Analysis by Evan Brooks | Numbers & Judgment Editorial
Published September 19, 2026. Evan Brooks is an editorial byline used by Numbers & Judgment for analysis produced by the publication’s editorial team.
Artificial intelligence is usually discussed as a race for chips, customers, talent and electricity. Increasingly, it is also a race for capital.
Goldman Sachs estimates that the AI infrastructure buildout could require roughly $7.6 trillion of cumulative capital spending between 2026 and 2031 across compute, data centers and power. Its baseline rises from about $765 billion of annual AI capital spending in 2026 to about $1.6 trillion in 2031. At that scale, financing is no longer just an input into the AI investment case. The investment boom can begin influencing the price of capital itself.
The same pool of savings
Governments are issuing debt to fund deficits, defense, energy security and infrastructure. At the same time, technology companies and the broader AI ecosystem are borrowing heavily to finance data centers, chips and power systems. Goldman’s rates strategists have described the result plainly: governments and AI companies are drawing from the same global pool of savings.
That does not mean AI borrowing alone explains higher bond yields. Inflation risk, fiscal deficits, resilient growth and term premiums all matter. But the AI buildout is large enough to become one contributor to capital scarcity rather than merely a beneficiary of capital markets.
Credit markets are already changing
Goldman Sachs Research estimates that nearly $500 billion of AI-related debt had already been issued in 2026 by early August. Hyperscalers accounted for about $194 billion of that amount, versus $108 billion in all of 2025. AI-related borrowers also represented roughly 18% of U.S. investment-grade supply at that point, and about 40% of issuance with maturities of 15 years or longer.
Those numbers matter because credit investors do not have unlimited appetite for exposure to a single theme. Even highly rated borrowers eventually encounter concentration limits, duration limits and portfolio-allocation constraints. Goldman’s credit team has argued that the binding constraint may not be how much additional debt hyperscalers can technically carry while retaining investment-grade ratings. It may be how much the investment-grade bond market can comfortably absorb.
A project can work operationally and fail financially
For CFOs, the lesson is broader than AI. A project can have attractive operating economics and still destroy value if its financing assumptions were formed in a cheaper-capital environment.
Consider a data-center or AI investment whose business case assumes a particular borrowing rate, terminal value and refinancing environment. If thousands of similar projects reach the market at once while governments are also increasing issuance, the discount rate used in the original model may no longer be realistic.
- Operating case: Does the project generate attractive cash returns before financing?
- Financing case: What happens if the cost of debt is 100 or 200 basis points higher?
- Absorption case: Can the relevant debt or equity market absorb the required financing without demanding a materially higher return?
- Refinancing case: Does the project still work if capital markets are less generous when the debt matures?
- Concentration case: Are investors already carrying significant exposure to the same technology, customer or infrastructure theme?
AI may raise its own hurdle rate
There is an important second-order effect here. Finance teams normally treat the cost of capital as an external assumption. But sufficiently large investment waves can help change that assumption.
If trillions of dollars of AI infrastructure investment require debt and equity financing, while public-sector borrowing is also elevated, then competition for savings can push required returns higher. A higher risk-free rate increases corporate borrowing costs and raises the discount rate applied to future cash flows. Projects whose economics looked compelling at yesterday’s hurdle rate may look much less attractive at tomorrow’s.
This is especially relevant to long-duration investments. Much of the expected value from AI infrastructure lies years in the future, while much of the capital must be committed now. That makes the economics unusually sensitive to both the price and availability of financing.
The judgment
The AI boom’s most visible constraints have been chips, power and data-center capacity. Capital should now be added to that list.
The key question is not whether capital will be available. Deep public and private markets can finance enormous amounts of investment. The more important question is at what price—and whether the returns from the projects being financed remain attractive after that price is paid.
For capital allocators, that changes the AI conversation. The investment case should not end with projected productivity gains, utilization or revenue. It should also ask whether the project survives a world in which the AI buildout itself has made money more expensive.
Numbers & Judgment 