AI Meets the Cost of Capital: Why Real Cash Flow Is Only the Beginning

The most important question about the AI investment boom is becoming a financial one: how much must the technology earn to justify the capital committed to it?

A company can have real customers, strong cash flow and excellent technology while still paying too much for capacity, building it too early or accepting returns that are too low. Higher bond yields make that distinction harder to ignore. They increase the cost of waiting for future profits and the appeal of investments with more predictable cash payments.

The AI debate needs to move beyond a binary choice between a revolution and a bubble. A genuine technological revolution can coexist with overinvestment in particular assets. The task for investors and business leaders is to identify where durable economic value will accrue, when it will arrive and whether the price paid leaves room for uncertainty.

What Bessent's argument gets right

In an October 4 report, Bizportal described U.S. Treasury Secretary Scott Bessent's view that rising U.S. bond yields reflect a wider international adjustment. It also reported his argument that major AI investors, including Microsoft, Alphabet, Meta and Amazon, have established businesses and substantial cash generation. The report placed the U.S. ten-year yield around 5.25% and thirty-year mortgage rates above 7%; these are attributed market observations, not live quotes. [1]

The distinction between a cash-generating incumbent and a venture dependent on its next funding round is meaningful. An established business may finance investment through a downturn, absorb disappointing projects and distribute new products through relationships it already owns. Those advantages change financing risk.

Microsoft's published fiscal 2026 results provide a concrete example of scale: annual revenue of $331.8 billion and operating income of $155.2 billion. These are company-wide figures, not a measure of AI profitability. They demonstrate a substantial existing business, but cannot establish the return on each new infrastructure project. [3]

The financing argument answers whether a company can afford to build. Investment analysis must also answer whether building creates sufficient value. Confusing those questions allows a strong balance sheet to become an excuse for weak capital discipline.

A global rise in yields still has local consequences

Calling higher yields a global phenomenon is a description of their breadth. It does not remove their economic effects or show that every country faces the same underlying problem. Governments differ in fiscal position, debt maturity and investor demand. Businesses differ in their funding needs and ability to pass costs to customers.

A useful framework separates a long-term bond yield into expected short-term rates over its life and a term premium for bearing longer-term risk. New York Fed research uses this decomposition and emphasises the importance of term premiums. A ten-year yield therefore cannot be read simply as a forecast of the next central-bank decision. [4]

Several forces can operate together. Stronger expected growth can support higher real rates. Inflation uncertainty can alter the compensation investors demand. Greater debt issuance can change the balance between supply and demand. It takes evidence to assign weights to those explanations; a shared direction across countries does not identify one cause.

For an AI infrastructure project, the consequence is practical. A higher financing rate increases interest expense on new or refinanced debt, all else equal. A higher required return also reduces the present value of distant cash flows, even if a project uses internal cash. A company does not need to borrow for the cost of capital to matter.

Internal cash has an opportunity cost

Cash generated by an existing business belongs economically to its owners. Management can reinvest it, acquire another business, repay debt or return it to shareholders. Choosing an AI project means forgoing those alternatives.

That is why the relevant hurdle is a risk-adjusted return, rather than the interest rate printed on a loan agreement. A project with uncertain utilisation, pricing and replacement costs should not be evaluated as though its future cash flows were equivalent to sovereign-bond payments.

Consider a deliberately simplified example. A project costs $100 million today and produces $15 million of annual free cash flow at the end of each of the next ten years. Assume that these cash flows already reflect operating costs, taxes and necessary reinvestment, with no terminal value. At an 8% discount rate, their present value is about $100.65 million, giving a slightly positive net present value. At 10%, their value falls to approximately $92.17 million, producing a negative net present value of about $7.83 million.

Nothing about the technology or customer forecast changed. Only the required return changed. These are illustrative assumptions, not a valuation of any named company, and the discount rates are project hurdle rates rather than Treasury yields. The example shows why the same operating plan can look acceptable in one financing environment and inadequate in another.

