Arthur Hayes: AI Bubble Likely to Burst as Credit Expansion Slows

This article is from: Arthur Hayes

Compiled by Odaily Planet Daily (@OdailyChina; Translator: Azuma (@azuma_eth)

Looking around, humans have transformed the natural environment of Earth into something entirely different. Some changes are admirable, others alarming, but without exception, they all began with an idea in the mind of one or more evolved primates—humans.

Because the brain processes vast amounts of information each day, we continuously construct various narratives to make the world feel coherent and meaningful. As a result, narratives themselves ultimately shape reality.

For investors, predicting future price movements in the market requires understanding the collective illusions shared by market participants. The same company, with unchanged future cash flows, can receive vastly different valuation multiples simply because the market believes different narratives.

The simplest way to achieve a re-rate for a previously dull and unexciting company is to give it a new narrative that aligns with current market trends, encouraging investors to chase it regardless of cost.

This also raises the central question of whether AI is currently in a bubble.

However, before discussing the AI bubble, it’s more important to first answer a more fundamental question: “What exactly are we investing in?”—to put it in terms of a romantic relationship, “What is our relationship really?”

At least from the perspective of someone with a somewhat Luddite inclination, the key question is how the market defines AI capital expenditure (AI CAPEX)—is it classified as Technology or Real Estate?

The prevailing market narrative holds that this multi-trillion-dollar AI infrastructure build-out falls under "technology" and therefore deserves a high growth valuation.

But my view is exactly the opposite. AI CAPEX is, at its core, just another dull real estate investment. The only difference is that instead of office buildings, these data centers are filled with computing power. This computing power will ultimately give rise to silicon-based lifeforms, driving the advancement of human civilization—potentially even surpassing the significance of the railway revolution.

It is essential to distinguish between "real estate" and "computing power" because today’s newly matured hedge fund managers, banks, private credit funds, and even governments mistakenly believe they are lending to tech giants like Apple, rather than providing real estate financing to Lehman Brothers.

I believe the reason the AI bubble will eventually burst is that financial intermediaries will overbuild data centers and all the supporting infrastructure required for them, including energy, power, and everything needed to support AI chip training and inference.

Therefore, the AI bubble resembles a 2008-style credit bubble more than a profit-driven bubble like the 2000 internet bubble.

During the 2000 dot-com bubble, most publicly traded internet companies had little to no revenue, let alone profits—such as Pets.com—making the bubble fundamentally an issue of an "earnings story" that couldn't be delivered.

The 2008 financial crisis was different. What truly triggered it was the slowing pace of U.S. housing price increases, which raised concerns among banks and financial institutions about the creditworthiness of mortgage-backed assets, making it a credit crisis (credit story).

The AI bubble will follow a similar logic. The real turning point will not be when leading AI companies stop making profits, but when the pace of data center construction begins to slow, or when hyperscalers lower their guidance for future data center builds.

Even though leading AI companies may still generate substantial profits, their forward multiples will contract due to lower growth expectations. The first to fall will be those AI companies with the weakest credit profiles and the highest leverage.

Subsequently, these risks will quickly spread to the balance sheets of financial institutions holding large amounts of AI-related debt assets and operating with similarly high leverage. Ultimately, the government will once again intervene under the banner of “national security” to ensure that these over-leveraged AI companies and the financial institutions behind them do not fail.

And this misallocated capital will ultimately flow into the crypto market… sending Bitcoin back to the moon.

Whenever someone claims that "AI is in a bubble," AI bulls almost always invoke the Jevons Paradox as a counterargument. Jevons argued that when the price of a commodity falls, its usage increases dramatically, causing the overall market size to continue expanding, even exponentially.

If you believe that AI capital expenditures alone represent demand for computing power, then according to Jevons' Paradox, there is indeed nothing to worry about. As the cost of computing power continues to decline, demand for AI token-consuming applications and AI agents will grow exponentially, making lending to AI infrastructure a naturally profitable (money good) business.

But I believe this is actually a misinterpretation of Jevons' paradox. To understand why Jevons' paradox does not imply that all credit flowing into AI CAPEX will be repaid, let’s examine what a hyperscaler is actually doing when building a data center.

