 40% for four consecutive years. I carefully thought about the software companies in history that have maintained this growth rate on a revenue base of tens of billions of dollars, and I couldn't find any.

There is only one condition for the 852 billion valuation to be established: someone is willing to take over at this price in the open market. That said, the IPO must be successful.

Once this layer is figured out, the entire financing structure will make sense.

Amazon's 50 billion and 35 billion are subject to IPO conditions. If it doesn't get listed, it won't get paid. SoftBank will invest 30 billion in three tranches. The first tranche will be paid when the financing is closed, and the subsequent two tranches will arrive in July and October, accurately laying out the key nodes in the IPO preparation period.

OpenAI sold 3 billion shares to retail investors for the first time through banks, and also entered ARK Invest's ETF. Retail investors buy shares and enter the ETF, and when the IPO opens, they will be the natural base buyers.

The wording in the financing announcement no longer sounds like a report to private equity investors. "We are the fastest platform to reach 10 million users, the fastest to reach 100 million users, and will soon be the fastest to reach 1 billion weekly active users." "Our revenue growth rate is four times that of Google and Meta in the same period." This set of words can be directly transferred to the first page of the prospectus without any modification.

PitchBook has a study that points out that among the three largest AI IPO candidates: OpenAI, Anthropic, and Databricks, OpenAI has the lowest business quality fundamental score but the highest valuation.

Every design detail of the 122 billion financing points in the same direction. Let this company go public and let the public market pick up this valuation.

Two companies compete for the same faucet

OpenAI needs an IPO, but it's not the only one. That’s the real drama in 2026.

Look at the queue list first. CoreWeave was listed in March last year, with an issue price of US$40. Now it is US$130, with a market value of over 46 billion, setting an example for the companies behind it. Databricks is valued at $134 billion in roadshows, with annualized revenue of nearly $5 billion. Cerebras resolves CFIUS review to resubmit IPO filing.

The real heavyweights are Anthropic and OpenAI. Anthropic is valued at $380 billion and has hired Wilson Sonsini for IPO legal preparations. Kalshi predicts a 72% chance that Anthropic will hit the market before OpenAI.

The odds are terrible for OpenAI. The pool of funds in the market that want to buy AI targets is limited. If Anthropic eats up this batch of funds and attention first, OpenAI’s IPO pricing will be compressed.

And Anthropic is indeed encroaching on OpenAI's territory. Enterprise API market share, OpenAI fell from 50% in 2023 to 25% in mid-2025, and Anthropic rose from 12% to 32% in the same period. Anthropic's revenue growth is roughly three times that of OpenAI. Some analysts extrapolate based on the current curve, Anthropic will exceed OpenAI's annualized revenue in mid-2026.

OpenAI dominated the enterprise market two years ago, and now Anthropic is the leader in the enterprise API market. Claude Code's annual revenue from one product is US$2.5 billion, and it contributes 4% of the world's GitHub public commits. This speed of reversal is also rare in the technology industry.

OpenAI certainly has its own trump card. With 900 million weekly active users and 50 million paid subscriptions, the advertising business’s revenue exceeded 100 million in the sixth anniversary of its trial. ChatGPT’s brand recognition and user habits remain the biggest moats in the AI ​​industry. But the stalling on the enterprise side is real.

Both companies are also spending money at an alarming rate.

OpenAI is expected to lose 14 billion in 2026, and the annualized cash burn rate may reach 57 billion by 2027. 122 billion in financing sounds astronomical, and it will last about 18 to 24 months. Anthropic is expected to spend 19 billion in 2026, 12 billion to train the model and 7 billion to run inference.

Whoever goes public first will be the first to survive. The money in the private market can no longer feed these companies, and the public market is the last faucet that has not yet been opened. Renaissance Capital predicts that there may be 200 to 230 IPOs in 2026, and the IPO financing scale of OpenAI, Anthropic, Databricks, and Cerebras alone may exceed US$200 billion.

This is the largest tech IPO window since 2000. The last time there was an IPO wave of this level was in 2000.

Can the speed of making money outpace the speed of spending money

All valuations, all financing structures, all IPO plans are ultimately bets on one judgment. AI can make money faster than it can spend money.

Outperformed, 122 billion in financing is foresight, and 852 billion in valuation is a discounted price.

Some people are modeling the scenario of not being able to win. Analysts call it the CapEx Cliff. Hundreds of billions of dollars in data centers are built, and the software running on them does not make enough money to cover the costs. The efficiency revolution will replace the competition for scale, and companies that have bet all on "bigger is better" will find themselves sitting on a pile of expensive but under-utilized hardware.

Efficiency is advancing faster than most people realize. Training a model of the same level as GPT-4 will cost approximately US$79 million in 2023. By 2026, with a new generation of hardware and distillation, quantification and other technologies, the cost has dropped to US$5 million to US$10 million.

Last year, DeepSeek R1 used less than 300,000 US dollars to train an inference model close to the cutting-edge level. In January this year, it issued a new training architecture paper to continue to make a fuss about efficiency. Google’s latest Gemini 3.1 Flash-Lite has lowered the price of inference to US$0.25 per million tokens. IBM researchers said in public that 2026 will be the year when cutting-edge large models and efficient small models are separated.

If the efficiency route continues to outperform the scale route, the computing power empire that OpenAI built with the money raised from a valuation of 852 billion may face depreciation before it is completed.

The Internet didn't disappear after the bubble burst in 2000. Google grew out of the ashes. Those who died were those companies that raised the most money and built the most infrastructure at the peak of the bubble, but never found a sustainable business model.

AI is not going away either. But whether the $122 billion and $852 billion valuation can survive until the day of profitability is far less certain than it seems.

The drum is still beating, and the beat is getting faster and faster.