Author: He Lüheng, Editor: Li Motian
Last November, Wall Street was in an uproar over the depreciation issues of computing power giants.
The controversy was sparked by renowned short seller Michael Burry, who posted on social media that computing power giants artificially inflated their profits by extending the depreciation cycles of AI servers.
The longer the equipment depreciation cycle, the less depreciation expense is allocated each year. For example, if a server costing 6 million RMB is purchased and depreciated over 4 years, 1.5 million in depreciation expenses must be deducted from the books annually; extending this to 6 years reduces the annual deduction to 1 million, resulting in a 500,000 increase in profit.
Last year, Meta increased the depreciation period for its servers from 4-5 years to 5.5 years, netting an additional $2.9 billion in book profits, accounting for 4% of its pre-tax profit for the year. Over the past five years, the other four giants have also used this method to boost their profits, and not just once.
The actual depreciation cycle of AI servers largely depends on the iteration speed of the core component, the GPU. While a GPU has a long physical lifespan, it can be replaced by a newer model if it fails to keep up with the changing demands of AI models.
As NVIDIA's GPU iteration cycle has shortened from 18-24 months to 12 months, it theoretically accelerates the pace of GPU upgrades. This means that the depreciation period should actually be shortened, rather than extended.
According to Richard Jarc, an analyst at Uncovered Alpha, the true economic lifespan of a GPU is closer to one to two years [1].
Prominent investor Jim Morrow sarcastically noted that the giants are well aware of this, stating, "They go to great lengths to tweak accounting treatments and depreciation schedules to get ahead of the curve—which is really about avoiding putting all these capital expenditures on the income statement [1]."
According to first-quarter reports, institutions expect the combined capital expenditures of the five major giants to exceed $700 billion in 2026, and to surpass $1 trillion next year. Given such massive investment, it is understandable that they might want to use extended depreciation as a slight buffer.
However, those truly engaging in GPU "speculation and profiteering" are actually other players.
Speculation and Profiteering
Not long after the depreciation controversy among computing power giants, a massive loan at the beginning of the year reignited the topic.
The borrower is CoreWeave, an AI computing power company that started by flipping GPUs and made a fortune renting them out during the peak of the crypto mining era. In 2022, it claimed to transform into an "AI cloud service provider," but it was essentially the same business under a new name—still renting out GPUs, except the tenants shifted from crypto miners to AI companies and cloud giants.
Where does the money to buy GPUs come from? Financing and borrowing. Therefore, how to borrow money and how to borrow more has always been CoreWeave's top priority.
In 2023, CoreWeave raised $2.3 billion for the first time using GPUs as collateral, which sparked controversy at the time.
Generally speaking, collateral for loans is chosen for its stable value and ease of liquidation. For instance, banks typically prefer real estate such as houses and land. Depreciating consumables like GPUs, which are obsolete every few years, were almost never considered before.
The loan at the beginning of the year also used GPUs as collateral. Besides the amount being nearly quadrupled, what makes it even more special is that it was rated as investment-grade by credit rating agencies Moody's and DBRS, a level of bonds that even pension funds and insurance companies can purchase.
This indicates that institutions judge the probability of default or failure to repay this loan over the next six years (maturing in 2032) to be extremely low. Even if they fail to repay, selling off the company's existing GPUs would be sufficient to cover the debt.
This brings us back to the original controversy: How long can a GPU actually retain its value?
Three years or six years is merely a matter of how the books look for computing power giants, but for companies like CoreWeave, which rely on GPUs both to make money and to borrow money, it is a matter of life and death.
Typically, the launch of a new GPU directly impacts the sales of the previous generation. When NVIDIA released its latest Blackwell architecture GPU, Jensen Huang half-jokingly and half-seriously said that once Blackwell goes into mass production, the previous-generation Hopper might be "impossible to give away even for free."
According to statistics from a tracking agency at the beginning of the year, the selling price of a three-year-old used H100 has already dropped to 45% of the price of a new model [2].
As the GPU iteration cycle has significantly shortened, the pace of this depreciation is also accelerating. Theoretically, a GPU from six years ago is not far from being considered electronic waste for today's most advanced AI data centers.
However, in the current market environment of extreme supply-demand imbalance, these "theories" seem to have started failing.
CoreWeave CEO Michael Intrator revealed last year that their A100 chips (released in 2020) were completely sold out, and a newly purchased batch of H100 chips (released in 2022) had already been fully pre-ordered at 95% of the original price [3].
Regarding the controversy over the $8.5 billion loan at the beginning of the year, Intrator was quite dismissive in an interview with Forbes, saying, "Duh, of course I'm going to borrow money [4]." Another co-founder, Brannin McBee, casually echoed, "There is absolutely no risk [4]."
