AI infrastructure is going crazy, but data centers are also burning through money quickly, and the debt bubble is starting to burst.

Article by: Long Yue

The speed of AI arms race spending has already outpaced the rate at which these tech giants are making money themselves.

Over the past week, Amazon, Google, Meta, Microsoft, and Oracle—five massive tech companies (with a combined market value of over $12 trillion)—all released their Q1 earnings reports, unanimously raising their capital expenditure plans.

Morgan Stanley’s Chief Cross-Asset Strategist Andrew Sheets immediately updated his forecast, raising the five tech giants’ combined capital expenditure prediction for 2026 to $800 billion, and further increasing it to over $1.1 trillion in 2027.

Sheets wrote:

We forecast that the capital expenditure of these mega-tech companies will be about $800 billion in 2026, nearly double the spending in 2025 and three times that of 2024. Next year, my colleagues estimate that U.S. mega-tech companies’ capital expenditure could reach $1.1 trillion.

The numbers are impressive enough, but questions also arise.

Where does the money come from? The answer is borrowing.

In recent years, these tech giants have accumulated large amounts of free cash flow through a “light asset” business model. But now, the situation has quietly reversed.

Data on Amazon and Meta’s free cash flow shows that both companies are approaching or have already fallen into negative territory.

What does this mean? Simply put: there’s not enough money, so they have to borrow.

Especially since these companies still need to maintain stock buybacks and dividends, additional capital expenditures can almost only be supported by issuing debt.

Morgan Stanley predicts that 2026 will be the busiest year in history for the U.S. investment-grade (IG) bond market:

Total issuance of about $2.25 trillion, up 25% year-over-year.

Net supply of about $1 trillion, up 57% year-over-year.

Among these, the tech industry has already contributed 18% of U.S. investment-grade bond supply this year—this is the highest proportion ever for the industry, double the same period in 2025.

The core logic driving all this, Morgan Stanley summarizes in four words: “AI Capex Driving Supply.”

The bond market is showing signs of fatigue.

The market has definitely felt it.

After experiencing a $300 billion AI debt frenzy, investors are beginning to show signs of fatigue.

The most straightforward example: Meta issued a $25 billion investment-grade bond last week, with peak order books around $96 billion. This number seems large, but compared to the $125 billion demand attracted by the same issuer’s $30 billion bond in October last year, it’s clearly shrunk.

More noteworthy details:

An issuer related to SoftBank Group had to raise its yield to complete financing due to insufficient demand.

Investors are beginning to demand stronger protective clauses—including a “backstop” provided by Google’s parent company Alphabet, which guarantees rent payments for data centers in case of tenant default.

Some investors have outright refused certain deals. One investor told Bloomberg they abandoned a $14 billion Oracle data center bond in Michigan because the bond included call provisions that were unfavorable to bondholders.

Robert Tipp, Head of Global Bond Funds at PGIM Fixed Income, said:

Ultimately, these companies are selling large amounts of debt, and they will have to pay a higher price to borrow. After corporate spreads narrowed sharply to historic lows, the market is facing a wall of concern.

John Servidea, Co-Head of Investment Grade Debt Capital Markets at J.P. Morgan, added:

We are seeing different investors value these financings differently, assessing risk and return in various ways. Demand for these deals remains quite strong, but as supply increases, we expect the terms and structures of transactions to continue evolving.

Banks are “struggling to hold on.”

The fatigue in the bond market is just the tip of the iceberg. Deeper pressures are building within the banking system.

According to the Financial Times on May 3, major lenders such as J.P. Morgan, Morgan Stanley, and Sumitomo Mitsui Banking Corporation (SMBC) are actively seeking ways to disperse the risks associated with data center-related debt to a broader range of investors, freeing up balance sheet space.

Matthew Moniot, Co-Head of Credit Risk Sharing at Man Group, bluntly stated:

The scale we’re talking about… far exceeds anything we’ve imagined before. Banks will soon be overwhelmed.

A concrete example illustrates the severity: J.P. Morgan and MUFG spent over six months trying to distribute $38 billion in construction debt related to Oracle’s Texas and Wisconsin data center projects. The result was—insufficient demand, forcing some banks to sell at a discount, offloading these loans to non-bank lenders.

$38 billion, a single project, unsold after six months.

Behind this is the strict internal risk limits of banks—once exposure to a single borrower or industry hits the cap, banks cannot finance new projects.

