Summary of Key Points
The long-term opportunities in the AI industry still exist, but the business model of buying GPUs to rent them out for a steady profit is on the decline. Over the past two years, the market has treated GPUs as a form of “digital real estate,” assuming that computing power would always be in short supply, leading many companies to borrow heavily to purchase GPUs and rent them out. However, the logic has changed: GPUs are rapidly depreciating assets (unlike houses that don’t become obsolete), and tech giants like Meta are adjusting their hardware depreciation schedules to adapt to technological obsolescence. They are even entering the rental market themselves, offering idle computing power at lower costs than rental companies. The investment focus has shifted from “who owns the most GPUs” to “who can generate the most cash flow from them.” The computing power rental industry is about to undergo a major reshuffle.
1. Are AI Giants Starting to Be More Frugal?
The essence of the AI race is spending money on hardware, but lately, these giants have begun to be more cautious with their expenses. For example, Meta increased the depreciation period for its servers from the previous years to 5.5 years, saving $2.9 billion in annual losses—this isn’t financial fraud; they realized that AI hardware is being replaced too quickly, and the theoretical lifespan of the equipment doesn’t match the actual technical life expectancy.
To illustrate: The H100 GPUs that were highly sought after a year ago are now outdated, as NVIDIA has released newer models like the H200 and Blackwell. This means that the billions of dollars spent on these GPUs by giants could be wasted before they even fully depreciate, reducing their asset value. Even financially strong companies like Meta and Microsoft are starting to worry about this, let alone smaller firms.
2. The High-Leverage Weakness of Computing Power Rental Companies
The business model of computing power rental companies is straightforward: they borrow money to buy GPUs and then rent them out to AI companies. In the past, when GPUs were scarce, customers were eager to rent them, driving up rents and seemingly ensuring stable profits. However, this model has a fatal flaw: their income depends on the scarcity of computing power, but their assets (GPUs) depreciate rapidly with technological advancements.
Take CoreWeave as an example. The company initially focused on cryptocurrency mining before transitioning to GPU leasing and borrowed $8.5 billion in 2025 to purchase equipment. If new GPUs become more readily available or the next-generation chips significantly improve performance, the rents for older GPUs will plummet. Unlike houses that can still be used despite new developments, older GPUs won’t be able to handle newer AI models, leading to a decline in demand and potentially insufficient profits to cover interest and depreciation costs.
3. Meta Entering the Rental Market: A Competitor That Threatens Rental Companies
The biggest threat to rental companies isn’t just GPU depreciation but the fact that their largest customers are becoming direct competitors. Meta is now exploring the possibility of renting out its own idle computing power. Since it bought these GPUs with cash, it doesn’t incur financing costs. In a price war, giants can set very low rents (since idle resources are essentially wasted), making it difficult for rental companies to compete.
Previously, rental companies profited by acting as intermediaries, buying GPUs at lower prices and reselling them at higher ones. Now that giants are selling computing power directly, the role of intermediaries is significantly diminished.
4. The Shift in AI Investment Strategy
Over the past two years, the market focused on who had the most GPUs and who could raise the most capital (NVIDIA’s stock price soared because of this belief in the permanence of AI infrastructure value). However, investors are now asking more practical questions: “When will these GPUs start generating profits?” and “Will rental profits be sustainable?”
As a result, the AI industry is becoming more selective. Companies with chip technology (like NVIDIA), ecosystem advantages (like Microsoft), or the ability to generate revenue from GPU-powered services (such as AI application developers) are still favored. In contrast, those that merely buy GPUs to rent them out are under increased scrutiny from investors, who are concerned about their cash flow: how much money they have borrowed, whether their rents cover interest and depreciation costs, and how quickly their assets are devaluing.
In short, the AI revolution is real, but not everyone will profit from it. In the past, the focus was on acquiring as much computing power as possible; in the future, it’ll be about generating actual cash flow. The major reshuffle in the computing power rental industry has just begun.