Summary of Key Points
The four major cloud computing giants (Microsoft, Oracle, Google Cloud, and AWS) collectively hold a "Reserve for Purchase Orders" (RPO) worth $2.1 trillion, half of which comes from two unprofitable AI startups, OpenAI and Anthropic. This model of relying heavily on a few customers is essentially a financial cycle that feeds itself: cloud providers invest in AI companies; the AI companies use these funds to purchase cloud services; the cloud providers then report improved financial results, driving up their stock prices; they use the increased capital to issue bonds to acquire NVIDIA GPUs. However, AI companies will need to generate astronomical revenues over the next few years to repay this "computing power debt." Given that the return on investment for enterprise AI applications is extremely low and coupled with geopolitical risks, this capital frenzy could repeat the mistakes of the 2000 fiber optic bubble.
I. Giant Customers: Cloud Giants Held by Two AI Companies
Previously, cloud providers relied on the "long-tail effect"—receiving revenue from millions of smaller customers. Now the situation has changed:
- Oracle is the most aggressive: Of its $553 billion in RPO, $300 billion comes from OpenAI's StarGate project (accounting for 54%). This order, which will begin to be delivered in 2027, is eight times Oracle's annual revenue. To build data centers, it has laid off 18% of its staff and seen its debt soar to six times its equity.
- Microsoft is heavily dependent: OpenAI accounts for $250 billion, and Anthropic for another $30 billion, totaling 49% of its commercial RPO. In exchange, Microsoft has given up some IP royalties and exclusivity rights to OpenAI.
- AWS and Google are closely aligned with Anthropic: Google Cloud's $200 billion in RPO comes from Anthropic; AWS has secured a 10-year, $100 billion contract with Anthropic, meaning these two AI companies account for 51% of its total RPO.
The result is that ordinary customers suffer: AWS has increased the price of its AI computing power by 15% as the giants deplete available capacity, forcing smaller businesses to pay higher premiums for the remaining resources.
II. The Financial Loop: Money Circulates Within the Ecosystem, with Only NVIDIA Making Real Profits
Pessimists describe this as a financial "Ponzi scheme," and the logic is straightforward:
1. Cloud providers (such as Amazon and Google) invest real money in Anthropic.
2. Anthropic uses this money to purchase computing power services from the cloud providers, which then becomes part of their RPO.
3. The cloud providers use the increased RPO figures to boost their financial reports on Wall Street; when their stock prices rise, they issue bonds.
4. The proceeds from these bonds are used to buy NVIDIA GPUs, which represent the actual cost of operations.
In this entire chain, the "profits" shown in the cloud providers' financial reports are merely self-deception—money leaves the companies and then returns in the form of purchase orders, with only NVIDIA actually receiving cash for its hardware.
The bearish investment firm Kerrisdale previously shorted CoreWeave by exploiting this logic.
III. The Growth Dilemma: Can AI Companies Repay the Trillion-Dollar Debt?
For OpenAI and Anthropic to fulfill their $1.05 trillion in computing power orders, they need to achieve unrealistic revenue targets:
- OpenAI currently has annual revenues of $25 billion and needs to grow by 11 times to reach $28 billion by 2030.
- Anthropic is valued at nearly a trillion dollars and must generate $148 billion in revenue by 2029 (an increase of 4-5 times).
However, the reality is bleak: According to MIT research, 95% of corporate investments in AI yield zero returns. Apart from simple tasks like writing copy or coding, businesses are reluctant to use expensive large-scale models for their core operations. Just as it took Tesla seven years to grow its revenue from $1 billion to $10 billion, it seems almost impossible for AI companies to achieve such growth in just a few years.
IV. The Shadow of a Bubble: More Dangerous than the 2000 Fiber Optic Bubble
This situation closely resembles the telecommunications bubble of 1996-2001: Wall Street believed the lie that internet traffic would double every 100 days, leading to an investment of $1.4 trillion in fiber optic infrastructure, which resulted in 85% of it being unused and the bankruptcy of WorldCom. The current issues are even more severe:
- Lack of Application Momentum: While the fiber optic bubble had at least a clear direction (it supported the development of mobile internet), AI applications are still far from becoming commercially viable.
- Geopolitical Risks: The three major cloud providers control 63% of high-end computing power. Iran's attacks on AWS' data centers in the Middle East, and the U.S. Department of Defense's concerns about Anthropic's supply chain risks, mean these $2 trillion in orders could be revoked at any time by governments.
V. Cloud Giants' Self-Defense: AWS' Strategy
AWS has played a clever move by securing a 10-year contract with Anthropic that requires the use of Amazon's proprietary Trainium chips instead of NVIDIA GPUs. This strategy has two benefits:
1. It reduces dependence on a single hardware supplier (NVIDIA).
2. If the bubble bursts, its data centers, liquid cooling systems, and proprietary chips will be tangible assets that won't remain unused like the fiber optic infrastructure.
However, this does not prevent the inevitable reckoning: AI companies will need to go public (IPO) between 2026 and 2027 to transfer the risks to secondary market investors. Once their financial reports are made public and the true cost of operations is revealed, the truth about this frenzy will emerge—whether it's a stepping stone to artificial general intelligence (AGI) or just another bubble in Silicon Valley. Time is running out for them to prove otherwise.