If the AI Bubble Bursts, It May Not Die from Technology, But from Accounting Methods
Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle
Regarding whether AI is a bubble, the most common debate over the past two years has centered on one question: are large models actually useful?
BitMEX founder Arthur Hayes recently offered a different perspective. He believes the key variable in this round of AI capital spending may not lie in AI itself at all, but rather in how the financial system accounts for it. Markets value data centers, power facilities, and chip clusters as technology investments, but the actual form of these assets more closely resembles real estate: one-time heavy capital construction, recouped through long-term lease-like cash flows, with massive credit propping it all up. Once credit expansion exceeds the pace of cash flow recovery, the narrative switches from a tech story to a credit cycle story.
This perspective deserves serious consideration, because the 2008 real estate crisis followed exactly this path: houses always had value, but what broke down was the leverage and credit chains built around them.
Technology Accounting or Real Estate Accounting
Let’s lay out both sets of books.
By technology accounting, data centers are next-generation infrastructure, similar to investing in a growth-stage tech company: large investments, rapid growth, and future profit margins that will continue to rise, so they deserve high-multiple valuations. The trajectory of NVIDIA and various computing power concept stocks over the past few years has basically been priced according to this accounting.
By real estate accounting, the same asset takes on a completely different appearance: construction capital comes from credit, recovery depends on renting out computing power, the asset itself has a long useful life but rapid technological depreciation, and cash flows are stable but growth is limited. Under this model, the valuation ceiling isn’t “how sexy the future is,” but whether rental returns can cover financing costs. Real estate financial history has repeatedly proven that these types of assets fear two things most: rising interest rates and oversupply.
Hayes’ judgment is that the market currently prices these assets using technology accounting, but their cash flow structure follows real estate accounting. The gap in between is where the bubble lies. What’s more troubling is that a considerable portion of the liquidity supporting this gap comes from private credit funds and leveraged financial arrangements that believe they’re buying “future technology infrastructure,” when in reality the debt they hold is collateralized by physical space and power contracts.
Inflection Points Look at Growth Rate, Not Absolute Values
The most operationally valuable judgment criterion in this analysis concerns the location of inflection points.
Asset prices typically peak during the acceleration phase of growth and begin falling when growth is still far from stopping. The reason is that valuations price in expectations, and the variable expectations are most sensitive to is the change in growth rate—the second derivative. As long as growth continues to accelerate, people will hold even absurd valuations; once growth rate slows, even if absolute values remain huge and companies are still growing, the market will immediately reprice.
Mapping this to AI infrastructure, this standard points to an easily overlooked possibility: sustained growth in computing power demand and sustained growth in data center construction are not the same thing. Chip efficiency is improving, with new generation hardware producing more computation with lower energy consumption. A situation could easily arise where token consumption grows year over year, but demand for new data center construction starts declining at some point, because the supply capacity of existing computing power is continuously amplified by efficiency gains.
If that happens, the first to take a hit would be data center debt and equity valued as “technology assets,” while the AI application layer business might actually benefit: excess computing power drives down inference costs, and the layer that uses computing power to produce value directly captures the price reduction dividend.
After the Burst, Where Does the Money Go
The final link in this chain of reasoning extends from AI to the monetary system, and is also the most controversial.
His reasoning is: historically, whenever an industry that has received substantial credit support experiences investment excess that threatens financial system stability, the government’s response has been to inject liquidity—the massive easing after 2008 is the most direct precedent. If AI infrastructure credit really does run into problems, policymakers will most likely follow the same playbook. And Bitcoin’s core narrative is precisely a hedge against this kind of monetary expansion—it was born after the 2008 financial crisis as a response to the credit expansion of fiat currency systems. So if the AI credit cycle ends with a liquidity rescue, when capital seeks inflation hedge assets, Bitcoin is a candidate.
My view on this link is: the directional reasoning holds, but the transmission chain is too long and the timing is completely unpredictable. From “AI infrastructure capital misallocation” to “financial institution asset repricing” to “policy intervention and liquidity injection,” each step could be interrupted by market self-correction, technological efficiency improvements, or industry structure changes. It’s more reasonable to treat it as a hypothesis worth continuous tracking than as a conclusion to bet on.
What This Means for Entrepreneurs and Investors
Stripped of the crypto narrative, the usable parts of this framework for ordinary practitioners and investors are actually three checking actions.
First, look at the capital expenditure of AI-related companies and distinguish between financing structure and cash flow structure. Companies that cover capital expenditure with operating cash flow versus computing power assets built on credit and financing leases have completely different survival capabilities when cycles reverse. The former are technology companies, the latter are real estate companies—don’t be fooled by the same AI concept face.
Second, watch for inflection points in growth rates, not absolute values. When NVIDIA’s data center revenue growth rate, North American cloud providers’ capital expenditure growth rate, and sequential changes in token consumption start turning downward, it’s often the starting point for valuation repricing, even though absolute values still look prosperous.
Third, if you truly believe “computing power will eventually be oversupplied,” the implication isn’t to buy Bitcoin, but to shift attention from companies selling shovels to companies using shovels to do work. Long-term declining inference costs benefit the application layer that uses inference to build products. This judgment doesn’t depend on any monetary narrative, only on the most basic physical trend of efficiency improvement.
As for whether the bubble will actually burst and when, the honest answer is nobody knows. But getting the accounting right at least ensures that when the avalanche happens, you know whether you’re standing on the technology accounting side or the real estate accounting side.