In their view, if everything goes according to plan, CoreWeave will not only retain the value of its GPUs but also generate a certain profit.
In February this year, Amazon, the first among the computing power giants to extend the GPU depreciation cycle, reduced the useful life of some servers from six years back to five years, citing a study that found "the pace of technological advancement is accelerating, particularly in the fields of artificial intelligence and machine learning."
Other giants have also started to be vague. In its annual report last year, Microsoft set the useful life of its equipment at two to six years. Such a flexible range is certainly not without reason.
But no matter how the giants and capital judge the lifespan of GPUs, they cannot stop this trend of speculation and profiteering from escalating.
GPU Brokers
Without this wave of AI, companies like CoreWeave would most likely not have survived until today.
In the first few years when CoreWeave just started offering cloud services, the entire market was already dominated by the three giants: Amazon, Microsoft, and Google. CPUs were still the absolute mainstay of data center servers.
The sudden AI boom made the GPUs gathering dust in CoreWeave's warehouses priceless overnight.
Traditional cloud giants did not have enough GPU stock, not even enough for training their own large models. With AI companies unable to find GPUs elsewhere, CoreWeave and similar companies became the first choice.
As an NVIDIA "Preferred Partner," CoreWeave and its peers can get priority access to GPU allocations, with a priority even higher than that of the cloud giants. In 2024, Microsoft, as CoreWeave's largest customer, accounted for over 60% of CoreWeave's annual revenue with its orders.
What was once considered "heavy asset junk" has turned into "tech nouveau riche," but everyone knows the real core competitiveness well—"having a good relationship with NVIDIA."
In March 2025, CoreWeave successfully went public, with its stock price surging over 300% within three months. Lambda Labs' valuation rocketed from $200 million to $6 billion in two years, much to the anger of highly respected veteran analysts.
Investment research firm New Constructs slammed CoreWeave as being "Rotten to the Core" [5], while D.A. Davidson analyst Gil Luria bluntly stated, "If NVIDIA doesn't want it to exist, it simply cannot continue to exist [4]."
Legendary short seller Jim Chanos believes CoreWeave doesn't even qualify as a tech company, but rather an equipment leasing and financial company. He earnestly advised, "Do not apply magical valuations to mediocre business models [6]."
What draws the most criticism from Wall Street regarding CoreWeave and its peers is their snowballing, ever-increasing debt and their "permanently negative" net profits.
In the first quarter of 2026, CoreWeave's total liabilities skyrocketed from less than $20 billion a year ago to over $50 billion. The high interest expenses brought by this debt have kept CoreWeave in a loss-making state to this day, with a trend of losing more and more.
Next door, Lambda Labs is also engaged in "continuous strategic losses," while Nebius, the latest entrant, added $4 billion in debt in one year to catch up.
The accumulating debt is the price of maintaining competitiveness: borrowing more money to hoard more assets, then renting out the assets to make money from the spread, repaying loans, and reinvesting. The last industry to do this was real estate.
This business model has been described by analysts as a "GPU Debt Trap," where capital requirements continue to grow, yet it never reaches the scale needed to achieve profitability.
Chanos believes that this business model of "immediately investing today's rental income into the next generation of more expensive hardware" will result in shareholders never seeing true free cash flow dividends. Even in the current situation of extreme supply shortage, the expected pre-tax return on investment for owning data centers and chips is only between 5% and 8% [6].
By the end of the first quarter, CoreWeave's Remaining Performance Obligations (RPO) reached an unprecedented $99.4 billion, and Nebius also signed a $27 billion forward supply agreement with Meta in May. All of this has encouraged CoreWeave and its peers to more boldly leverage up to buy GPUs.
The rent collected hasn't even had time to warm up in their own pockets before being thrown into the ever-rising capital expenditures.
CoreWeave has set its full-year capital expenditure for 2026 at $31 billion to $35 billion, more than doubling compared to 2025. Lambda Labs has also explicitly stated that it will use all the billions of dollars raised from its latest financing to expand its computing centers.
Fairly speaking, companies like CoreWeave are merely minor players in the entire industry; whether they live or die cannot affect the whole chessboard. However, their prosperity reflects the illusion brought to the industry by the rapidly expanding market that "computing power will always be scarce."
Conversely, when this illusion is shattered, these companies will be the first to die.
Shattering the Illusion
It is Meta that shattered the illusion of CoreWeave and its peers.
On the first day of July, foreign media reported that Meta was going to lease out its idle AI computing power. Subsequently, an audio recording of an internal company meeting was leaked, in which Mark Zuckerberg admitted that AI R&D was slower than expected.
The chemical reaction of these two pieces of news directly crashed US chip stocks. As direct targets of the blow, CoreWeave and Nebius both plummeted by over 10%.
The reason this has such a devastating impact is, first, because Meta occupies a very special ecological niche.