Moniot said:

If I were a bank’s Chief Risk Officer, facing billion-dollar credit requests for individual projects, I would ask how they plan to distribute these risks.

To address this, banks are exploring “Significant Risk Transfer” (SRT) tools—cutting high-concentration data center loans into parts, transferring the riskiest portions off-balance sheet, and selling them to private credit funds, insurance companies, and other investors.

David Lucking, a partner at law firm Linklaters, said:

Banks usually still retain a certain percentage of exposure. SRT investors want to see that the bank still has some stake.

Frank Benhamou, Portfolio Manager at Cheyne Capital, pointed out that data center-related SRT differs fundamentally from traditional products:

The number of operators is limited, highly concentrated, and there are significant construction risks. Naturally, you demand higher returns for this.

Goldman Sachs warns: Investment-grade bond market is becoming “stock-like.”

This wave of AI debt is also changing the structure of the entire investment-grade bond market.

Amanda Lynam, Strategist at Goldman Sachs Investment Grade Bond Strategy, noted that since 2026, U.S. investment-grade bond issuance has had its strongest start on record—by April 20, issuance reached $794 billion, roughly in line with Morgan Stanley’s full-year forecast of $2.25 trillion.

But what’s more noteworthy is the structural change.

Jeffrey Papai, a trader in Goldman Sachs’ investment-grade bonds, wrote in a recent report that among the 660 issuers that issued investment-grade bonds over the past year, only 11 contributed about 25% of the adjusted issuance after duration. Among these, four mega-tech companies (Meta, Amazon, Oracle, Google) plus four large data center financings accounted for nearly 20% of the total duration-weighted issuance.

To put it in perspective: Oracle (ORCL) is now the largest single issuer in the investment-grade index after risk adjustment; Meta, in less than a year, jumped from the 51st to the 8th largest issuer.

Last week, Meta issued a $25 billion bond, the largest single data center financing transaction (RPLDCI), with a duration-weighted size approaching the total of Boeing’s (BA) outstanding bonds, even surpassing Ford’s (F) or General Motors’ (GM) total outstanding bonds.

Goldman Sachs thus issued a warning:

“We are now facing a market increasingly concentrated on AI infrastructure, similar to the stock market, but with a more negative convexity—because fixed income has no upside space.”

In other words: betting on AI in stocks can lead to gains; betting on AI in bonds earns interest, but if problems arise, losses are real.

When funds are insufficient, go borrow globally.

Faced with the capacity limits of the U.S. investment-grade bond market (typically no more than 2-3% per issuer), tech giants have begun seeking financing in global markets.

Goldman Sachs data shows that since 2024, issuance of bonds denominated in euros, pounds, and Swiss francs by mega-tech companies has risen significantly.

At the time of writing, Alphabet, Google’s parent company, just launched a euro bond issuance of at least €9 billion, and also started a Canadian dollar bond sale—both setting new records in their respective markets.

Meta has taken a different approach: setting up off-balance-sheet special purpose vehicles (SPVs) to share debt burdens. After partnering with Blue Owl last year to complete a $27 billion “Beignet Project” financing for Louisiana data centers, Meta is now working with Morgan Stanley and J.P. Morgan on a $13 billion “Sopaipilla Project” to finance its Texas El Paso data center.

The essence of this structure is to disperse debt as widely as possible among more parties.

Morgan Stanley: When will the AI bubble burst? Look for four signals.

As the entire AI supercycle increasingly depends on smooth debt markets, Morgan Stanley has listed four warning signs that could trigger a surge in credit spreads and cause the “house of cards” of AI to collapse:

Debt growth outpacing profit growth.

Leverage financing market expanding faster than high-quality credit markets.

M&A activity exceeding long-term trend levels.

Accelerating private equity-backed deals and declining equity contribution ratios.

Another more direct market signal worth noting: last week, on the same day the U.S. stock market hit record highs and several “Big Seven” tech stocks soared, Meta’s credit default swap (CDS) spreads hit a record high and widened daily.

Stock prices hitting new highs while CDS spreads also reach new highs—this simultaneous occurrence is itself a warning signal.

Morgan Stanley’s final, concise conclusion:

The credit market is providing financing for AI infrastructure.

In other words: once the credit market shuts down, the AI supercycle will come to an end.

This article does not constitute personal investment advice, does not represent platform views, markets are risky, invest cautiously, and make independent judgments and decisions.