This may be the strangest scene in the tech industry in more than two decades.

Microsoft, Google, Meta, and Amazon are among the most profitable companies with the strongest cash flows in human business history. They year-round (chángnián, 'constantly') have tens of billions of dollars in cash on their books, which in the past was either used for stock buybacks, dividend payouts, or even acquiring companies left and right.

But now, these companies, which are least in need of money, are also lining up to borrow.

In June this year, Google announced a new financing plan worth up to $80 billion, setting a record in U.S. stock market history. Around the same time, Nvidia also borrowed $25 billion in one go, marking its return to the bond market for the first time in nearly five years.

The reason is simple: AI is incredibly cash-intensive.

This year, capital expenditures by the world's top five cloud providers are expected to reach $814 billion, up about 80% year-on-year.

The operational cash flow they generate is being rapidly consumed. According to Morgan Stanley estimates, the free cash flow of Amazon, Oracle, and Meta will approach zero or even turn negative this year.

The visible borrowing is just the tip of the iceberg.

According to Morgan Stanley, by May 2026, Google, Microsoft, Amazon, Meta, Oracle, and Nvidia will collectively shoulder nearly $1.8 trillion in off-balance-sheet liabilities.

What does $1.8 trillion mean? Last year, the U.S. government's total fiscal revenue was $5.24 trillion, meaning this debt is equivalent to one-third of the federal government's annual revenue.

More importantly, this astronomical debt is entirely hidden off the balance sheet. In other words, the risks are concealed where you can't see them.

Today, we'll follow Morgan Stanley's report to comb through all the visible and hidden debts of these tech giants.

/ 01 / $1.8 Trillion in Liabilities Hidden Beyond Financial Statements

Morgan Stanley's analysis reveals that this $1.8 trillion bomb primarily consists of two parts:

First, $982 billion in 'hardware procurement commitments.'

To secure scarce chip production capacity, tech giants are frantically placing 'long-term locked-volume orders' upstream. Contracts are signed, but chips haven't been delivered yet. Under accounting rules, these are called 'procurement commitments,' mentioned only briefly in the footnotes of financial reports and invisible on the main balance sheet.

The problem is, these are rigid obligations that must be paid. Currently, Google alone carries $332 billion, Meta $238 billion, and Microsoft $142 billion.

As the 'chief dealer' of the entire supply chain, Nvidia not only receives hundred-billion-dollar long-term orders from cloud providers but also locks in production capacity with upstream TSMC. Its own procurement commitments reach $155 billion. Its inventory plus long-term procurement obligations already account for 32% of its expected revenue for fiscal 2027.

Historically, this ratio has typically hovered around 20%—now it's 12 percentage points higher.

What does that mean? Suppose a restaurant expects to sell $1 million worth of dishes next year. Normally, it might pre-order $200,000 worth of ingredients. But Nvidia is now pre-ordering $320,000 worth of ingredients before even selling the dishes.

If customer demand continues to grow, that's fine. But once demand slows, these inventories and procurement costs will quickly pile up.

Second, $822 billion in 'long-term leasing arrangements.'

AI development requires massive computing power and data centers. Cloud providers, to save time, have signed numerous leasing contracts in advance.

However, under regulations, leased spaces not yet in use don't count as current liabilities. This has created substantial off-balance-sheet debt.

By the first quarter of 2026, the scale of unactivated data center leases among the top five cloud providers had climbed to $822 billion, including $261 billion for Oracle, $197 billion for Microsoft, and $183 billion for Meta.

In other words, the leasing liabilities visible on the books don't even represent a fraction of what they actually owe.

Morgan Stanley calculated: If these unactivated leasing commitments are included, Microsoft's capital expenditures as a percentage of revenue would surge from 33% to 44%, while Oracle's would jump from 76% to 189%.

What does 189% mean? For every $1 you earn, you'd need to invest nearly $2 to build data centers. Any traditional industry would see investors walk out at such ratios.

/ 02 / Amazon and Meta Are Borrowing to Survive

Beyond off-balance-sheet financing, visible on-balance-sheet financing is also looking grim. Morgan Stanley's latest report predicts that the top five cloud providers (Amazon, Google, Microsoft, Meta, Oracle) will collectively pour $814 billion into capital expenditures this year, up more than 80% year-on-year.

The cash generated from operations can't cover the computing power spending hole.

According to Morgan Stanley estimates, the free cash flow of Amazon, Oracle, and Meta will approach zero or even turn negative this year.

Google isn't faring much better. By 2026, its free cash flow is expected to shrink by 69%. Among the top five cloud providers, only Microsoft maintains positive cash flow growth thanks to its diversified business.

This means every dollar Amazon, Oracle, and Meta spend going forward will need to be borrowed.

For companies of Amazon and Meta's scale to rely on bond issuance for daily capital expenditures would have been almost unthinkable in the past.

Starting this year, the debt snowball from AI has begun rolling wildly.

