Analysis by Evan Brooks | Numbers & Judgment Editorial

Published September 19, 2026. Evan Brooks is an editorial byline used by Numbers & Judgment for analysis of AI, technology and finance produced by the publication’s editorial team.

The AI investment boom is usually framed as a competition for customers, chips, engineers and electricity. Increasingly, it is also a competition for capital.

Goldman Sachs has estimated a baseline of roughly $7.6 trillion of cumulative AI infrastructure capital spending from 2026 through 2031, spanning compute, data centers and power. Its model rises from about $765 billion of annual AI capital spending in 2026 to roughly $1.6 trillion in 2031. Goldman explicitly describes the estimate as scenario-based rather than a forecast, but the scale is enough to change the finance question.

Recent reporting on Goldman’s credit-market analysis adds another dimension. U.S. convertible issuance has reached about $135 billion in 2026, with AI-related borrowers accounting for roughly 44%. Goldman has also reportedly lifted its forecast for U.S. investment-grade issuance to $2.3 trillion, with AI-related issuers expected to represent about one-quarter of supply.

The cost of capital is not just an input

A standard capital-allocation model treats the cost of capital as an external assumption. Management estimates cash flows, chooses a discount rate and decides whether the expected return clears the hurdle.

At sufficient scale, however, an investment cycle can begin influencing the financing environment used to evaluate the investment itself. AI companies, hyperscalers, data-center developers and infrastructure providers are raising debt and equity while governments are simultaneously borrowing for deficits, defense, energy and conventional infrastructure. All of them ultimately draw from overlapping pools of global savings.

That does not mean AI borrowing alone determines global interest rates. It does mean the financing side of the AI buildout can no longer be treated as incidental.

A good project can become a bad investment

This distinction matters for CFOs because operating success and investment success are different tests. A data center can fill. An AI product can grow. A productivity project can generate measurable savings. Yet the investment can still disappoint if the financing assumptions embedded in the original return model prove too optimistic.

Higher borrowing costs affect interest expense directly, but the consequences extend further. A higher weighted average cost of capital reduces present values, raises hurdle rates, pressures terminal values and makes refinancing more consequential. Long-duration projects are particularly sensitive because much of their expected value sits years in the future.

The $7.6 trillion number is conditional

Goldman’s infrastructure work is useful precisely because it does not present the headline number as inevitable. The total depends heavily on assumptions about the useful life of AI chips, the cost and complexity of next-generation data centers, chip architecture and bottlenecks in power, labor and equipment.

That conditionality matters to capital allocators. Faster hardware obsolescence can destroy the economics of assets that remain perfectly functional. More expensive power and cooling can raise the cost per megawatt. Delays can leave committed capital waiting longer before it produces revenue. Cheaper compute may lower unit costs but stimulate enough additional demand that aggregate spending still rises.

A CFO stress test for AI capital

Finance teams evaluating large AI commitments should stress-test more than adoption and productivity. At minimum, the model should ask:

• What if the refinancing rate is 150–250 basis points higher?
• What if utilization reaches plan one or two years later?
• What if the hardware becomes economically obsolete before its accounting life ends?
• What if competitors receive the same productivity benefit and pass part of it to customers?
• What if the project clears its operating targets but no longer clears the company’s higher hurdle rate?

The last question is easy to miss. Management teams naturally focus on whether a project is working operationally. Capital allocation asks a stricter question: is it working well enough relative to what the capital now costs and what else the organization could do with the money?

The judgment

AI’s physical constraints are increasingly visible: chips, power, grid connections, construction capacity and specialized labor. Capital belongs on that list.

The technology can generate extraordinary demand and still produce uneven returns for the institutions financing the infrastructure underneath it. The larger the buildout becomes, the more important it is to distinguish technological success from capital-allocation success.

The defining finance question of the AI boom may therefore become less about whether demand exists and more about who can finance that demand at a cost that still leaves an adequate return.

Sources

• Goldman Sachs Global Institute, “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out,” May 1, 2026.
• Goldman Sachs, “Private Markets Are Expected to Have a Growing Role in Data Center Financing,” June 12, 2026.
• Financial Express, September 19, 2026, reporting on Goldman Sachs credit-market analysis and 2026 issuance estimates.