For much of the AI boom, one of the strongest arguments against calling it a bubble was that the companies doing most of the spending could afford it.

This was not the late 1990s, when startups with little revenue raised money from investors to pursue business models that might never work. Nor was it the telecom boom, when companies borrowed heavily to lay far more fiber than the market needed at the time. The AI revolution was being funded by Microsoft, Alphabet, Amazon, Meta and some of the richest, most profitable companies on Earth.

They were spending their own money. If AI revenue took longer than expected to materialize, their enormous cash-generating businesses could carry the investment.

That was largely true when the buildout began. It is becoming much less true now.

The latest evidence comes from SoftBank, which plans to issue ¥1 trillion, or approximately $6.3 billion, in seven-year retail bonds. It will be the largest retail bond offering ever by a Japanese company. The bonds are expected to carry a coupon of between 4.3% and 4.9%, with the proceeds going toward both AI investments and the repayment of existing debt.

SoftBank has committed more than $60 billion to AI investments, including its enormous OpenAI wager. It has sold assets, borrowed against its holdings in Arm and explored other ways of financing Masayoshi Son’s determination to make SoftBank one of the central players in the AI economy. The new bond offering is another large piece of that funding campaign.

SoftBank has always been willing to make bold, highly concentrated technology bets, so borrowing billions to pursue Son’s AI vision should not surprise anyone. But this offering is part of a much larger change taking place across the industry.

The AI revolution is no longer being paid for primarily in cash.

The Cash-Flow Argument is Weakening

The scale of the shift is easy to see.

According to Vanguard, Alphabet, Amazon, Meta, Microsoft and Oracle issued an average of approximately $35 billion in debt annually between 2020 and 2024. Their issuance rose to $93 billion in 2025 and reached approximately $132 billion through the end of July 2026.

Goldman Sachs, using a broader measure, puts hyperscaler issuance at approximately $194 billion so far this year and estimates that debt will finance roughly one-third of hyperscaler capital expenditures in 2026. Estimates for total AI-related debt issuance—including chipmakers, data center developers, utilities and other parts of the infrastructure ecosystem—range from approximately $300 billion to $570 billion this year.

The different estimates reflect what is being counted. Some measure conventional corporate bonds. Others include project finance, private credit and other obligations. The precise total is less important than the direction. The cash-funded phase of the AI buildout is giving way to a financing structure that increasingly depends on creditors and outside investors.

Oracle is the most visible example. The company announced plans to raise between $45 billion and $50 billion during calendar 2026 through a roughly balanced combination of debt and equity. Oracle spent approximately $55.7 billion on capital expenditures in fiscal 2026 and now projects as much as $95 billion in fiscal 2027.

Oracle’s free cash flow is negative, and it expects to raise nearly another $40 billion through debt and equity in fiscal 2027. The company can point to $638 billion in remaining performance obligations as evidence that the demand behind its infrastructure spending is real. But much of that revenue will arrive over several years. The money required to build the capacity is needed now.

Alphabet reported negative free cash flow of $5.9 billion in its June quarter after capital expenditures reached $44.9 billion. It remains an extraordinarily strong company and generated $53.3 billion in free cash flow over the trailing 12 months, so this should not be presented as Alphabet running out of money. It does show how quickly AI infrastructure spending can consume even Alphabet-sized operating cash flow. The company has also tapped bond markets around the world, including a recent $3.9 billion Australian-dollar offering.

Meta produced only $784 million in free cash flow during its second quarter, while its long-term debt increased from approximately $58.7 billion at the end of 2025 to $83.7 billion by June 2026. Meta still has nearly $90 billion in cash and marketable securities, but its financing profile looks very different from the company that ended 2024 with less than $29 billion of long-term debt.

This is not a U.S. or Western phenomenon, and the funding pressure is not confined to debt. Alibaba has finalized a HK$80 billion, or $10.2 billion, placement of 710 million new shares in Hong Kong. The shares were priced at HK$112.70, an 8.4% discount to the previous close, making it the largest primary follow-on offering by a Hong Kong-listed company. Alibaba said it will use 100% of the net proceeds to invest in its full-stack AI capabilities, including expanding and enhancing its AI infrastructure.

Alibaba chose dilution rather than borrowing, but the underlying message is similar. Internal cash generation is no longer the only source of capital these companies are willing to rely on to maintain the pace of the AI race. Alibaba has already spent nearly half of its three-year capital-expenditure plan, while quarterly net profit fell 75% as AI-related spending surged. Its Hong Kong shares fell approximately 8% after the placement was priced, providing an immediate measure of what shareholders thought about being asked to finance the next stage of the buildout.

The borrowing extends beyond the hyperscalers. AI infrastructure provider Nebius recently raised $5 billion in convertible debt after initially seeking $4.5 billion. Broadcom is involved in discussions around another AI financing package that could exceed $60 billion and potentially approach $100 billion.

The Broadcom structure requires some care in how it is described. This would not simply be $100 billion placed on Broadcom’s corporate balance sheet. Much of the debt would be issued through a special-purpose vehicle, with Broadcom potentially guaranteeing part of the senior financing. Private-credit firms would provide capital backed by customer commitments, equipment and other project assets.

These different financing structures do not carry identical risks. A corporate bond, a lease, an SPV obligation, a residual-value guarantee and a private-credit loan are not interchangeable. But they are all being used to finance different layers of the same AI investment cycle.