At its core, it begins with a real estate development project: constructing a building to house server racks, then procuring the latest-generation semiconductor chips for AI model training and inference. As industrial technology—particularly semiconductor manufacturing—continues to advance, the number of floating-point operations (FLOPs) per kilowatt-hour will continue to grow exponentially.

Several years from now, whether it’s Nvidia, AMD, Intel, Huawei, or SMIC, all will launch next-generation AI chips with far greater efficiency than today’s. At that time, the same data center will be able to generate over 1,000 times more intelligence while consuming significantly less power.

This means two things can be true simultaneously: on one hand, physical infrastructure such as AI data centers could experience complete supply saturation; on the other hand, the consumption of AI tokens could still grow exponentially.

So, the real question to consider is: do you want to hold the real estate business—the very thing today’s cloud giants are doing—or the AI application layer?

The common rebuttal from AI bulls is that hyperscalers are both landlords and tenants. They rely on the massive cash flows generated by their Web2.0-era “attention-selling” businesses to provide credit support for issuing debt to build data centers; at the same time, they leverage their own AI capabilities to sell the “apples of wisdom” from Eden to the world.

If you truly believe this story, I hope you hold their stocks, not their bonds. Bonds are called fixed income for a reason—no matter how successful the company ultimately becomes, the best outcome for bondholders is simply getting their principal back, plus a little interest.

If Google succeeds in betting on AI and creates a revolutionary product capable of changing the course of human civilization, causing its stock price to soar, shareholders will certainly have reason to celebrate; but bondholders will still receive only their principal.

Conversely, if Google ultimately becomes nothing more than a landlord renting out large quantities of heavily depreciated NVIDIA GPUs, yet fails to generate sufficient revenue to repay its debts and interest, creditors will suffer severe losses. Moreover, it is highly questionable how much value a data center filled with outdated chips could still hold.

The CFOs of cloud computing giants and Wall Street financiers are not foolish. They know they are actually in the real estate business. Therefore, they must find some "latecomers" who believe they are investing in high technology, not real estate.

These buyers, including insurance companies under alternative asset management giants like Apollo, as well as taxpayers across various countries who will ultimately bear the cost of the government’s implicit guarantee of AI-related credit.

If you carefully review those deliberately obscure financial statements, you’ll find that the substantial debt issued to finance AI CAPEX is almost entirely kept off the balance sheet, with little clear connection to the core profit-generating businesses that underpin stock valuations.

How we define AI CAPEX determines how we understand the entire AI investment cycle. It is this narrative that explains why capital is severely misallocated and why the scale of this bubble may surpass that of the railway bubble of the past.

More importantly, since the AI bubble is a credit bubble rather than a profit bubble, when the crisis strikes, the government will inevitably step in to rescue the last buyers who mistakenly treated traditional real estate debt as new technology equity assets.

Don’t assume the AI bull market is over because of recent corrections in the AI sector, especially in highly leveraged markets like Korea. On the contrary, the truly wild “blow-off top” phase may have just begun.

Last week, the Federal Reserve had the opportunity to raise interest rates to address inflation that remained above trend levels across various measures, but it chose to hold steady, with even former Chair Powell voting in favor of maintaining rates unchanged.

So, for crypto traders who have been forgotten by the market and are struggling through sideways bear markets, does AI credit allocation even matter? It matters because it determines how, why, and how much money governments around the world will print to fill the financial gaps created by out-of-control AI CAPEX investments.

The following content of this article will expand on this theory and explain why governments ultimately can only choose to print money to rescue the market.

As the growth rate of AI CAPEX slows while credit expansion continues, Bitcoin will complete its base formation and enter a long-term upward trend. When policymakers finally realize that the AI-driven GDP growth they have pinned their hopes on is merely another ordinary real estate bubble, they will be forced to initiate monetary easing on a scale even larger than the 2008 Global Financial Crisis (GFC).

And ultimately, this will drive Bitcoin past $1 million—and even higher.

I constantly remind myself: "What you're really trading is not growth itself, but the acceleration of growth."

In other words, what we’re really focusing on is the second derivative—whether growth is accelerating or decelerating.

