For most of the past year, the debate around AI capital spending has focused on one question: can Google, Meta, Amazon, Microsoft, and Oracle keep spending at this pace?
Oracle's latest earnings pushed that debate forward. Revenue reached $19.3 billion, up 30% year over year, OCI infrastructure revenue rose 121% to $7.4 billion, and remaining performance obligations climbed to $664 billion. Oracle also said it signed more than $30 billion of new AI cloud contracts during the quarter.
That makes it harder to argue that the AI buildout is being funded against imaginary demand.
But the same quarter also showed why the next question matters more. Oracle spent $28.5 billion on capital expenditures and still reported negative $5.4 billion of free cash flow. Its operating cash flow was helped by $11.36 billion of customer prepayments with a significant financing component.
AI demand may be real. The harder question is how much capital must be deployed before that demand turns into durable free cash flow.
Can Big Tech Still Afford the AI CapEx Boom?
For the strongest hyperscalers, the answer is still yes.
Microsoft, Alphabet, and Meta continue to generate enormous operating cash flow and retain access to investment-grade funding. A rise in debt issuance does not, by itself, mean the AI buildout is financially unsustainable.
But affordability is not the same as unlimited capital at the same price.
Over the past year, Alphabet, Amazon, Meta, Microsoft, and Oracle have issued roughly $220 billion of debt. By early July, Amazon, Alphabet, Meta, and Oracle alone had issued about $194 billion, already 79% more than their combined issuance in all of 2025. Goldman Sachs estimated that the five largest hyperscalers could issue around $250 billion in 2026 and as much as $400 billion in 2027.
The capital is still available. The price is starting to change.
Why Are AI Credit Spreads Starting to Widen?
The more important signal is not the absolute amount of debt. It is how investors are absorbing it.
Hyperscaler bond order books were close to 5x covered in February, but fell to below 2x by July. Median new-issue concessions rose from roughly 2.25 basis points in 2025 to about 12 basis points in 2026. Among 91 comparable hyperscaler bonds issued this year, 78 later traded at higher yields, with a median increase of roughly 22 basis points.
Reuters also found a more unusual pricing distortion.
Under normal conditions, a larger bond from the same issuer with similar maturity and credit risk should often trade at a slightly lower yield because the larger issue is more liquid. Instead, matched-bond comparisons for Oracle, Alphabet, Meta, and Nvidia showed that some large new issues were trading at spreads roughly 15–22 basis points wider than comparable older bonds.
That is not a credit crisis. A 15–22bp spread difference is nowhere near enough to stop Microsoft or Alphabet from building data centers.
But it is a price signal.
The bond market is not refusing to fund AI. It is beginning to demand more compensation to absorb the supply.
Why Do Treasury Yields Matter for AI Financing?
Corporate borrowing costs do not start at zero.
A simple way to think about the cost of new corporate debt is:
Treasury yield + credit spread + new-issue concession
The macro backdrop tightened again on September 11. August headline CPI rose 0.4% month over month and 3.4% year over year, both in line with consensus. But core CPI increased 0.3% month over month, above the 0.2% expected pace.
Before the release, futures markets were pricing roughly a 70% probability of a 25-basis-point Fed hike at the September meeting. After the CPI report, live market pricing moved to around 90%.
At the same time, the 10-year Treasury yield was already pressing toward 5%.
That matters because AI borrowers are now facing two repricings at once.
The first is the base rate. Even if a company's credit quality does not change, a higher Treasury yield raises the starting point for corporate borrowing costs.
The second is the AI-specific premium. Heavy hyperscaler issuance is pushing investors to demand wider spreads and larger new-issue concessions at the same time.
So the financing pressure is no longer just:
AI debt supply → wider spreads
It is increasingly:
higher risk-free rate + wider AI credit spread + higher issuance premium
That combination raises the hurdle rate for the next data-center project much faster than a 15–22bp spread change would suggest on its own.
The distinction matters. The matched-bond anomaly is measuring the AI-specific credit premium after controlling for the Treasury benchmark. The CPI and Fed repricing, by contrast, are lifting the benchmark itself. One is a sector-specific repricing; the other raises the financing floor for nearly every borrower.
It is also important not to overstate the causality. AI debt issuance is not the main reason the 10-year Treasury yield is near 5%. Fiscal deficits, inflation, energy prices, and monetary-policy expectations remain much larger drivers.
But when the risk-free rate is rising at the same time that AI credit spreads and issuance concessions are widening, capital becomes more selective even before it becomes scarce.
Why Oracle Is the Best AI Financing Stress Test
Oracle is especially useful because its latest quarter separated the demand question from the financing question.