Timing matters as well. If the same ten payments shift back by one year, their value at 10% falls to about $83.79 million, before considering any additional construction expense. A delayed grid connection can therefore hurt economics even when the eventual demand remains intact.

Follow cash after investment, not just before it

Microsoft reported $55.4 billion of operating cash flow and $35.8 billion of cash spending on property and equipment in fiscal Q4 2026, leaving $19.6 billion of free cash flow. Its broader capital-expenditure figure was $41 billion, including finance leases. These measures have different scopes and should not be substituted for one another. [2]

The analytical point is wider than one company. Operating cash flow indicates the cash generated by the business before capital investment. Free cash flow, under a stated definition, helps reveal what remains after that investment. A company can grow revenue and operating profit while spending a larger share of its cash on infrastructure.

Neither pattern is automatically good or bad. During an attractive expansion, lower near-term free cash flow may be the price of profitable future growth. The concern arises when investment keeps accelerating but the connection to incremental, sustainable cash generation becomes less clear.

Analysts should also distinguish capacity serving existing demand from capacity built in anticipation of future demand. The first can have a shorter path to monetisation. The second requires stronger assumptions about adoption, pricing and utilisation. A headline capital-expenditure total does not provide that distinction.

Finance leases, purchase commitments and long-term service obligations deserve attention alongside conventional debt. Their accounting presentation may differ, but they can all constrain future cash. Comparing companies requires consistent definitions and a clear statement of what each measure includes.

Hardware, buildings and customers run on different clocks

AI infrastructure combines assets with very different economic lives. A building or power connection may be useful across multiple generations of computing equipment. A specialised accelerator may face faster competition from more efficient hardware. A customer contract may be shorter than either asset's expected payback period.

An investment model should reflect those differences. Applying one useful life to an entire campus can conceal the cost of replacing computing equipment. Assuming that current prices per unit of compute persist throughout that life can be equally misleading.

Depreciation is an accounting allocation, while economic obsolescence concerns the asset's actual earning power. A longer estimated accounting life can reduce annual depreciation expense without improving the asset's competitive position. Conversely, new hardware does not instantly make all older equipment worthless; the question is which workloads it can still serve profitably.

The same scrutiny applies to contracts. Contracted revenue is more informative when the customer can pay, cancellation provisions are understood and delivery obligations are achievable. A large order book is a starting point for evaluating cash flows, not a replacement for that evaluation.

The most useful management discussion would connect each spending cohort to capacity delivered, paid utilisation, realised pricing and expected replacement needs. That is much closer to the economic question than celebrating aggregate spending as an achievement in itself.

Power turns an AI strategy into an infrastructure strategy

The International Energy Agency's 2025 Energy and AI report projected global data-centre electricity demand of about 945 terawatt-hours in 2030, compared with roughly 415 in 2024. It also estimated that unresolved grid constraints could put around one-fifth of planned data-centre projects at risk of delay. Those are dated forecasts and conditional assessments, not realised 2030 outcomes. [5]

For investors, the implication is that a chip-delivery schedule is only one part of the deployment schedule. Power availability, connection approvals, cooling and network capacity can determine when purchased equipment begins earning revenue.

Financial commitments may start earlier. A business might pay suppliers, carry construction costs or service debt while capacity remains unavailable. If equipment improves rapidly during that delay, the operator can face both lost revenue and a weaker competitive position when the facility opens.

Location therefore becomes a capital-allocation decision. A site with cheaper land may be less attractive if power arrives late or reliability is poor. A more expensive location can be economically superior if it enables earlier revenue and lower operational risk.

This does not mean that every infrastructure bottleneck defeats the AI thesis. It means that execution quality belongs in the valuation, with explicit assumptions about timing and contingency costs.