This is actually quite intuitive. When an asset is still in a phase of accelerated growth, people constantly weave stories about its limitless potential. As a result, the market is filled with bold statements like, “I’d rather see a cloud computing giant go bankrupt than miss the chance to be part of building AGI.”

However, any growth eventually enters a slowdown phase. The issue is that the price behavior of most assets often follows:

• Acceleration phase: Prices continue to make new highs;
• Slowing growth phase: Price consolidates sideways;
• The price only truly begins to decline after the growth itself (i.e., the first derivative) turns negative.

No one can accurately predict how long it will take from the moment growth begins to slow down until it truly enters negative growth, but many investors, including myself, instinctively assume that even if growth has started to slow, asset prices can continue rising indefinitely.

If the AI bubble is fundamentally a credit bubble, then the importance of the second derivative becomes even more pronounced, because society’s willingness to continuously finance AI CAPEX rests on the assumption that AI investment will always accelerate.

Once this acceleration disappears, continuing to increase debt becomes increasingly dangerous, but the reality is that no one knows when to stop until they actually take a punch.

Either wait for a financial crisis to unfold; or wait for Ken Griffin (Odaily Planet Daily note: Here referring to Ken Griffin, who recently acquired the positions of the AI stock guru at low prices) to sweep up all your assets at the market bottom.

Therefore, even as investment growth begins to slow, credit volumes often continue to rise. Only when AI CAPEX budgets truly start to decline will the market experience that classic “Wile E. Coyote” moment—when the character has already run off the cliff but doesn’t realize there’s no ground beneath until he looks down, and then plummets instantly.

At that time, the market will begin to identify who has become over-leveraged due to holding large amounts of junk AI CAPEX debt.

Let’s apply this logic to the U.S. subprime mortgage crisis. My favorite course in college was on U.S. housing policy and the mortgage market; the professor had previously served as Deputy Secretary of Housing under the Clinton administration. Coincidentally, I took this course in the spring of 2008—the very time when Bear Stearns collapsed—making it incredibly timely.

The core message of this course is that the government continuously encouraged more people to buy homes to achieve the social goal of "homeownership for all," leading to sustained credit expansion. However, by 2006, many first-time homebuyers were already unable to afford their monthly payments after interest rate resets; their only way to keep paying was if home prices continued to rise at an accelerating pace.

Of course, I'm still waiting for the government to deliver on its promise of "forty acres and a mule." If that's the case, why not just print money to build houses?

The four-panel image below shows:

• S&P 500 Index;
• U.S. building loans and construction activity;
• Case-Shiller U.S. National Home Price Index.

By the end of 2005, the pace of U.S. housing price increases had begun to slow, coinciding with the peak in actual construction investment spending (the orange line in the first chart). However, real estate credit (the purple line) continued to flow into the market until the stock market reached its peak and began to show slight corrections.

From 2006 to 2007, it was essentially a “no man’s land” before the full crisis erupted—housing prices were still rising, but the rate of increase had been steadily slowing; then, the stock market peaked in mid-2007 (the pink dashed line in the chart); the true “Wile E. Coyote moment” occurred in August 2007—with the collapse of three credit hedge funds managed by BNP Paribas; the crisis then spread relentlessly, ultimately toppling Bear Stearns and Lehman Brothers in September 2008… Prior to that, the S&P 500 had already declined by approximately 50% from its peak.

What truly triggered the financial collapse was investors finally realizing who held those toxic "Frankenstein-like" financial derivatives. In the end, governments had to take over both the debt and equity of these institutions to prevent a new Great Depression.

This is very important, as we will return to this logic later when discussing how governments rescue the AI industry.

The second chart is also worth noting. It shows the starting point of capital misallocation, precisely when housing price increases began to slow. If new credit were still being used to build more housing, the problem would not be severe; however, if the system begins to rely on rolling over debt to pay off existing debt, risks begin to accumulate. The rising ratio of construction loans to construction investment spending is the clearest indicator of this process.

Now, apply the same analytical framework to AI. Here, the key variable corresponds to each company’s CAPEX spending plans.

The market currently believes that real estate (here referring to AI) is technology; the more technology is invested in, the higher the future profits will be. Therefore, the market rewards cloud giants that announce increased capital expenditure budgets by driving up their stock prices.

