AI Has Real Revenue. Higher Interest Rates Still Change the Investment Case. By Dor Arad | October 4, 2026 Real customers. Real cash flow. Real infrastructure. Those are strong arguments for taking the AI economy seriously. They are not enough to prove that every AI investment will earn an attractive return. As bond yields rise, the distinction between financing a project and creating value from it becomes more important. A business can afford to build something that its shareholders would have been better off not funding. The latest debate around Treasury Secretary Scott Bessent's comments brings that distinction into focus. Bizportal's October 4 report describes his view that higher U.S. yields are part of a global trend and that major AI investors benefit from established, cash-generating businesses. [1] My view is that both observations can be true while the investment hurdle continues to rise. The strength of the balance sheet is the starting point Large technology companies have advantages that a startup awaiting its next funding round does not. Existing products can finance expansion. Distribution can lower the cost of reaching customers. A diversified business can absorb the failure of an individual project. Microsoft's fiscal 2026 results show the scale involved: $331.8 billion in annual revenue and $155.2 billion in operating income. These are company-wide numbers, not evidence of a standalone return on AI infrastructure. [3] They help answer whether investment is financially possible. Shareholders still need to know whether the incremental return is sufficient. Internal cash is not free capital. It could be used for another project, debt repayment or a distribution to shareholders. Spending it on AI requires a commercial case that survives comparison with those alternatives. Higher rates change the value of waiting Consider a simplified project costing $100 million today. Assume it produces $15 million of annual free cash flow for ten years, with no terminal value. Those assumed cash flows include the costs of operating and maintaining the investment. At an 8% required return, the payments are worth about $100.65 million today. At 10%, they are worth about $92.17 million. The operating forecast is identical. The investment changes from slightly positive to negative in net-present-value terms because the required return is higher. This is an illustration, not a valuation of any company. The rates are project discount rates, not Treasury yields. But the message is directly relevant: an investment can become less attractive without any deterioration in the technology. If every payment starts one year later, its value at 10% falls again, to about $83.79 million. Deployment delays are financial events. The cash-flow statement deserves as much attention as the product demo Microsoft's fiscal Q4 2026 disclosures reported $55.4 billion in operating cash flow, $35.8 billion of cash spending on property and equipment, and $19.6 billion of free cash flow. The broader $41 billion capital-expenditure measure included finance leases. Comparing those figures requires respecting their different definitions. [2] The wider lesson is to follow cash after investment. Revenue growth may be strong while the infrastructure required to support it absorbs an increasing share of cash. That can be sensible during a profitable expansion. It becomes harder to defend if investment keeps growing without a clear path to adequate future returns. The questions I would ask management are straightforward. Which spending supports committed demand? Which anticipates future demand? How much capacity is earning revenue? What happens to margins if compute prices fall? What commitments remain if growth slows? The physical bottleneck belongs in the valuation A data centre is not ready simply because its chips have arrived. It needs power, cooling, connectivity and an operating schedule that matches customer requirements. The IEA's 2025 Energy and AI report identified grid constraints as a source of possible delays in planned projects. That is a reminder to evaluate deployment assumptions alongside demand forecasts. [5] A delayed facility can consume cash before generating revenue. Equipment can also become less competitive while it waits to be used. The project then suffers from both later income and a potentially weaker market position. Different assets require different assumptions. A building may serve several hardware generations. Specialised computing equipment may need replacement much sooner. Customer commitments may be shorter than the expected investment payback. One optimistic useful-life assumption should not stand in for an analysis of the entire system. Demand must survive beyond the financing cycle Partnerships among investors, cloud providers and AI developers can be commercially valuable. They can also make the source of demand harder to interpret. In a hypothetical arrangement, a cloud provider invests in a model company, and that company spends some of the funding on the provider's services. The transactions may be entirely legitimate. They still leave an important question: how much spending would continue if investment funding slowed? The answer requires evidence about end customers, cash collections, renewals and willingness to pay. This is a test of demand durability, not an accusation of misconduct. Investors should also avoid treating every stage of a supply chain as separate final demand. A dollar spent on an application can support revenue elsewhere in the infrastructure stack without representing several dollars of independent customer value. Technology success and investor returns can diverge AI may improve productivity even if some infrastructure investments disappoint. Customers can benefit from intense competition and falling prices while suppliers struggle to recover their capital. That possibility does not settle the bubble debate in either direction. It makes company selection, project economics and entry price more important than a broad label. A stock with a price-to-earnings ratio of 25 has a 4% earnings yield. That is not directly comparable to a bond's yield: earnings can grow, require reinvestment and may never be distributed in full. Bond investors face their own interest-rate, inflation and holding-period risks. The useful comparison concerns future cash, its uncertainty and the price paid to obtain it. Strong technology cannot compensate for every valuation. What I would want to see next First, evidence that customers renew because AI improves a measurable business outcome after all relevant costs. Second, a clear connection between each phase of infrastructure spending, capacity delivered and cash generated. Third, financing and contractual commitments that leave room to adjust when demand, pricing or deployment dates change. Finally, management incentives tied to durable returns rather than the scale of spending announced. The ability to fund the AI buildout is an advantage. The ability to earn a sufficient return on that buildout is the test. Real money makes investment possible. Disciplined capital allocation determines whether it creates lasting value. Dor Arad (דור ארד) · October 4, 2026 Full website analysis: https://doraradinsights.com/en/ai-cost-of-capital-real-cash-flow Sources [1] Bessent: higher yields are a global trend, and AI is backed by real money — Bizportal, October 4, 2026 https://www.bizportal.co.il/wallstreet/news/article/20043910 [2] Fiscal 2026 fourth-quarter earnings conference call — Microsoft Investor Relations, July 29, 2026 https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 [3] Fiscal 2026 fourth-quarter and full-year results — Microsoft Investor Relations, July 29, 2026 https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast [4] The Term Structure of Expectations and Bond Yields — Federal Reserve Bank of New York, May 2016; revised April 2018 https://www.newyorkfed.org/research/staff_reports/sr775.html [5] Energy and AI: executive summary — International Energy Agency, 2025 https://www.iea.org/reports/energy-and-ai/executive-summary