Although it ranks as one of the five major hyperscalers, unlike Amazon, Microsoft, and Google, which have cloud service revenues to amortize data center expenses, all of Meta's self-built data centers serve its own business, with advertising being almost its sole source of revenue.
Therefore, Meta's business condition provides a more "unvarnished" reflection of the true supply and demand situation across the entire industry.
In the past few years, Meta's aggressive hardware investment has been well known. Jensen Huang's evaluation of this major client was, "No company is deploying AI at a scale larger than Meta." Now that the enthusiast wants to lie flat and collect rent, anyone would suspect that the times are getting tough and the demand for computing power is about to saturate.
Second, it means that the computing power leasing market, which has been growing wildly, has now seen the entry of a colossal legitimate force.
According to Bernstein's estimates, Meta currently has a data center capacity of approximately 20 GW [7]. For comparison, CoreWeave's operational computing power by the end of 2025 was 0.85 GW, and Lambda Labs is reportedly even less.
Although Meta is only leasing out a portion of its computing power, the scale will still be quite substantial.
Moreover, Meta bought its chips with its own money, so all the rent collected is purely its own. In contrast, CoreWeave and its peers bought theirs with borrowed money; every bit of rent earned must first be partially handed over to banks to pay off interest. Achieving break-even is already difficult, and once Meta gets tough and starts a price war, they will most likely only be able to take a beating passively.
A more direct blow lies in the loss of potential orders. Over the past two years, Meta has been one of the biggest patrons of computing power leasing.
In March this year, Meta signed a $27 billion supply agreement with Nebius, and in April, it added a super-sized $21 billion order to CoreWeave. Coupled with the $14.2 billion contract signed last year, Meta has already become one of CoreWeave's largest customers.
Already signed orders cannot be breached, but with a major customer turning into a competitor, it remains to be seen whether such massive orders will be signed in the future.
According to Morgan Stanley's calculations, assuming Meta leases out 0.25 GW of computing power for one year at a price of $40 per watt, it would bring Meta an upside of $3 in EPS by 2028. If the leasing scale reaches 1 GW, this upside would be increased to nearly $12 [8].
For such a foolproof business, it is only natural that Meta is tempted. Shareholders and the capital market are also fully supportive; on the day the news broke, Meta's stock price once surged against the market trend by over 10%.
Analysts who have recovered from the shock are also gradually realizing that Meta collecting rent is not necessarily bad news for the industry.
On the one hand, according to the first-quarter reports, the RPO of cloud giants has not dropped but increased instead. Microsoft's RPO doubled year-on-year to $627 billion, and Google's doubled quarter-on-quarter to $462 billion.
RPO represents revenue that customers have already signed contracts and paid for, but the services have not yet been provided, which is revenue to be recognized in the future. In other words, these revenues are already secured and will not be affected by changes in Meta's demand.
On the other hand, Meta has actually provided its peer giants with a decent template for admitting mistakes and remedial measures—
Even if computing power is truly overbuilt now, the excess capacity can be leased out in the future, turning sunk costs into substantial revenue. Before Meta, Elon Musk's xAI had already done this.
The Next Web predicts that this model will become increasingly common in the future: building resources far exceeding current computing needs, betting that they will be needed in the future, while leasing out the surplus resources to offset the bills [9].
Analysts who were anxious about the capital expenditures of the giants just half a month ago have now breathed a huge sigh of relief.
When all is said and done, the only ones left hurt are CoreWeave and its peers.
There is no fallback business, major customers can turn into competitors in the blink of an eye, and every single GPU is burdened with high-interest loans. One more idle GPU means being one step closer to bankruptcy.
An overinflated market will eventually correct its course, and CoreWeave and its peers will most likely be the first "mistakes" to be eliminated.
References
[1] The ‘Big Short’ investor betting $1 billion against the AI bubble says Meta and Oracle’s accounting is hiding the brutal truth, Fortune
[2] GPU Debt Has Gone Investment Grade. Here's Who Holds The Risk, Fortune
[3] The question everyone in AI is asking: How long before a GPU depreciates?, CNBC
[4] CoreWeave’s $29 Billion Bet That Its Debt-Fueled AI Boom Won’t Go Bust, Forbes
[5] IPO Report: Valuation Rotten to the Core, New Constructs
[6] Legendary Short Seller Jim Chanos Warns: AI Infrastructure Hides a Financial Time Bomb, Compute Leasing Is Essentially Financial Intermediation, finance.biggo
[7] Meta’s $48bn cloud question: Structural competitor or bursty-capacity opportunist?, Capacity Global
[8] 5 Points of Analysis on Meta's New Cloud and Hyperscaler Opportunities, Morgan Stanley
[9] Meta wants to rent out its spare AI compute, and Wall Street likes the idea, The Next Web