I anticipate that the growth rate of AI CAPEX will begin to slow from mid- to late-2027, with the market clearly entering a "slowing growth phase" by 2028.

At the same time, a seemingly contradictory phenomenon occurs: although the growth rate of CAPEX begins to slow, the volume of credit flowing into AI continues to expand. This is because lenders believe they are investing in technology, not real estate. Coupled with governments around the world consistently emphasizing the need to lead in the global AI competition, continuing to finance all projects related to AI CAPEX appears to be the most rational choice.

Thus, 2027 will become the "no man's land" similar to 2006 to 2007. The current sharp correction in AI stocks is merely a normal adjustment within a bull market. The true peak of the AI bubble will occur next year.

After that, the market will begin to reward cloud computing giants that take the lead in exiting the arms race by proactively cutting their CAPEX budgets. Unlike during the early stages of the bubble from 2022 to mid-2026, future cloud computing giants will find it increasingly difficult to fund AI investments using their own free cash flow and will need to rely more heavily on issuing bonds and raising equity.

The pressure on the balance sheet will also force management to seriously ask: “Is it really worthwhile to keep borrowing money to build more data centers just to house additional chips that will continuously lose value?”

At least for American cloud computing giants, China’s cutting-edge AI models—offering comparable performance at significantly lower prices—will shatter their illusions of being “silicon gods.” After all, if two products are of equal quality, or only slightly inferior, the vast majority of people will choose the cheaper one.

As AI chips continue to generate exponentially more intelligence per kilowatt-hour, and as the cost per token continues to decline under competitive pressure in China, a rational CFO of a major cloud computing company would not further deteriorate their balance sheet merely to build more data centers.

Even under Jevons' Paradox, demand for AI tokens will eventually surge, but this growth won't come fast enough to offset the negative impact of the large volume of debt issued years ago. Ultimately, the market will first punish the participants with the weakest credit. Only then will people truly realize how much capital has been wasted in this AI investment wave.

I can't predict which cloud computing giant will be the first to go too far and trigger a collective cry from bond investors: “Oh shit!”

However, before discussing why banks, despite knowing the immense risks of AI investments, still have to continue lending, let’s first look at the chart below, which compares the committed CAPEX investments by major cloud giants with their cash on hand.

What underpins the entire AI bull market narrative is leverage on a scale of trillions of dollars. Eventually, one of these companies will fall from its pedestal, just as the AI prodigy Leopold Aschenbrenner once did when the market turned. But this time, the rescuers won’t be the greedy and neurotic hedge fund managers of Wall Street; it will be the printing presses in the hands of Warsh and U.S. Treasury Secretary Bessent.

Many believe that the growth rate of AI CAPEX is about to slow, suggesting that the AI bubble is nearing its end. If that’s true, why are banks still continuing to lend?

The reason is actually quite simple: first, because it is profitable; second, because the government wants them to do it; third, because they know that even if the loans eventually turn bad, the government will step in to rescue them.

Louis-Vincent Gave of Gavekal Research published an interesting article last week, arguing that Walsh’s interest rate policy actually follows a very simple logic—actively steepening the yield curve.

There are two benefits to this approach. First, it can make lending more profitable for banks; second, it can gradually dilute the United States’ large debt burden through inflation.

Ultimately, banks will continuously create new loans—that is, new money. This newly created capital will finance the re-industrialization of American manufacturing while continuing to support AI development. This approach aligns closely with the “Hamiltonian Economics” frequently mentioned by Secretary of the Treasury Bessent recently.

According to all objective economic indicators, the Federal Reserve should have raised rates at its most recent monetary policy meeting, but it did not. Instead, long-term Treasury yields subsequently surged.

• Odaily note: The yield on 30-year U.S. Treasury bonds rose rapidly after the Federal Reserve held rates steady.

Many believe this was a policy mistake by the Federal Reserve, but from the banks' perspective, it’s nothing short of a windfall.

The reason is simple. Banks can finance themselves at a cost close to the federal funds rate, and the Federal Reserve deliberately keeps this rate below the nominal economic growth rate—and even below the actual inflation rate.

Subsequently, banks lend the funds to AI data center developers, rare earth mining companies, defense manufacturers, and others in the form of long-term loans. The steeper the yield curve, the higher the net interest margin banks can earn.