The demand is strong. OCI infrastructure revenue rose 121%, RPO reached $664 billion, and new AI cloud contracts exceeded $30 billion.
But the capital intensity is also extreme.
Oracle generated $23.1 billion of operating cash flow in the quarter while spending $28.5 billion on CapEx. More importantly, that operating cash flow included $11.36 billion of customer prepayments with a significant financing component.
Those prepayments should not be treated as a negative signal. In fact, customers willing to pre-fund capacity can strengthen the economics of large infrastructure contracts.
But they show that AI financing is no longer just a hyperscaler using its own cash to build capacity and then billing customers later.
The structure is increasingly becoming:
customer contract or prepayment → corporate capital → debt/equity financing → data-center construction
That changes what investors should track.
The most important Oracle question is no longer whether the $664 billion RPO is real.
It is:
How much capital must Oracle deploy before that RPO becomes free cash flow?
AI Infrastructure Is Being Financed Far Beyond Corporate Bonds
Looking only at the roughly $220 billion of hyperscaler bond issuance understates the size of the financing system.
AI infrastructure is increasingly funded through a mix of:
operating cash flow, corporate bonds, equity, customer prepayments, long-term leases, SPVs and JVs, project finance, private credit, and asset-backed loans.
One of the largest hidden commitments is leasing.
Reuters calculated that Microsoft, Meta, Oracle, Amazon, and Alphabet had committed roughly $1.09 trillion of future leases that had not yet commenced, much of it related to data-center expansion. Because those facilities had not yet entered service, the commitments had not yet appeared as ordinary lease liabilities on the balance sheet.
The largest disclosed amounts were approximately:
- Microsoft: $329.1 billion
- Meta: $279.0 billion
- Oracle: $260.0 billion
- Amazon: $137.2 billion
- Alphabet: $85.2 billion
These figures should not simply be added to corporate debt. A lease commitment is not the same thing as debt principal, and not every dollar is purely AI-related.
But economically, they are long-term fixed commitments that future data-center cash flows must support.
That is why balance-sheet debt alone no longer captures the full financing burden of the AI buildout.
AI Is Becoming a Credit Ecosystem
The Bank for International Settlements has now started treating this as a financial-stability issue.
The BIS estimates that AI-related capital spending by the five largest technology companies will exceed $1 trillion across 2025 and 2026. It also warns that CapEx is beginning to exceed internally generated cash flow at some firms, increasing reliance on debt, private credit, and more complex off-balance-sheet structures.
That matters because the economic exposure can sit outside the hyperscaler itself.
A third-party SPV or joint venture may borrow to build a data center, while the hyperscaler signs a long-term lease, commits to purchase capacity, or provides a guarantee. Legally, the debt may belong to the project vehicle. Economically, the project still depends on the AI customer generating enough cash to honor those commitments.
The AI buildout is no longer funded by Big Tech cash flow alone.
It is becoming a broader credit ecosystem.
Why the Same GPUs Can Carry Very Different Financing Costs
CoreWeave shows how quickly this financing system can start differentiating between projects.
In March, CoreWeave completed an $8.5 billion financing facility backed by GPU and HPC infrastructure tied to high-quality customer contracts. The floating-rate portion was priced around SOFR + 225bp.
A later $3.1 billion facility tied to lower-rated customers was priced around SOFR + 450bp. Another financing later reached roughly SOFR + 550bp.
The hardware was not the only thing the lenders were pricing.
They were also pricing:
- counterparty quality
- contract duration
- utilization certainty
- guarantees
- collateral structure
That creates a useful new operating variable for AI infrastructure investors:
bankability of contracted demand
An AI order is not equally valuable simply because the headline contract size is large. The financing market will increasingly distinguish between contracts that can support cheap capital and contracts that cannot.
Which AI Companies Will Feel the Pressure First?
Financing pressure will not hit the AI industry evenly.
Microsoft and Alphabet are unlikely to be the first casualties of higher borrowing costs. Their balance sheets and cash flows remain strong enough that a higher cost of capital mainly raises the return threshold for new projects.
Oracle sits in a more sensitive position because its current CapEx burden is much higher relative to internally generated free cash flow.
The most exposed group may sit one layer below the hyperscalers: neoclouds, independent data-center developers, and infrastructure companies that rely heavily on project finance or private credit.
That is where rising capital costs can become operational.
Lenders are already asking for stronger permits, leases, guarantees, customer commitments, and other backstops before funding new projects.
The first sign of a financial bottleneck is therefore unlikely to be Microsoft suddenly cutting GPU purchases.
It is more likely to be marginal projects losing access to cheap capital.
Will Higher Financing Costs Slow AI CapEx?
