AI infrastructure debt exceeds $236 billion in 2026, as tech giants shift to Wall Street financing.

AI infrastructure is evolving from a technology industry "CapEx cycle" into a financing cycle supported by the entire financial system.

Written by Jim, MSX MaiTong

Edited by: Frank, MSX Maotong

Over the past two years, the most important number in Wall Street’s discussion of AI has been GPUs, data centers, and CapEx.

But by 2026, another number began rapidly entering the market spotlight: debt.

In the past, some of Silicon Valley’s most profitable companies routinely used their cash to purchase GPUs and build data centers. Today, as the scale of AI infrastructure continues to grow toward hundreds of billions—and even trillions—of dollars, even the world’s most prolific cash-generating companies like Alphabet, Amazon, and Meta are increasingly turning to the bond market.

Morgan Stanley estimates that global AI-related debt issuance could approach $570 billion in 2026, more than doubling from last year; as of the end of May, the total had already reached approximately $236 billion, four times the amount from the same period last year.

Meanwhile, Alphabet, Amazon, Microsoft, and Meta are expected to spend approximately $700 billion on related expenses this year, and by 2027, hyperscalers’ capital expenditures could exceed $1 trillion.

And financing methods are also continuing to expand outward.

• On August 10, NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion in third-party capital for AI infrastructure.
• At nearly the same time, the future lease payment commitments disclosed by Microsoft, Meta, Oracle, Amazon, and Alphabet—尚未开始执行—have reached the trillions of dollars;

When the world’s wealthiest companies begin changing their financing methods, it signals that the AI story is entering a new phase. The question arises: why, after holding hundreds of billions of dollars in cash, are these companies suddenly so eager to borrow?

Alphabet is the best example of understanding this change.

From a business perspective, its most recent quarter was nearly strong—Q2 revenue reached $119.8 billion, a 24% year-over-year increase; Google Cloud revenue reached $24.8 billion, a remarkable 82% year-over-year increase.

On the other hand, Alphabet’s quarterly capital expenditures have reached approximately $44.9 billion, causing the company to report negative free cash flow for the first time, despite continued strong business growth—free cash flow in Q2 fell to -$5.9 billion. At the same time, Alphabet has raised its 2026 capital expenditure guidance to $195–205 billion.

This is precisely the biggest difference between AI capital expenditures and the traditional software era.

In the software era, major tech companies are essentially cash machines: after the initial research and development is complete, the marginal cost of adding each new user is limited, allowing a large portion of revenue to ultimately be converted into free cash flow.

In other words, AI is making tech companies heavier again.

GPU, servers, high-speed networks, data centers, power substations, cooling systems, and land all require substantial upfront cash investments before revenue is generated.

Amazon CEO Andy Jassy once explained that data centers often begin incurring construction costs about two years before they are officially launched, while revenue only starts to materialize once the facilities are fully operational.

Thus, a natural maturity mismatch arises: cash is needed today, but income must be recovered gradually over many future years. In this situation, even if a company has substantial cash on its balance sheet, relying entirely on internal cash flow for financing may not be the most rational choice.

In February, Alphabet completed approximately $31.5 billion in global bond financing, including rare 100-year bonds; in August, it issued another $25 billion in investment-grade dollar-denominated bonds. Amazon acted even more aggressively, raising approximately $37 billion in the U.S. bond market in March, followed the next day by a €14.5 billion bond issuance—totaling nearly $54 billion—and then issued another $25 billion in dollar-denominated bonds in July.

Meanwhile, Amazon has increased its 2026 capital expenditure plan to $220 billion, with AWS's latest quarterly revenue growing 37% year-over-year—the fastest pace in over four years—yet its free cash flow over the past 12 months has declined from a positive $18.2 billion a year ago to -$7.6 billion.

Meta has shown the same trend. In April this year, Meta completed a $25 billion bond issuance; revenue in Q2 still grew 28% to $60.8 billion, but free cash flow plummeted from $8.55 billion in the same period last year to just $784 million, with 2026 CapEx guidance raised to $130–145 billion.

These companies have not suddenly lost their ability to generate profits, and their earnings remain substantial; however, the amount of cash available for discretionary use is decreasing.

This is why, in the AI era, free cash flow is becoming an indicator that can no longer be ignored compared to EPS alone.

But borrowing itself does not imply danger.

For companies like Alphabet and Amazon, debt is primarily a capital structure tool.

They have substantial core operations, stable cash flows, and high credit ratings; issuing long-term bonds to spread construction costs over time under conditions of highly front-loaded capital expenditures is a normal maturity matching strategy.

What truly matters is whether financing is optimizing the balance sheet or beginning to put pressure on it, once capital expenditures consistently exceed the company’s ability to generate cash.

Oracle is one of the most extreme examples currently available.

By the end of fiscal year 2026, Oracle's annual capital expenditures reached approximately $55.66 billion, while operating cash flow amounted to only about $32 billion, resulting in a free cash flow of -$23.69 billion for the year. Meanwhile, the company completed approximately $43 billion in debt financing and $5 billion in equity financing during FY2026, with total future principal repayments on all borrowings reaching approximately $130.1 billion as of the end of May.

On July 9, S&P Global Ratings downgraded Oracle's long-term credit rating from BBB to BBB-, the lowest tier within investment grade, one step below which enters speculative grade.