Morgan Stanley's tracking data shows that as of May 2026, the cumulative issuance of global AI-related corporate bonds year-to-date has reached $236.3 billion, 3.57 times that of the same period last year. Full-year issuance is expected to exceed $570 billion, surging 162% from 2025.

Breaking it down, investment-grade dollar bond issuance by top cloud providers reached $82 billion in the first five months, up 1,540% year-on-year. Even data center operators and computing service providers with relatively weaker credit profiles issued $29.7 billion, up 1,384% year-on-year.

Morgan Stanley noted an interesting detail: As of May, there were 12 high-yield bonds related to data center construction in the public market, compared to just 4 for all of 2025—nearly tripling in less than six months.

The result of dense (mìjí, 'intense/frequent') bond issuance is that it has completely altered the 'asset-light' nature of tech companies.

The gross profit leverage (representing how much capital investment is needed for every $1 of gross profit) of hyperscale cloud enterprises surged from 0.9x to 1.8x in just two quarters—doubling.

This level now exceeds that of the energy sector, where oil giants typically have gross profit leverage of 0.2–0.5x, and electric utilities around 0.6–1.2x.

On the surface, this is just a numerical change. But it means these tech giants' financial structures are shifting from 'asset-light, high-profit, low-leverage' good businesses to 'asset-heavy, high-investment, high-debt' traditional infrastructure models.

/ 03 / Phantom Leverage Through Circular Financing

If off-balance-sheet debts can at least be traced in financial report footnotes, what comes next is 'phantom leverage' that's invisible even there.

This tactic is called 'SPV financing' (Special Purpose Vehicle).

Simply put, instead of borrowing directly, a tech giant sets up an 'independent shell company.' This shell issues bonds and buys assets, with all debt recorded under the shell's name, keeping the parent company's balance sheet clean.

Meta is a master of this game. For its August 2025 data center financing, Meta partnered with asset manager Blue Owl to create an SPV. Blue Owl contributed the majority, while Meta contributed a minority stake. Through this shell, they issued $27 billion in bonds, drawing in long-term institutional funds like BlackRock.

The result? Meta's balance sheet only showed a small equity contribution, while the $27 billion in debt vanished like a ghost.

This 'off-balance-sheet financing' tactic is now being replicated crazy (fēngkuáng, 'wildly/rampantly') across the computing power industry.

In June 2026, Apollo and Blackstone teamed up with Broadcom to launch the AI XPV financing platform, with an initial $35 billion investment directed at Anthropic's computing infrastructure.

This money was raised from the market by private credit institutions, secured by chip assets delivered by Broadcom, with Google—as Anthropic's core shareholder—providing implicit credit support, while investment banks offered matching (pèitào, 'supporting/ matching ') credit facilities.

The entire capital flow completed a full circle: flowing out from investors, passing through AI companies, chip manufacturers, and investment banks, before returning to the financial system.

Every participant in the chain offloaded risk one layer further through contracts, guarantees, or equity relationships, yet no one truly absorbed the underlying computing power's revenue risk.

If we must find someone paying for this cycle, it's retail investors. Morgan Stanley revealed a more conceal (yǐnbì, 'hidden/concealed') detail: Insurance companies are indirectly injecting funds into this market through annuity products.

What does this mean? The ultimate bag-holder for this trillion-dollar leverage chain is likely ordinary investors who bought annuity products expecting a peaceful retirement. The parallels to the 2008 subprime mortgage crisis are striking.

/ 04 / Conclusion

While AI financing seems crazy (fēngkuáng, 'wild/unrestrained') and shows no signs of slowing, two factors could potentially detonate this time bomb.

The first factor is massive potential depreciation.

The enormous construction projects will eventually depreciate. Over the past year, Oracle's construction in progress has surged 200%, while Meta's grew 90%. Once these data centers are completed, depreciation will devour profits like a black hole.

Morgan Stanley predicts that over the next three years, cumulative depreciation for Microsoft, Oracle, Meta, and Google will exceed $520 billion. Oracle's depreciation-to-revenue ratio will skyrocket from 4% to 28%, while Meta's will rise from 9% to 19%.

This means for every $100 earned, $28 must be spent on amortizing equipment costs. Profit margins will be severely compressed.

The second factor is revenue-expenditure mismatch.

Simply put, money is being spent faster than it's being earned.

According to previous forecasts, by 2026: Google's capital expenditure forecast has been revised up 139%, Oracle's by 170%. But revenue forecast revisions are 'nowhere near those levels.'

In Morgan Stanley's words: 'The gap between capital expenditure growth and revenue growth is widening.'

This means these cloud providers are burning cash at an accelerating pace, but revenue growth isn't keeping up proportionally. While financing can plug gaps short-term, if ROI doesn't recover long-term, this will evolve from a liquidity crisis into a solvency crisis.

And most of this isn't reflected in those glossy financial reports.

Now, a gray rhino is already lurking at Silicon Valley's doorstep. The only question is: When will it charge?

By Yuan Yuan