Not necessarily.
Oracle's latest results show that AI cloud demand remains strong, while Microsoft, Alphabet, Meta, and Amazon still have enormous financing capacity.
The more likely near-term effect is greater project selection.
As the cost of capital rises, hyperscalers and infrastructure developers will place more weight on utilization, customer credit quality, contract duration, time to revenue, and cash conversion.
The transmission mechanism is straightforward:
AI CapEx rises → external financing needs rise → debt, leases, and private credit expand → investor absorption weakens → financing premiums rise → project hurdle rates rise
The result is not an immediate collapse in AI spending.
It is a higher bar for which projects get funded.
Capital Is Becoming AI's Fifth Bottleneck
The AI infrastructure debate has mostly focused on four physical constraints:
GPU, memory, optics, and power.
Each answers the same question:
Can this infrastructure be built?
Capital answers a different one:
Should this infrastructure still be built at this price?
When AI investment was measured in tens of billions of dollars, capital could be treated almost like an unlimited input.
That assumption is becoming harder to defend.
Big Tech is now issuing hundreds of billions of dollars of debt, signing more than $1 trillion of future lease commitments, using customer prepayments, private credit, SPVs, and GPU-backed financing, while Treasury yields remain close to multi-decade highs.
The financial market is beginning to influence which AI projects are economically viable.
That is why a 15–22bp matched-bond spread anomaly matters.
It may be one of the first visible signs that the AI buildout is developing a financial bottleneck, not just a physical one.
The Bottom Line
There is no evidence yet of an AI credit crisis.
The largest hyperscalers still have powerful balance sheets, strong cash generation, and access to deep capital markets. Oracle's latest results also reinforce that AI cloud demand is real.
But the marginal dollar of capital is becoming more expensive.
Bond order-book coverage is falling. New-issue concessions are rising. Matched AI-related bonds are trading at wider spreads. Treasury yields are near 5%, and the August CPI report pushed markets sharply closer to expecting another Fed hike in September. Project financing can differ by hundreds of basis points depending on the quality of the underlying customer contract.
That combination matters more than any one variable by itself. AI-specific credit repricing is now happening on top of a higher macro financing floor.
That changes the next phase of the AI investment cycle.
The most important question is no longer only how many GPUs, HBM stacks, optical modules, or megawatts the industry can secure.
It is increasingly:
Which AI projects can generate returns high enough to justify their cost of capital?
If that question begins to determine which data centers are built, which companies can keep expanding, and which contracts can actually be financed, then capital is no longer just funding the AI boom.
Capital itself has become part of the bottleneck.
Sources
- Reuters, "AI debt splurge is warping credit spreads"
- Reuters, "Hyperscaler debt binge pushes yields up as investor demand cools"
- Reuters, Oracle Q1 FY27 earnings coverage
- Reuters, "AI data-centre race builds $1 trillion lease burden for Big Tech"
- Reuters, "AI construction crunch widens credit fault lines"
- Reuters, August 2026 CPI coverage and September Fed-hike pricing
- Bank for International Settlements, AI financing and financial-stability commentary
- Oracle Q1 FY27 earnings release
- Microsoft, Alphabet, Meta, Amazon, Oracle and CoreWeave SEC filings
Related Reading
Is the AI Bubble About to Break? The Real Risk May Be Hiding in the Credit System
AI demand is still growing, but as GPUs, long-term compute contracts, project finance, and customer credit become more tightly connected, the real risk may be hiding in how the expansion is funded.
What Is AI Infrastructure Credit Risk? A Beginner’s Guide to the AI Bubble Debate
AI demand is real, but the infrastructure behind it is increasingly financed through debt, long-term contracts, project finance, and GPU-backed lending.
Can AI CapEx Pay Off? Start With Token Cost, Then Watch the Payback Period.
Nvidia is showing that the unit cost of AI production is falling. Investors are asking whether cheaper tokens can turn into visible profits and free cash flow fast enough.
Is HBM Demand Really This Strong? Testing the AI Memory Boom Against Orders, Utilization, and 2027 Supply
AI server orders, cloud utilization, and HBM contracts all point to real demand. The harder question for 2027 is whether high prices and new qualified supply begin to change the cycle.
Invest better with thoughtful research.
Join readers who receive our best ideas and insights straight to their inbox.
Disclosure
This article is for research and education only. It is not investment advice. References to specific debt issuance amounts, credit spreads, new-issue concessions, lease commitments, Oracle Q1 FY27 figures and CPI / Fed pricing reference public reporting (Reuters, SEC filings, BIS commentary) as of the article date; investors should consult primary filings and current market data before making any investment decision.
Comments