Real transactions do not guarantee independent end demand

An AI ecosystem can contain legitimate investments, cloud contracts, equipment purchases and financing relationships among the same participants. Such arrangements can accelerate innovation and improve coordination. They also require analysts to distinguish financial support from demand ultimately funded by customers outside the investment cycle.

Imagine, purely as an illustration, that a cloud provider invests in a model company, which uses part of that funding to buy cloud services from the provider. Both transactions may be genuine. But the resulting cloud revenue does not, on its own, establish that independent end users will support the same spending once investment funding slows.

This is a question about the durability and concentration of demand, not an allegation of improper accounting. Relevant evidence includes cash collections, customer concentration, contract enforceability, ongoing financing needs and the share of revenue supported by external business use.

Double counting also deserves care. Spending can appear as revenue at several stages of a supply chain. Adding chip sales, cloud revenue and application revenue together without considering intermediate transactions can exaggerate the economic value ultimately purchased by end customers.

A healthy market discussion should ask where new customer value is created, how much customers are willing to pay for it and which participant can retain a sustainable margin.

A 4% earnings yield is not a 5% bond yield

A stock trading at twenty-five times earnings has an earnings yield of 4%. That arithmetic is useful, but it is not a direct expected-return comparison with a bond yielding 5%.

Earnings are not necessarily distributed to shareholders. Some must be reinvested, and future earnings can rise or fall. A bond has a contractual payment structure, but its quoted yield is not a guarantee of a particular investor's realised return: purchase price, holding period, reinvestment, credit risk and inflation matter.

The right comparison considers the timing and uncertainty of the cash flows, as well as growth and the price paid. A growing company may justify a low current earnings yield; it still needs a credible path to enough future cash. Higher bond yields make optimistic assumptions more expensive to defend.

This is especially relevant where market prices embed several favourable developments at once: strong demand, stable margins, rapid capacity deployment and a long period of technological advantage. If those expectations leave little room for delay or competition, even a successful business can produce a disappointing investment return.

Three paths investors should be prepared to test

In a productive-expansion scenario, customers renew because AI creates measurable value, infrastructure fills as it becomes available and efficiency gains support margins. Spending remains substantial, but cash generation grows enough to justify it.

In a useful-but-overbuilt scenario, adoption continues while capacity grows faster than profitable demand. Customers benefit from lower prices, yet infrastructure owners face weaker utilisation and returns. AI can succeed economically for its users while disappointing some of its investors.

In a funding-and-execution squeeze, expensive capital combines with project delays, customer concentration or weaker pricing. Firms with greater liquidity and flexible commitments may absorb the pressure; heavily committed projects have less room to adjust.

These are analytical scenarios, not probability estimates or predictions. Their purpose is to expose which assumptions a valuation needs. If a project only works under the first path, investors should know how much they are paying for that dependence.

What capital discipline should require

Boards should approve major investments against explicit commercial milestones: customer commitments, power availability, deployment dates, utilisation and cash returns. The option to stage a build or delay a later phase can have real value when demand and financing conditions are uncertain.

Compensation should reward economically useful capacity and durable results, rather than spending volume alone. Reporting should show what changed when costs rose, deployment slipped or utilisation fell short. A forecast becomes more credible when management explains the conditions under which it would reduce or redirect investment.

For enterprise buyers, the parallel test is practical: does the application improve a measured business outcome after software, compute, integration and review costs? Adoption statistics are helpful, but productivity should eventually appear in better output, lower unit costs, faster delivery or stronger revenue.

My view is that the AI boom can remain commercially powerful while becoming financially more demanding. Established cash flow gives the leaders resources and time. It does not suspend the obligation to earn an adequate return.

The next phase should be judged by the quality of capital allocation: what gets built, when it starts earning, who pays for it and how much cash remains after keeping the system competitive. Real money makes the investment possible. Sustainable returns make it worthwhile.

Dor Arad (דור ארד) · Independent analysis and opinion · October 4, 2026