Is AI infrastructure a bubble, or is it a "collective buy for time"? Analyzing the financial structure behind the $30 trillion.

This is not a simple binary debate of “bubble vs non-bubble”; the answer may be more complex and nuanced than you think. I do not have a crystal ball to predict the future. However, I try to delve into the underlying financial structure of this feast and construct a set of analytical frameworks.

The article is long and detailed, so let's start with the conclusion:

In terms of direction, I don't think this is a big bubble. However, there are high risks in certain segments.

To be more precise, the current AI infrastructure resembles a long march of “banding together + buying time.” Major companies (( Microsoft, Google, Meta, Nvidia, etc.) leverage financial engineering to unlock massive leverage, but outsource the main credit risk to special purpose vehicles (SPVs) and the capital market, tightly binding the interests of all participants.

The so-called “buying time” refers to betting on their own cash flow and the patience of external resources, and whether they can hold on until the day when “AI truly enhances productivity” arrives.

If the bet wins, the AI will fulfill its promise, and the big companies will be the biggest winners. If the bet loses (AI progress does not meet expectations or costs are too high), the first to be hurt are the external resources providing financing.

This is not the kind of “bank leverage excess, single point detonation” bubble from 2008. This is a giant experiment in direct financing, led by the smartest and most cash-rich entrepreneurs on the planet, using complex “off-balance-sheet financing” strategies to break risk into many tradable fragments, distributing them for different investors to digest.

Even if it's not a bubble, it doesn't mean that all AI infrastructure investments can achieve a good ROI.

01 Understanding the Core: The Benefit Binding Mechanism of “Banding Together”

The so-called “banding together” refers to the close binding of the interests of the five parties in this AI infrastructure.

Tech giants (Meta, Microsoft, Google) and their large model partners (OpenAI, xAI): need computing power but don't want to spend a large amount of money at once.

Chip supplier (Nvidia): Needs a continuous stream of large orders to support its valuation.

Private equity funds (Blackstone, Blue Owl, Apollo): Need new asset classes to expand asset management scale and collect more management fees.

Neocloud (CoreWeave, Nebius) and hybrid cloud service providers (Oracle Cloud Infrastructure): provide infrastructure and computing power, but at the same time require large companies to sign long-term contracts to leverage financing.

Institutional investors (pension funds, sovereign wealth funds, traditional funds like BlackRock): require stable returns higher than government bonds.

And these five parties form a “community of shared interests”, for example:

Nvidia prioritizes supply for CoreWeave while investing in its equity.

Microsoft gives CoreWeave a long-term agreement while assisting with its financing.

Blackstone provides debt financing while raising funds from pension funds.

Meta and Blue Owl jointly established an SPV to share risks.

OpenAI and other large model vendors continue to raise the standards for model parameters, inference capabilities, and training scales, effectively increasing the overall industry's computational power demand threshold. Particularly under its deep integration with Microsoft, this “technology outsourcing, pressure internalization” collaboration structure allows OpenAI to become a catalyst for accelerating the global capital expenditure competition without spending money. It is not a funding party, but rather the actual curator driving the leverage increase across the board.

No one can be self-sufficient, and that is the essence of “banding together.”

02 Capital Structure - Who is funding? Where is the money flowing?

To understand the overall architecture, we can start with the flowchart of funds below.

Tech giants need astronomical computing power, and there are two paths:

Self-built data center: This is the traditional model. The advantage is complete control, while the drawback is slow construction, and all capital expenditures and risks are placed on one's own balance sheet.

Seeking external supply: The giants are not simply renting servers, but have given rise to two core “external supplier” models. This is the new trend of the moment and is the focus of our analysis.

The first type is SPV (Special Purpose Vehicle) / Special Purpose Entity, which is a purely financial instrument. You can think of it as a special entity established specifically for “a single project, a single client.”

Business model: For example, if Meta wants to build a data center but does not want to spend a large sum of money at once, it can collaborate with an asset management company to form an SPV. The sole task of the SPV is to construct and operate this center exclusively for Meta. Investors receive high-quality debt instruments (a mix of corporate bonds and project financing) backed by rental cash flow.

Customer Type: Extremely Singular, usually only one (e.g. Meta).

Risk level: Life and death completely depend on a single customer's credit.

The second type is Neocloud (, such as CoreWeave, Lambda, Nebius ), which are independent operating companies (Operating Company, OpCo) with their own operating strategies and full decision-making authority.

Business model: For example, CoreWeave raises funds (equity and debt) to purchase a large number of GPUs and rents them out to multiple clients, signing “minimum guarantee/reservation” contracts. It is flexible but the equity value fluctuates greatly.