And, as shown in the chart below of commercial and industrial loan volumes, the more banks lend, the more incentive they have to continue creating new money through lending.

• Odaily note: The white line represents the 10-year U.S. Treasury yield minus the effective federal funds rate (reflecting the steepness of the yield curve); the yellow line represents the outstanding balance of commercial and industrial loans by U.S. commercial banks.

From a political perspective, this is a sustained Federal Reserve policy. Even though the Trump administration’s Department of Justice investigated and even prosecuted some Federal Reserve governors (such as Lisa Cook and Powell), these voting members continue to support keeping short-term interest rates at negative real levels.

In other words, Warsh actually already had a “coalition of volunteers,” including both individuals from Trump’s camp and those who had been deeply affected by Trump Derangement Syndrome (TDS) but remain within the system.

From a monetary policy perspective, the Fed’s recent actions have enabled Treasury Secretary Bentsen to issue short-term Treasury bills (T-Bills) at yields below the nominal growth rate. If the market cannot absorb the massive weekly volume of Treasury issuances, the RMP (Reserve Management Program) will fill the demand gap by printing money.

To suppress the yields of long-term Treasuries that remain "uncooperative" and continue to rise, Bessent could also implement Treasury buybacks—issuing short-term Treasuries monetized by the Federal Reserve and using the proceeds to repurchase 10-year or 30-year Treasuries, thereby lowering long-term rates.

Notably, Wash, who has long been known for advocating the reduction of the Federal Reserve’s balance sheet, now shows no intention of limiting or even halting the expansion of the RMP program. This is nothing more than a kabuki-style UFC performance taking place on the White House lawn.

If you were a credit approval officer at a "too big to fail" (TBTF) bank and hoped for a promotion and raise in the future, you would almost certainly approve loan applications from "critical industries" such as AI and defense.

The reason is simple: this boosts bank profits while aligning with the policies of the Federal Reserve and the Treasury. Even if the loans ultimately default—which, mathematically speaking, is quite likely—the government will swiftly deploy a bazooka-sized bailout.

There is almost no downside risk here. This is how "window guidance" actually works in the United States.

I believe the 2026 version of the "Treasury-Fed Accord" has already quietly occurred, even though it has not been formally announced. Otherwise, how else could you define the current situation?

• The Federal Reserve maintains negative real interest rates;
• The Federal Reserve prints money to purchase Treasury securities issued by the Department of the Treasury;
• The Ministry of Finance encourages banks to lend to key industries;
• When loans go bad, the ruling government will again step in to cover the losses through implicit guarantees.

If this isn't fiscal and monetary policy coordination, I don't know what is. So, I'm extremely bullish on the market right now—large-scale money printing is far from over.

Now, let’s stretch our imagination a bit further. What if the U.S. government didn’t just bail out banks after a crisis occurred, but instead proactively bought shares in AI companies at the first signs of trouble? After all, imaginative thought experiments are always fascinating.

In fact, the bailout of AI during the Trump era has already begun. Under the banners of “national security” and “U.S.-China competition,” the U.S. government has started borrowing money and directly purchasing equity in companies deemed “critical industries,” such as rare earth elements and semiconductors.

This is essentially an operation to increase dollar liquidity, also understood as equity QE. The dollars that were previously sitting in government accounts are now being directly injected into financial markets.

Below are some examples where the U.S. government directly acquired equity stakes in companies using funds borrowed from the CARES Act, the CHIPS Act, and the Department of Defense budget.

Unfortunately, for cryptocurrency investors whose wealth is entirely dependent on changes in the money supply, there is now very little room left for governments to continue similar equity investments under the existing legal framework.

However, the Trump administration and Treasury Secretary Bentsen have clearly demonstrated a willingness to use borrowed funds to buy AI stocks at a discount, as long as the law permits it.

Thus, a new question arises: is there a way to print money and buy AI stocks before a crisis erupts, without requiring congressional approval?

The answer is yes! And that’s precisely what makes it most interesting.

Under the Federal Reserve Act, the Federal Reserve may directly print money and provide unlimited liquidity loans to a special-purpose vehicle (SPV) established by the U.S. Treasury under so-called “emergency and exigent circumstances.”