So, while Alphabet is primarily leveraging its balance sheet in its AI spending, Oracle has begun challenging it—this is the new framework that must be established for the next phase of the AI market.

Previously, the market focused most on how much AI revenue grew, how many orders increased, and how fast the cloud business expanded. In the future, several additional questions need to be added: How much did the company spend to achieve this growth? How much additional capital expenditure is required for every additional dollar of revenue? How much free cash flow remains? How much debt is needed? After interest, depreciation, and lease costs all begin to appear on the income statement, how much profit is ultimately left?

Ultimately, what truly determines valuation is not growth itself, but the capital efficiency of that growth.

If Alphabet, Amazon, and Meta issuing bonds merely represented a change in financing methods, NVIDIA’s latest move has taken this a step further.

On August 10, NVIDIA announced strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish an independent compute financing platform aimed at mobilizing over $500 billion in third-party capital for AI infrastructure over the long term.

NVIDIA retains the option to provide a backstop of up to approximately 25%, or up to $125 billion, for potential transactions, though the specific contributions and timing of fund deployment by the institutions have not yet been disclosed.

What truly matters is how NVIDIA defines this model. In its official announcement, NVIDIA directly describes AI Compute and AI Factory as a new "investable asset class."

The logic isn't actually complicated. In the past, a Neocloud needing to purchase billions of dollars' worth of GPUs had to raise substantial capital on its own. In the future, this model may increasingly resemble traditional infrastructure financing: GPUs and data centers become assets, customers lease computing power over the long term to generate cash flow, and institutions like Apollo, BlackRock, and KKR provide long-term capital, with the project repaying financing costs using future computing revenue.

Thus, GPUs are no longer merely chips sold in a one-time transaction; instead, the complete AI Factory—comprising GPUs, data centers, power, and long-term computing power contracts—is now being packaged as an infrastructure asset capable of generating sustained cash flows, and is eligible for valuation and financing in capital markets.

The significance behind this is enormous.

Once AI infrastructure becomes eligible for investment by pension funds, insurance capital, private credit, infrastructure funds, and asset management firms, the pool of capital accessible to AI will no longer be limited to the cash reserves of technology companies.

This could make this round of AI infrastructure development last longer than the market expects.

At the same time, it also transforms the nature of risk. After all, the greatest risk in the first phase of the AI cycle was that no one would use AI; but in the latest second phase, a more pressing concern may be that while AI is being used, computing power prices are falling too rapidly, causing revenue growth to lag behind debt, depreciation, and financing costs.

Especially as GPU update speeds remain extremely fast, asset return models designed today for five-year or longer cycles must be based on the critical assumption that these devices will continue to maintain sufficiently high utilization and economic value in the coming years.

This is a new variable that the AI industry must confront after financial capital truly began to enter.

In addition, there is another capital commitment that is more easily overlooked.

According to Reuters’ analysis of company filings, Microsoft, Meta, Oracle, Amazon, and Alphabet have disclosed approximately $1.09 trillion in future lease payment commitments not yet executed, with Microsoft at approximately $329.1 billion, Meta at approximately $279 billion, Oracle at approximately $260 billion, Amazon at approximately $137.2 billion, and Alphabet at approximately $85.2 billion; Meta subsequently signed new data center lease agreements worth approximately $68 billion in July, raising the total known scale to approximately $1.16 trillion.

Of course, these figures cannot be simply interpreted as "tech companies have already incurred $1.16 trillion in debt." Many contracts span over a decade and have not yet been fully implemented, so not all are currently recorded as lease liabilities on the balance sheet; Amazon's related disclosures also include assets such as warehouses, offices, aircraft, and vehicles, not all of which are AI data centers.

But it still reveals something important: a significant portion of AI infrastructure investment over the coming years has already been locked in advance. As long as demand for computing power continues to grow, these commitments form the foundation for future revenue growth. However, if model efficiency improves rapidly, the cost per unit of computing power continues to decline, or enterprise AI commercialization proceeds slower than expected, the long-term capacity locked in today could become fixed costs that are difficult to reduce quickly in the future.

And this is precisely the biggest difference between the AI funding cycle and a purely technological cycle.

It is still difficult to interpret these changes simply as negative signals.

On the contrary.

The entry of bond markets, private credit, pension funds, insurance capital, and global asset management institutions is likely to further expand the scale of capital accessible to AI, prolonging this wave of infrastructure development.

As of August 12, U.S. stocks remain close to their all-time highs. Following its latest earnings report, Amazon surged nearly 9% in after-hours trading due to a 37% growth in AWS; Microsoft also received strong market recognition after demonstrating growth in its cloud business and cash generation capabilities. Meanwhile, Alphabet faced pressure after announcing further increases in CapEx, while Meta experienced selling pressure following a 91% plunge in free cash flow.

The market hasn't begun to reject AI investments; it is simply transitioning from pure "demand-driven trading" to a more rigorous "return on capital trading":

• In the first phase, it's a test of who dares to spend the most.
• In the second phase, the competition is about who can acquire more computing power at a lower capital cost and generate sufficient cash flow from each dollar invested;

But when AI shifts from being a capital expenditure for tech companies to an asset that Wall Street can buy, finance, and price, the real question this competition must eventually answer is: “Whose creation will ultimately generate profits?”

And this may be the true dividing line that determines the valuation gaps between Google, Amazon, Meta, Oracle, CoreWeave, and even NVIDIA over the next one to two years.