Customer Type: Theoretically diverse, but in practice, it is highly dependent on large companies (for example, Microsoft's early support for CoreWeave). Due to its smaller scale, unlike SPV relying on a single wealthy backer, Neocloud has a higher dependence on upstream suppliers (Nvidia).

Risk level: Risk is distributed among multiple clients, but operational capabilities, technology, and equity value all affect survival.

Although fundamentally different in terms of legal and operational structures, the commercial essence of both is ultimately the same: they are both “external suppliers of computing power” for the giants, effectively “offloading” the massive GPU procurement and data center construction from the giants' balance sheets.

So where does the money for these SPVs and Neoclouds come from?

The answer is not traditional banks, but private credit funds. Why?

This is because after 2008, the Basel III Accord imposed stringent requirements on banks' capital adequacy ratios. Banks taking on such high-risk, high-concentration, long-term massive loans need to set aside reserves that are excessively high relative to their costs.

The businesses that banks “cannot do” or “dare not do” have created a huge vacuum. Private equity giants like Apollo, Blue Owl, and Blackstone have filled the gap—they are not subject to banking regulations and can offer more flexible, faster financing, but with higher interest rates. Guaranteed by project rent or GPU/equipment with long-term contracts.

For them, this is an extremely attractive pie - many have traditional infrastructure financing experience, and this theme is enough to grow the scale of asset management several times, significantly increasing management fees and carried interest (.

So where does the money for these private credit funds ultimately come from?

The answer is institutional investors (LPs), such as pension funds, sovereign wealth funds, insurance companies, and even general investors (for example, through the private credit ETF issued by BlackRock - which includes the 144A private bond under the Meta project Beignet Investor LLC 144A 6.581% 05/30/2049).

The transmission path of the risk chain is thus established:

) ultimate risk bearer ( pension fund/ETF investors/sovereign funds → ) intermediary institutions ( private credit funds → ) financing entities ( SPV or Neocloud ) such as CoreWeave ( → (ultimate users) tech giants ) such as Meta (

03 SPV Case Study — Meta's Hyperion

To understand the SPV model, Meta's “Hyperion” project is an excellent case (with enough public information):

Structure/Equity: Meta and Blue Owl manage a fund group JV (Beignet Investor LLC). Meta holds 20% equity, Blue Owl holds 80%. Bonds issued under SPV 144A structure. The JV owns the assets, and Meta leases them under a long-term agreement. Capital expenditures during the construction period are in the JV, and assets gradually transfer to Meta's balance sheet after financing leases begin.

Scale: Approximately $27.3 billion in debt (144A private placement bonds) + approximately $2.5 billion in equity, making it one of the largest single corporate bond/private credit financing projects in U.S. history. The maturity date is in 2049, and this long-term amortization structure essentially locks in the most challenging time risk first.

Interest Rate/Rating: The debt has been rated S&P A+ (high rating allows insurers to allocate), with a coupon rate of approximately 6.58%.

Investor composition: PIMCO subscribed to 18 billion; BlackRock's ETFs totaled over 3 billion. For this group of investors, this represents a highly attractive high-quality stable yield.

Cash Flow and Leases: Blue Owl is not focused on the potentially depreciating GPUs (I believe some people in the market are concerned that the assumption of a long depreciation period for GPUs is misplaced because GPUs are just hardware, while the overall value of AI lies in hardware + models. The price of older hardware decreases due to iteration, but that does not mean the value of the final AI model application decreases as well). Instead, they are focused on the SPV cash flow supported by Meta's long-term leases (starting from 2029). The construction period funding is also allocated to U.S. Treasury bonds to reduce risk. This structure integrates the liquidity of corporate bonds with project financing protection clauses, and it is also 144A-for-life (restricted to a limited circle of investors).

So why is the short-term risk of this architecture extremely low?

This is because under this structure, the Hyperion task is straightforward: the left hand collects Meta rent, while the right hand pays Blue Owl interest. As long as Meta does not collapse (the likelihood in the foreseeable future is extremely low), the cash flow is as stable as a rock. There is no need to worry about fluctuations in AI demand or GPU price drops.

This 25-year ultra-long maturity, rent amortization debt structure locks in all recent refinancing risks as long as rental income is stable and interest payments are normal. This is the essence of “buying time” (allowing the value created by AI applications to gradually catch up with the financial structure).

At the same time, Meta uses its own credit and strong cash flow to secure substantial long-term financing that bypasses traditional capital expenditures. Although under modern accounting standards (IFRS 16), long-term leases ultimately appear on the balance sheet as “lease liabilities”, the advantage is that the pressure of capital expenditures amounting to billions of dollars during the initial construction phase and the associated construction risks and financing operations are first transferred to the SPV.