During the 2008 financial crisis and the COVID-19 pandemic in 2020, the Treasury used the Exchange Stabilization Fund (ESF) to support first-loss equity, after which the Federal Reserve provided loans to the SPV to purchase various financial assets and stabilize the market.

Currently, approximately $28 billion remains in the ESF account. Bessent could fully use this funding as initial capital for a new SPV, citing the continued need to safeguard national AI security.

Historically, the Federal Reserve has been willing to provide up to 10x leverage to SPVs. This means Bessen could theoretically mobilize around $280 billion to invest in AI companies that are not yet profitable. Of course, compared to today’s AI companies with market valuations in the trillions, $280 billion is far from “heavy artillery.”

So, could it be scaled up further? For example, could the Treasury simply establish an SPV with no first-loss capital buffer, allowing the Fed to lend indefinitely? Technically, this is possible, but it would mean the Fed would face immense political pressure—because outsiders would perceive it as effectively engaging in unlimited equity quantitative easing.

So, does the Federal Reserve really care about political pressure?

The answer is both yes and no. New Chair Wash has consistently emphasized that AI will soon become a miracle for boosting U.S. productivity. In other words, ideologically, he himself believes in the grand narrative painted by AI entrepreneurs.

If Trump told him that in order to save Sam Altman and OpenAI, the government must directly step in to buy shares, because there are no longer enough retail investors willing to put up real money to purchase a cutting-edge AI company that hasn't turned a profit—while at the same time, Anthropic, founded by Dario Amodei, is already profitable and has even stronger models.

Then Wash would likely do so without hesitation. Of course, according to procedure, approving the SPV loan still requires the affirmative votes of three other Federal Reserve governors. But given that at the most recent FOMC meeting, including Cook and Powell, others had already aligned with Wash (in support of maintaining interest rates unchanged), if Wash truly pushes the Fed down this path, I can hardly see any substantial resistance arising.

After all, personal investment gains in a stock account are always more convincing than theoretical concerns about whether money should be printed.

If the Treasury uses newly printed money to support newly listed AI star companies in issuing new shares, it is essentially realizing the unrealized gains on the books of early investors and employees. This is the purest form of “liquidity creation.”

These capital assets didn't exist until the government stepped in to support them. It was the government’s willingness to provide buyers for valuations in the primary market that lacked fundamental justification that truly turned these paper wealth gains into realizable value.

From an accounting perspective, the government can gain two benefits.

First, whenever an AI company has government backing, investors rush in. After all, trading stocks alongside those with the power to print money almost always yields profits—at least in the early stages. As a result, this SPV quickly accumulates massive unrealized gains on its books. Trump could easily package these paper gains as government “profits,” even claiming they could theoretically offset the fiscal deficit. If AI truly is the most important technological revolution in human history, then the mere paper gains on the stock market could, from an accounting perspective, be enough to “eliminate” the entire U.S. fiscal deficit.

Second, the millionaires, billionaires, and even trillionaires created as a result must pay federal and state capital gains taxes when selling their shares. This additional tax revenue can similarly reduce the budget deficit, allowing the government to borrow less and further claim that the U.S. debt-to-GDP ratio has declined. At least initially, bond markets will believe this narrative, causing Treasury yields to fall and rewarding the government’s “accounting magic.”

However, I must emphasize that Trump did not invent the Philosopher's Stone. He merely delayed the problem, hoping that the next administration—preferably another Republican one—will pick up the tab.

Why will this model inevitably lead to disaster? Let’s conduct a simple thought experiment. Suppose you want to become a billionaire overnight and don’t want to do any work. So, you spend a few thousand dollars to register a company and issue a total of 1,000,000,001 shares.

Then, you sell one share to your mother for $1. Since the latest trade price is $1, your remaining 1 billion shares appear to be worth $1 billion on paper. Next, you take this “wealth” to the bank and ask for a $100 million loan to buy a mansion, a Lamborghini, and other luxury items. The bank will simply tell you: “Not a chance.”

You might feel confused. In your view, the loan-to-value (LTV) ratio for this loan is only 10%, so the risk seems clearly low, but the bank’s response is simple:

If these shares need to be sold in the future to repay the loan, there would be no liquidity in the market.