Moving an obligation away from a corporate balance sheet does not make the risk disappear. It determines where the risk resides and who absorbs the loss if the assumptions behind the financing prove wrong.

Nvidia’s recent agreement with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR offers another glimpse of where this is heading. The group intends to establish independent financing platforms capable of mobilizing more than $500 billion in third-party capital for Nvidia-based AI infrastructure.

The idea is to make AI compute into an investable asset class. Goldman Sachs CEO David Solomon described it as an opportunity to create a market for credit backed by Nvidia compute. Future usage revenue would service the debt, while the Nvidia systems and associated infrastructure would provide collateral and residual value.

These are still memorandums of understanding, not $500 billion of committed funding. Nvidia may provide residual-value support of up to 25% on selected transactions, but it has not issued a blanket $125 billion guarantee. Individual projects would still need to be underwritten.

Even so, the ambition is clear. Wall Street wants to make AI compute recognizable, financeable and ultimately tradable. That could unlock capital from asset managers, insurers, pension funds, private-credit funds and infrastructure investors. It would also spread the financial exposure to AI far beyond the technology companies whose names dominate the headlines.

Debt Puts AI on a Clock

None of this means an AI crash is inevitable.

Debt is how societies finance infrastructure. Railroads, electrical grids, telecommunications networks and cloud computing were not built entirely from retained earnings. Borrowing allows companies to spread the cost of a productive asset over the years in which it generates revenue.

The leading hyperscalers also remain exceptionally strong borrowers. Much of their debt has long maturities and relatively low financing costs. AI demand is real, data center vacancy rates are low, cloud backlogs are growing and a significant portion of the new capacity is supported by customer contracts.

The problem is not that all this debt comes due on one ominous date. There is no single bill sitting in a drawer waiting to arrive.

The risk is the timing mismatch between the obligations created today and the revenue expected tomorrow.

Interest has to be paid throughout the life of the debt. Loans eventually need to be refinanced. Customer contracts expire. AI providers have to convert compute into products that customers will continue buying at prices that support the underlying infrastructure costs.

The equipment also has to remain valuable long enough to support the financing.

An AI data center combines the capital requirements of a railroad or power plant with the obsolescence cycle of semiconductors and the pricing pressure of software. The building may remain useful for decades, but the GPUs, networking systems and cooling configurations inside it can lose economic value much more quickly.

Nvidia argues that CUDA updates extend the life of its installed systems and that older generations such as the A100 remain commercially productive. That is a credible argument. At the same time, Nvidia’s growth depends heavily on convincing customers that each new generation delivers enough additional performance and efficiency to justify another round of spending.

The financial model needs older systems to retain substantial residual value. The product model benefits when newer systems make the old ones less competitive. There is tension between those two requirements.

AI efficiency creates another uncertainty. Models may become more capable while requiring less compute. Inference costs may continue to decline. Competition could drive token prices down faster than utilization rises. The world could consume vastly more AI while producing less revenue per unit of compute than today’s financing models anticipate.

AI does not have to fail for AI infrastructure investments to disappoint. It only has to monetize more slowly, or at lower margins, than the debt assumes.

If the financing cycle turns, it probably will not begin with Microsoft or Alphabet defaulting. It is more likely to begin with an AI lab, neocloud, data center developer or highly leveraged SPV missing its utilization or revenue assumptions.

A major tenant might be unable to fulfill an enormous compute commitment. A project could arrive late or over budget. A facility could struggle to secure enough power. A customer might renegotiate a contract when market prices fall. A lender could conclude that the GPUs backing a loan are worth substantially less than originally projected.

Once a few transactions disappoint, lenders will reprice the category. Interest rates and required equity contributions rise. Residual-value assumptions become more conservative. Projects that worked on a spreadsheet at one financing cost no longer work at another.

The result would be a pullback in construction that moves rapidly upstream. Data center developers order fewer servers. Cloud providers buy fewer GPUs. Networking demand slows. Utilities and power developers reconsider projects built around projected data center loads. The companies that supplied the boom feel the effects before many of the bonds themselves mature.

AI Can Win While Investors Lose

History offers plenty of examples of transformative technologies producing destructive investment cycles.

Railroads changed America, but many railroad investors lost money. Fiber-optic networks became essential to the internet economy, but that did not prevent the telecom financing bubble from bursting. The infrastructure survived, and later buyers often acquired it for a fraction of what the original investors spent building it.

AI may follow a similar path. The technology can transform business, science, health care, manufacturing and daily life while still leaving behind overbuilt facilities, impaired debt and investors who paid too much for capacity too early.

The wider concern is that exposure to this investment cycle is spreading throughout the financial system. Investors may own the AI wager through technology equities while also holding the bonds and private-credit vehicles financing AI infrastructure. Pension plans and insurers may provide capital to funds backed by compute leases and data center revenue.

If the economics of AI disappoint, the equity and credit sides of those portfolios may fall together. The diversification investors thought they owned may prove to be multiple claims on the same underlying assumption.

SoftBank’s bond offering does not prove that the AI boom is nearing its end. Neither does Oracle’s borrowing, Alphabet’s bond sales or Wall Street’s effort to turn Nvidia compute into an investable asset. Together, however, they show that the financing foundation of the boom has changed.

The AI industry has not necessarily borrowed more than it can ultimately repay. But it is increasingly borrowing against a future whose timing, margins and winners remain uncertain.

When AI was being financed from surplus cash, the industry could afford to be early. As debt assumes a larger role, being right eventually may no longer be enough.

AI now has to be right on schedule.