Transform one-time massive capital expenditures into lease payments amortized over the next 25 years, greatly optimizing cash flow. Then gamble on whether these AI investments can generate enough economic benefits in 10-20 years to pay back the principal and interest (considering a bond with a coupon rate of 6.58%, the ROI calculated based on EBITDA must be at least 9-10% to provide equity holders with a decent return rate).

04 Neocloud's Buffer Pad — OpCo's Equity Risk

If the SPV model is “credit transfer”, then CoreWeave and Nebius, which are Neocloud models, represent “further layering of risk”.

Taking CoreWeave as an example, the capital structure is much more complex than that of an SPV. Multiple rounds of equity and debt financing involve investors such as Nvidia, VC, growth funds, and private debt funds, creating a clear sequence of risk buffers.

If AI demand falls short of expectations, or new competitors emerge, and CoreWeave's revenue plummets and cannot pay high interest, what will happen:

The first step is the evaporation of equity value: CoreWeave's stock price plummets. This is the “equity buffer” - the first to absorb the shock. The company may be forced to finance at a discount, significantly diluting the equity of existing shareholders, or even resulting in total loss. In contrast, the equity buffer of the SPV is thinner, as it cannot directly raise funds in the public market.

The second step is that creditors suffer losses: only after the equity is completely “burned out” will it be Blackstone and other private debt creditors' turn to bear the losses, if CoreWeave still cannot repay its debts. However, these funds typically require excellent collateral (latest GPUs) and strict repayment priority when lending.

CoreWeave and Nebius both adopt the strategy of 'first securing long-term contracts, then financing those contracts', rapidly expanding through refinancing in the capital markets. The brilliance of this structure lies in the fact that large clients can achieve better capital utilization efficiency, leveraging future procurement contracts to stimulate more capital expenditure without contributing funds, and the likelihood of risk contagion to the entire financial system is limited.

On the contrary, Neocloud shareholders need to be aware that they are sitting in the most turbulent yet exhilarating position in this gambling game. They are betting on rapid growth while also praying for the management's financial operations (debt extensions, equity issuance) to be nearly flawless, and they must pay attention to the debt maturity structure, pledge range, contract renewal windows, and customer concentration in order to better assess the risk-reward ratio of equity.

We can also speculate about who would be the marginal capacity most easily abandoned if the demand for AI really grows slowly. SPV or Neocloud? Why?

05 Oracle Cloud: The Counterattack of an Atypical Cloud Player

While everyone is focused on CoreWeave and the three major cloud giants, an unexpected “dark horse” in the cloud sector is quietly rising: Oracle Cloud.

It does not belong to Neocloud, nor is it part of the first tier of the three major tech giants, but it has secured contracts for a portion of the computing load from Cohere, xAI, and even OpenAI, thanks to its highly flexible architecture design and close collaboration with Nvidia.

Especially when Neocloud's leverage is gradually tightening and traditional cloud space is insufficient, Oracle, with its positioning of “neutral” and “substitutable,” has become an important buffer layer in the second wave of AI computing power supply chain.

Its existence also shows us that this battle for computing power is not just a showdown among the three giants; there are also atypical yet strategically significant suppliers like Oracle quietly vying for position.

But don't forget, the table of this game is not just in Silicon Valley, but extends to the entire global financial market.

The government “implicit guarantee” coveted by many.

Finally, in this game dominated by tech giants and private equity finance, there is a potential “trump card” - the government. Although OpenAI recently stated that it “does not have and does not wish” the government to provide loan guarantees for data centers, the discussions with the government are about potential guarantees for chip factories rather than data centers. However, I believe that their (or similar participants') original plan must have included the option of “bringing the government in to team up.”

How to say? If the scale of AI infrastructure is so large that even private equity cannot bear it, the only way out is to upgrade it to a national power competition. Once the leadership position of AI is defined as “national security” or “the moon landing race of the 21st century,” government intervention becomes a matter of course.

The most effective way to intervene is not to directly provide funds, but to offer “guarantees.” This approach brings a decisive benefit: significantly reducing financing costs.

Investors around my age should still remember Freddie Mac ) and Fannie Mae (. These two “Government Sponsored Enterprises” (GSEs) are not official departments of the U.S. government, but the market generally believes they have “implicit government guarantees.”

They purchase mortgages from banks, package them into MBS and provide guarantees, and after selling them in the open market, redirect the capital back to the mortgage market, increasing the funds available for lending. Their existence also amplified the impact of the financial tsunami in 2008.