Applying this logic to the AI SPV is the same. If the SPV has become the largest single shareholder of an AI company, and other investors are buying shares solely because the government is involved, then once the government prepares to exit, there will be no real buyers in the market. Moreover, when politicians—such as Ro Khanna, Nancy Pelosi, and others—begin selling their shares, all investors will rush to sell before the government does, causing the paper gains supposedly meant to “offset national debt” to vanish instantly. These gains won’t just turn into actual losses; they will further increase government debt.

Worse still, the Treasury will still need to repay the money it originally borrowed from the Federal Reserve in the future. Thus, for this SPV, this is effectively an investment that can only be bought, not sold. The Federal Reserve can only continuously roll over the SPV’s loans to ensure that a margin call is never triggered.

Ultimately, to sustain the entire system, the expansion of the Federal Reserve’s balance sheet will become permanent.

However, this is not an issue Trump needs to worry about, as politically, he has gained on both fronts: on one hand, U.S. AI companies that have yet to turn a profit have received funding to continue competing with China; on the other hand, the paper "wealth" created by AI has boosted tax revenues and stimulated current economic activity.

Meanwhile, unrealized gains combined with new tax revenues create the illusion that the U.S. debt-to-GDP ratio is declining, causing markets to continue lending to the U.S. government at lower interest rates.

The U.S. government could easily do this right now, preventing the AI bubble from bursting in advance. Of course, it could also wait until the growth rate of AI CAPEX slows down and the market begins to sell off AI stocks en masse before stepping in to rescue it.

Since the U.S. government has already begun directly purchasing corporate equity, why not buy more? By combining "bank window guidance-style lending" with "government direct purchases of AI stocks," it would theoretically ensure that an AI credit crisis never occurs—at least not before the 2028 U.S. presidential election.

Some might wonder, given all this, if so much money has been printed since 2022, why hasn't Bitcoin broken $126,000 yet? Don’t worry—the next section will provide the answer.

The bottom of the previous cycle occurred after the market discovered that the white boy Sam Bankman-Fried had stolen client funds from FTX, and CZ helped facilitate this discovery.

Meanwhile, ChatGPT was commercially launched, marking the beginning of the AI wave.

Starting in October 2023, the U.S. liquidity environment shifted as funds continuously flowed out of the overnight reverse repurchase agreement (RRP) facility, leading to an increase in dollar liquidity. Subsequently, bank credit and government borrowing also began to rise. Bitcoin rose accordingly and peaked in October 2025; however, it did not continue to climb further, increasing by only about double its previous all-time high, as AI-related credit and AI stocks absorbed the newly added fiat liquidity.

As AI capital expenditures (CAPEX) expanded rapidly, consuming all available fiat liquidity, Bitcoin — which in hindsight was obvious — dropped by 50%.

In mid-2026, the liquidity environment reverses. The rate of announced AI capital expenditures over the next 18 months will begin to slow, while bank and government credit channels are only just beginning to create dollars and channel them into the AI industry.

If banks fail to fulfill their "patriotic duty" by continuing to provide credit, the government will strongly push them to lend to AI. If they still fail, the government will reduce the risk for banks lending to the AI industry by providing equity support to specific AI companies and securing offtake agreements similar to those made by Intel and IBM.

Bitcoin will bottom out in the early stages of this credit mismatch, and the financialization of AI — where the amount of U.S. dollars and RMB chasing high-quality AI projects exceeds the actual scale of truly quality projects — will ultimately lead to capital misallocation.

As I wrote this article in late July 2026, I did not know at what price Bitcoin would ultimately bottom; the bottom may already have occurred.

The market needs time to digest concerns stemming from Strategy (formerly MicroStrategy) selling its Bitcoin. At the same time, the market needs to find a new narrative: if Strategy can no longer issue shares or find investors to purchase its preferred stock and use those funds to buy more Bitcoin, why should Bitcoin continue to rise?

Bitcoin may oscillate between $60,000 and $70,000 for a period and could even drop to $50,000. However, during this time, AI capital waste will continue to accelerate, laying the foundation for Bitcoin to find support and subsequently rise gradually.

