Computing Power Competition Enters Credit Expansion Phase

Author｜Qingyun

Editor｜Xiaobai

Illustration｜AI Generated

Produced by｜Qiangdiao Next

On the evening of July 29, Meta's Q2 earnings report unveiled a stark contrast in figures: Q2 revenue reached $60.801 billion, up 28% year-on-year. Operating cash flow hit $31.862 billion, a 25% increase. However, capital expenditures surged to $31.078 billion during the same period, leaving free cash flow at just $784 million, a 91% year-on-year decline.

Meta also narrowed its 2026 capital expenditure guidance from '$125 billion to $145 billion' to '$130 billion to $145 billion.' At the start of the year, this range was '$115 billion to $135 billion.' Long-term debt on the company's books rose from $58.744 billion at the end of 2025 to $83.664 billion, with net bond financing reaching $24.91 billion in Q2.

This set of earnings data explains why Meta transferred its El Paso, Texas, data center into a joint venture the day before. The project's total development cost is approximately $14 billion, with BlackRock's fund holding an 80% stake and Meta holding 20%. Of this, $12.5 billion will be financed through project debt. Meta will continue to handle construction and management and will lease the entire campus upon completion, with this 1-gigawatt project expected to go live in 2028.

Meta also provided an initial residual value guarantee of approximately $13 billion, gradually decreasing thereafter. The lease has an initial term of four years, extendable up to 20 years after four renewals. If Meta exits within the first 16 years and the asset value falls below the threshold, it may need to cover the shortfall.

In 2025, Meta employed a similar approach for its Hyperion data center in Louisiana: the project's development cost is around $27 billion, with Blue Owl holding an 80% stake and Meta holding 20%. Meta leases the entire facility and provides a residual value guarantee for the first 16 years of operation.

The two transactions total $41 billion. Meta retains only a minority stake but still controls construction, operation, and usage rights. The trade-off is transforming one-time capital expenditures into multi-year rent, guarantees, and contractual obligations.

This is not merely a case of 'running short on cash.' As of late June, Meta still held $90.26 billion in cash, cash equivalents, and marketable securities, with its advertising business also experiencing rapid growth. More accurately, the scale and construction timeline of current AI investments have become so substantial that even Meta is unwilling to rely solely on current operating cash flow to cover them. By placing data centers into project companies, long-term leases, credit guarantees, and future usage demands become the basis for financing, with the bond market beginning to cover technology companies' computing power construction costs in advance.

01. Even Cash Cows Need to Borrow

Meta is not alone.

According to LSEG data, Amazon, Alphabet, Meta, and Oracle had issued approximately $194 billion in bonds by July 7 this year, a 79% increase from the entire year of 2025. Goldman Sachs projects that, including Microsoft, the five hyperscale cloud providers' bond issuance could reach $250 billion this year, rising to $400 billion by 2027.

Funds can still be raised, but the cost is rising. The median spread on 2- to 4-year bonds from these companies relative to risk-free rates has increased from 30 basis points in 2025 to 40 basis points. For long-term bonds over 20 years, the median spread has risen from 108.5 basis points to 118.5 basis points. The oversubscription ratio has dropped from nearly 5x in February to less than 2x in July. Credit expansion is not without limits; the more bonds are supplied, the higher the financing costs large firms pay for their next data center.

Amazon launched approximately $37 billion in bond issuance in March this year and raised another $25 billion in July. Oracle financed $43 billion in debt and $5 billion in equity in FY2026, with operating cash flow reaching $32 billion but free cash flow dropping to negative $23.7 billion. The company expects to raise another approximately $40 billion in FY2027.

On July 9, S&P downgraded Oracle's rating from BBB to the lowest investment grade of BBB-, with a stable outlook. The reason was not a lack of orders but massive upfront investments and long-term leases driving up leverage and sustaining negative free cash flow. Oracle's remaining performance obligations reached $638 billion, with S&P estimating roughly half stemming from OpenAI. If clients fail to fulfill obligations, Oracle may still be stuck with hard-to-transfer data center leases. Oracle has thus become one of the first major tech companies to face an explicit rating downgrade in this round of AI infrastructure expansion.

Alphabet's approach is more multidimensional. In the first half of 2026, the company issued $51.8 billion in bonds. As of late June, its long-term debt reached $98.2 billion. Signed but not yet commenced data center lease payments amounted to $85.2 billion, obligations not yet included in existing lease liabilities.

Microsoft offers another example. In Q4 FY2026, its capital expenditures reached $41 billion, a more than 70% year-on-year increase. Operating cash flow was $55.441 billion, with free cash flow still at $19.6 billion. Annual capital expenditures totaled approximately $145 billion.

Even larger figures lie in leases. As of late June, Microsoft had signed but not yet commenced data center leases worth $329.1 billion, more than tripling from $92.7 billion a year earlier. These leases will begin between FY2027 and FY2033, with terms up to 20 years, locking in future cash outflows.

Microsoft also extended the estimated term of long-term data center leases from 15 to 25 years, reducing its 2026 capital expenditure estimate from approximately $190 billion to $175 billion. Adjusting the caliber (which means 'criteria' or 'scope' in this context) lowered the disclosed figure but did not cancel signed contracts.