Imagine if in the future there were a “National AI Computing Company” backed by the government. The bonds it issues would be regarded as quasi-sovereign debt, with interest rates approaching those of U.S. Treasuries.

This will completely change the previously mentioned “buying time for productivity to rise”:

The cost of financing is extremely low: the lower the borrowing cost, the lower the requirement for the “speed of AI productivity improvement”.

Unlimited extension of time: More importantly, it allows for continuous rolling over at an extremely low cost, equivalent to buying nearly infinite time.

In other words, this practice significantly reduces the likelihood of the bet directly “blowing up”. However, once it does blow up, the impact could expand by several dozen times.

$6 trillion bet — the truly key “productivity”

All the aforementioned financial structures - SPV, Neocloud, private debt - no matter how sophisticated, are merely answering the question of “how to pay.”

The fundamental question regarding whether AI infrastructure will become a bubble is: “Can AI truly increase productivity?” and “How fast is it?”

All financing arrangements lasting 10 or 15 years essentially “buy time.” Financial engineering provides giants with a breathing space, without the need for immediate results. However, buying time comes at a cost: investors in Blue Owl and Blackstone (pension funds, sovereign funds, ETF holders) require stable interest returns, while equity investors in Neocloud seek multiple valuation growth.

The “expected rate of return” of these financing parties is the threshold that AI productivity must cross. If the productivity improvement brought by AI cannot cover the high financing costs, this delicate structure will begin to collapse from its most vulnerable point (“equity buffer”).

Therefore, in the coming years, special attention should be paid to the following two aspects:

The speed of launching “application solutions” in various fields: having a powerful model (LLM) is not enough. We need to see real “software” and “services” that can make companies spend money. This type of application needs to be widely popularized, generating cash flow large enough to repay the principal and interest of the huge infrastructure costs.

External constraints: AI data centers are electricity monsters. Do we have enough power to support the exponentially growing demand for computing power? Is the upgrade speed of the power grid keeping pace? Will the supply of Nvidia's GPUs and other hardware face bottlenecks, causing them to be “slower than” the timelines required by financial contracts? Supply-side risks may drain all the “bought time.”

In short, this is a race between finance (financing costs) and physics (electricity, hardware) and business (application landing).

We can also roughly estimate how much productivity improvement AI needs to bring to avoid a bubble in a quantitative way:

According to Morgan Stanley's estimate, this round of AI investment should accumulate to 3 trillion dollars by 2028.

The aforementioned SPV bond issuance cost of Meta is approximately 6-7%, while according to a report by Fortune, CoreWeave's current average debt interest rate is around 9%. Assuming that most private debt in the industry requires a return of 7-8% with a debt-equity ratio of 3:7, the ROI of these AI infrastructures (calculated based on EBITDA and total capital expenditure) needs to be at 12-13% to achieve an equity return rate of over 20%.

So the required EBITDA = 3 trillion × 12% = 360 billion USD; if calculated at an EBITDA profit margin of 65%, the corresponding revenue is approximately 550 billion USD;

With the GDP of the United States estimated at approximately 29 trillion, an additional output equivalent to about 1.9% of GDP needs to be supported long-term by AI empowerment.

The threshold is not low, but it is not a fairy tale. In 2025, the global cloud industry is projected to have a total revenue of approximately 400 billion dollars. In other words, we need to see at least one or two cloud industries being revitalized through AI empowerment. The key lies in whether the speed of application monetization can be synchronized with the physical bottlenecks.

Risk Scenario Stress Testing: When is there not enough “time”?

All the financial structures mentioned above are betting that productivity can outpace financing costs. Let me use two stress tests to simulate the chain reaction when the speed of AI productivity realization is slower than expected:

In the first scenario, we assume that AI productivity is realized “slowly” (for example, scaling up only after 15 years, but many financings may be on a 10-year term):

Neocloud was the first to collapse: Independent operators like CoreWeave, which operate with high leverage, faced debt defaults or discounted restructurings because their revenues could not cover the high interest, and their “equity cushion” was burnt out.

SPV faces extension risk: When SPV debts like Hyperion mature, Meta must decide whether to refinance at a higher interest rate (the market has already witnessed the failure of Neocloud), eroding core business profits.

Private credit fund LPs have suffered huge losses, and tech stock valuations have been significantly revised down. This is an “expensive failure,” but it will not trigger a systemic collapse.

In the second scenario, we assume that AI productivity has been “falsified” (technical progress stagnates or costs cannot be reduced and scaled):

Tech giants may opt for “strategic default”: this is the worst-case scenario. Giants like Meta may judge that “continuing to pay rent” is a bottomless pit, and thus choose to forcibly terminate leases, forcing SPV debt restructuring.