If my view is correct—that the scale of truly valuable AI capital expenditure projects is smaller than the credit flowing into "AI"—then Bitcoin's price will eventually reflect this excess liquidity. This will help Bitcoin find a floor, even if digital asset treasury companies like Strategy can no longer purchase Bitcoin in a Bitcoin-per-share accretive manner through equity and corporate bond markets.

I will continue to monitor several indicators to verify this logic:

• Is AI capital expenditure growth slowing down?
• Has the AI loan volume increased?
• Are hyperscalers increasing off-balance-sheet commitments?

If we enter the capital waste phase of this AI credit boom, the next question is: What will regulators and governments do? Will they print money in advance, or will they wait for a final crisis to erupt, due to a lack of political space, and then intervene with bailout packages?

Fortunately, as long as we hold Bitcoin without leverage, we don’t care when the bailout comes, because we know that, due to distorted government incentives, they will ultimately choose to print money to save the system. The AI capital expenditure credit boom is now at a scale equivalent to the railroad construction era’s share of GDP, meaning the scale of capital misallocation has surpassed that of the U.S. subprime mortgage crisis.

Therefore, the scale of future bailouts will exceed the trillions of dollars printed by the Federal Reserve and major central banks worldwide between 2009 and 2013. Bitcoin was born precisely as a response to the "irresponsible bailouts of bankers" during the subprime crisis. If you think about it carefully, this is an astonishing thing. And this time, Bitcoin already exists—it may fulfill the dreams of many by rising to $1 million or even higher.

Given the current bleak state of the crypto capital markets, it’s not easy to imagine such a future. But in my view, this creates interesting asymmetric opportunities. Maelstrom has been holding a significant amount of Bitcoin for the long term.

Besides Bitcoin, what new narrative could drive a large-cap token upward over the next six months? Ethereum is currently the most hated and forgotten large-cap "altcoin" in the market. It hasn’t even broken its historical high of $5,000 from 2021, while most of the top ten market-cap altcoins already have.

In my view, the next narrative is enterprise RWA (real-world asset) chains, similar to Robinhood, that will use customizable Ethereum Layer 2s like Arbitrum. Ethereum will serve as the securities settlement layer for these chains. Therefore, even if the actual gas fees flowing to Ethereum represent only a small portion of the overall system, ETH remains the coin powering the tokenization of everything.

I am a critic of RWA. Maelstrom frequently receives a flood of spam project pitches, where teams claim they’re riding the wave of asset tokenization, while on the other hand, TradFi enthusiastically discusses: “All assets will be tokenized and run on some private or public blockchain in the future.”

I strongly believe that if this future comes to pass, these TradFi RWA projects must run on public blockchains. By launching its own chain on Arbitrum, Robinhood reduces the professional risk for TradFi practitioners—they can replicate the same model and ultimately build Ethereum-based solutions.

This narrative is very strong. ETH, as a meme coin, is the second-largest crypto asset by market cap and has existed since 2015, giving it the second-strongest Lindy effect after Bitcoin (the longer something has existed, the higher the probability it will continue to exist). Added to this, Tom Lee from Bitmine has endorsed ETH allocations for institutional investors, enabling fund managers to bet on the tokenization of capital markets.

My rough target price for ETH by the end of 2026 is $5,000, which is approximately a 2.6x increase from the current price. I like this trade because I can take on a substantial notional position while being comfortable with the very low risk that ETH could drop 75% in a single day due to some technical vulnerability.

In addition, ETH is highly liquid, so even if it makes up a large portion of my Maelstrom portfolio, I can still exit within minutes. Finally, I also sell out-of-the-money puts to generate additional income, while accepting the risk of buying ETH at a discount if its price falls below the strike price.

The AI bubble once drained liquidity from the crypto market, but that phase has ended. As the market narrative shifts from “invest in AI no matter the cost” to “what is my return on investment?” and eventually to “when will I get my principal back?”… governments that have bet their entire economic policy on AI will begin to worry—perhaps this bubble really could burst.

To avoid acknowledging their mistakes and prevent this outcome, they will engage in massive capital misallocation, which will ultimately create a cryptocurrency bull market unlike any we’ve seen since 2021.