These companies are not short on cash, but AI infrastructure expansion has surpassed what operating cash flow can comfortably cover. Data centers require upfront commitments to land, power, and chips, with revenue realized only after computing power goes live. Bond issuance is just the most visible layer; leases, finance leases, project company liabilities, and client prepayments are also locking in expenditures in advance.

Chinese companies are adopting similar methods. In September 2025, Alibaba issued $3.2 billion in zero-coupon convertible bonds, with roughly 80% of funds earmarked for data center expansion, technology upgrades, and cloud services. Two months prior, Alibaba also issued approximately $1.5 billion in exchangeable bonds. The difference is that U.S. tech companies rely more on corporate bonds, project bonds, and private credit, while Chinese enterprises currently still rely on convertible bonds, bank loans, and operating cash flow.

02. Computing Power Contracts Become Collateral

The crux of AI financing expansion is that computing power has transformed from a technical product into a financial asset.

As of late March 2026, CoreWeave had $11.8 billion in delayed drawdown loans, along with $6.4 billion in notes and $4.7 billion in equipment financing. Collateral includes not just GPUs but also long-term 'take-or-pay' contracts signed with clients like Microsoft. Future computing power revenue is being converted into construction funds in advance.

Oracle is also involving clients in financing. By the end of FY2026, among its large AI contracts, clients had prepaid $75 billion for GPU purchases or directly provided GPUs. Cloud computing contracts are now fulfilling roles previously handled by bank loans.

Chip companies are also extending credit downstream. AMD provides up to $4.1 billion in guarantees for partners' data center leases. Alphabet's lease default guarantees for third-party operators using TPUs have risen to $44 billion, up from just $6.5 billion at the end of September last year.

The Financial Times also revealed that NVIDIA may be the actual tenant of Hut 8's Texas data center, with a base contract value of $19.6 billion and potential subleasing to cloud service providers purchasing its GPUs. If true, NVIDIA's role would extend from chip sales to credit intermediation.

A financing chain is forming among suppliers, cloud providers, data centers, and model companies. Contracts secure loans, loans build data centers, data centers purchase GPUs, and GPUs support the next round of computing power contracts.

03. The Bond Market Begins to Price AI

These financing methods alleviate construction-phase funding pressures but do not reduce the projects' ultimate costs.

Credit ratings are just the first health check. On July 29, Oracle's five-year credit default swap (CDS) spread was approximately 200 basis points, Meta's around 93, NVIDIA's about 78, while the investment-grade company CDS index stood at roughly 53. NVIDIA's CDS briefly hit an all-time high this week.

CDS acts as Default insurance (default insurance) for corporate bonds, with wider spreads indicating higher risk compensation. Q2 CDS trading volume related to the tech sector neared $650 million, a nearly sixfold year-on-year increase, as bond investors purchase more protection.

Meta's free cash flow nearing zero, Oracle's downgrade to the lowest investment grade, and NVIDIA's rising credit risk premium reflect three issues: cash coverage, client-lease mismatches, and suppliers extending credit downstream. The bond market is not predicting AI investment failures but demanding higher prices for uncertain returns.

Meta's El Paso project relies on leases to support $12.5 billion in debt, with a residual value guarantee threshold of approximately $13 billion. If AI demand falls short, Meta must still pay rent or cover asset value declines.

The project company holds the data center, Meta holds usage obligations, and bond investors bear project credit risk. While Meta no longer fully owns the asset, risks persist through leases and guarantees.

AI is already improving Meta's recommendation and advertising efficiency. The question is whether advertising gains can long-term cover $130 billion to $145 billion in annual capital expenditures, along with subsequent depreciation, cloud service fees, and data center operating costs hitting the income statement.

More importantly, Meta lacks the large-scale cloud business of Amazon, Microsoft, and Alphabet, making it difficult to sell excess computing power directly to external clients. Zuckerberg has raised developing external computing power services in earnings calls, and Meta was also reported to be in talks with Anthropic for up to $10 billion in computing power leases, but until these materialize, they remain supplementary explanations for massive investments.

That same evening, Microsoft's after-hours stock price briefly rose over 8%, while Meta's fell around 10%. Both companies are increasing AI investments, but the difference is that Microsoft's Azure revenue grew 43%, with commercial remaining performance obligations reaching $678 billion and free cash flow still at $19.6 billion. Meta's external computing power revenue has yet to scale.

Capital markets are beginning to price credit expansion based on return speed, contract quality, and solvency.

Computing power competition has thus raised the financing cost threshold. Measuring AI investments requires not just capital expenditures but also factoring in project liabilities, long-term leases, procurement commitments, client prepayments, and supplier guarantees.

Meta is far from a cash crunch, but with free cash flow at just $784 million, project financing is no longer merely a financial tactic but a crucial funding source to maintain construction momentum.

Oracle's rating downgrade, widening tech company CDS spreads, and declining bond oversubscription ratios indicate that while AI infrastructure can still secure financing, cheap credit is dwindling. The real differentiator will be who can first convert computing power, leases, and guarantees into sustained revenue rather than merely covering old investments with new credit.

- END -