SPV bond collapse: Bonds like Hyperion, which are considered A+ rated, will instantly decouple from Meta, causing a price crash.

It could completely destroy the private credit “infrastructure financing” market, and there is a high chance that through the aforementioned interconnectedness, it could trigger a confidence crisis in the financial markets.

The purpose of these tests is to transform the vague question of “whether it is a bubble” into specific situational analyses.

07 Risk Thermometer: A Practical Observation Checklist for Investors

As for the changes in market confidence, I will continuously monitor five things as a risk thermometer:

The speed of achieving productivity in AI projects: including the acceleration or deceleration of expected revenue from major model vendors () linear growth or exponential growth), and the application status of different AI products and projects.

Neocloud company's stock price, bond yield, announcements: including large orders, defaults/amendments, debt refinancing (some private bonds are due around 2030 and need special attention), capital increase pace.

SPV bond secondary price/spread: whether 144A private placements like Hyperion maintain above par, whether trading is active, and whether ETF holdings increase.

Changes in the quality of long-term agreements: take-or-pay ratio, minimum retention period, customer concentration, price adjustment mechanism (adjustments for electricity prices/interest rates/inflation pricing).

Power progress and possible technological innovations: As the most likely external factor to become a bottleneck, attention should be paid to the policy signals regarding substations, transmission, distribution, and electricity pricing mechanisms. Additionally, whether there are new technologies that can significantly reduce electricity consumption.

Why isn't this a repeat of 2008?

Some people may draw parallels to the bubble similar to that of 2008. I believe this approach may lead to misjudgment:

The first point is the essential difference in core assets: AI vs. housing.

The core asset of the 2008 subprime mortgage crisis was “housing.” Housing itself does not contribute to productivity (rental income growth is very slow). When housing prices deviate from the fundamentals of residents' income and are packaged into complex financial derivatives, a bubble burst is just a matter of time.

The core asset of AI is “computing power.” Computing power is the “production tool” of the digital age. As long as you believe that AI has a high probability of substantially increasing overall productivity in society (software development, drug research and development, customer service, content creation) at some point in the future, there is no need to worry too much. This is a “prepayment” for future productivity. It has real fundamentals as an anchor point, but it has not yet fully materialized.

The second point is the difference in key nodes of the financial structure: direct financing vs. banks.

The 2008 bubble spread significantly through key nodes (banks). Risk was transmitted through “indirect financing by banks.” The bankruptcy of one bank (like Lehman) triggered a trust crisis in all banks, leading to a freeze in the interbank market and ultimately igniting a systemic financial crisis (including a liquidity crisis) that affected everyone.

Currently, the financing structure for AI infrastructure is primarily based on “direct financing.” If AI productivity is disproven, CoreWeave collapses, and Blackstone defaults on its $7.5 billion debt, it will result in significant losses for Blackstone's investors (pension funds).

The banking system has indeed become stronger since 2008, but we cannot oversimplify and think that risks can be completely “contained” in the private equity market. For example, private credit funds may also use bank leverage to amplify returns. If AI investments fail broadly, these funds could still suffer significant losses that may spill over through two pathways:

Leverage default: The fund's default on leveraged financing from the bank will transmit the risk back to the banking system.

Impact on LPs: Pension funds and insurance companies have experienced significant investment losses leading to deterioration of their balance sheets, triggering them to sell off other assets in the public market, causing a chain reaction.

Therefore, a more accurate statement would be: “This is not the type of interbank liquidity crisis that explodes at a single point and leads to a comprehensive freeze like in 2008.” The worst-case scenario would be “expensive failures,” with lower contagion and slower speed. However, given the opacity of the private equity market, we must remain highly vigilant about this new type of slow contagion risk.

Insight for investors: At which layer of this system are you?

Let's return to the original question: Is AI infrastructure a bubble?

The formation and bursting of bubbles come from the significant gap between expected benefits and actual results. I believe that, in a broad sense, it is not a bubble, but rather a precise high-leverage financial layout. However, from a risk perspective, aside from certain aspects that require special attention, we must also be cautious about the “negative wealth effect” that small-scale bubbles may bring.

For investors, in this multi-trillion dollar AI infrastructure race, you must know what you are betting on when holding different assets:

Tech giant stocks: You're betting that AI productivity can outpace financing costs.

Private credit: you earn stable interest, but bear the risk of “time may not be enough.”

Neocloud equity: You are the first cushion for the highest risk and highest reward.

In this game, position determines everything. Understanding this series of financial structures is the first step to finding your own position. And understanding who is “curating” this show is key to judging when this game will end.