Who Is Financing
the AI Boom?

01 / FOLLOW THE MONEYFour pools of capital, one connected trade

Microsoft, Amazon, Alphabet and Meta are not interchangeable AI investments. Their established cloud, advertising, software and commerce businesses support the buildout, but infrastructure spending competes with buybacks, dividends and other investment. Amazon illustrates the tension: trailing-year operating cash flow rose to $161.4 billion, while free cash flow fell to a $7.6 billion outflow after rising property-and-equipment purchases. A profitable business can still consume cash after investment. [2]

Bloomberg highlights AI infrastructure’s growing reliance on outside funding. [7] The distinction that matters is who owes the money. A corporate bond relies on the issuer’s wider business; a project loan may depend on a particular tenant, site and contract. Moving assets into a separate vehicle does not remove the economic risk. It changes where that risk sits.

Compute financing adds an unusual underwriting problem: the building may last for decades, while the most valuable equipment can become commercially outdated much sooner. Reuters reported lenders challenging Nvidia’s assumptions about GPU residual values and seeking stronger contractual support. [9] My interpretation is that an impressive chip is not sufficient collateral by itself. Repayment still needs dependable customers and cash flow.

02 / EARNINGS CHECKRevenue growth is only the first test

The latest reported quarters show strong demand, but they also show why investors need several measures of success. Cloud growth, operating profit and cash after investment answer different questions. These periods are identified individually; they are not a perfectly synchronized set of earnings. [1] [2] [3] [4] [5]

Amazon’s $62.6 billion net income included $53.4 billion of non-operating pre-tax other income, primarily from Anthropic investments. The different tax bases prevent directly subtracting these figures to find “underlying profit.” Operating results and cash flow better distinguish business performance from investment gains. [2]

For the company-specific earnings and expense discussion, see Note 004: Meta’s Muse Rally. Here, the focus is on how the wider buildout is financed.

03 / THE EQUITY TAPEThe market rewards different parts of the chain

On October 5, Nvidia rose 2.1%, Microsoft about 1.5%, and Meta about 2%, while the Nasdaq gained 1.05%. Reuters described a session supported by lower oil and changing rate expectations as investors approached earnings. [10] A stock move is not a clean measure of AI fundamentals: discount rates, positioning and the broader market also matter.

The longer picture can diverge from one session. Yahoo Finance reported on October 5 that Amazon had fallen 2.4% over the preceding month while Meta had rallied 23%. Its discussion also highlighted Amazon’s spending burden and company-specific regulatory concerns. [12] My inference: buying several large AI names does not guarantee the same exposure, but owning the supplier, its customer and its infrastructure financier can still leave a portfolio dependent on the same spending cycle.

04 / THE PHYSICAL BOTTLENECKYou cannot finance your way around a missing grid connection

A server generates no customer revenue while it waits for power. Grid connections, generation equipment and reliable supply can therefore affect when an investment becomes productive. The IEA’s updated central projection puts global data-center electricity consumption at roughly 950 TWh in 2030, compared with 485 TWh in 2025. These figures cover all data centers, not AI alone. [8]

The IEA flags energy-equipment constraints and slower electricity investment. [8] My implication: contracted power access deserves scrutiny alongside chip orders. For utilities, shareholder returns depend on tariffs, cost allocation and execution. A bigger load does not automatically mean a better stock.

For credit investors, I would ask: does interest accrue before the site earns revenue? Who absorbs a construction delay? Is the customer contract firm enough to support debt service? Those questions connect a technology story to familiar infrastructure and project-finance risks.

05 / THE REVENUE AT THE ENDAI as a tool: who pays for the productivity?

Infrastructure ultimately needs a paying use case. AI may help employees draft documents, search information, assist customers and write code, but the economic value depends on how much useful work gets completed after checking the output. Time saved is not automatically revenue earned.

The ILO’s 2025 assessment found that one in four workers globally was in an occupation with some generative-AI exposure. It emphasized transformation of jobs, rather than wholesale redundancy, because human input remains necessary. Exposure is not a predicted layoff rate. [13] An NBER study of customer-support work found roughly 14% higher productivity with AI assistance, with benefits concentrated among less-experienced workers. That is evidence from a particular setting, not a universal productivity assumption. [14]

My view for finance is that AI’s strongest near-term role is assisting research and preparation while people retain responsibility for valuation, execution and client judgment. The investment question is who captures the benefit: the employer through lower costs, the worker through higher output, the customer through lower prices, or the AI provider through subscription and usage fees. These outcomes imply very different returns on the same infrastructure.

06 / FORWARD VIEWPredictions need an assumption, not just a target

Nvidia guided to $108 billion of revenue, plus or minus 2%, for Q3 FY2027. That is management’s outlook, not a realized result. [1] Meta’s quarterly guidance and valuation sensitivity are covered separately in Note 004.

Microsoft’s July call put calendar-2026 capex at approximately $175 billion after a shift from finance to operating leases changed its earlier $190 billion expectation. Accounting presentation matters: a smaller capex headline need not represent a proportionate reduction in economic spending. [15] Separately, Yahoo Finance reported Evercore ISI estimates of Amazon capex at $320 billion in 2027 and $370 billion in 2028, with roughly $50 billion of negative free cash flow in each year. That is one analyst’s outlook—not consensus, management guidance or my forecast. [12]

07 / THE S&T LENSTrade the cash-flow chain

I would organize this theme around relative performance and credit quality rather than a single “AI is bullish” call. Strong AI revenue can help one company while the corresponding spending constrains another’s free cash flow. The price paid for exposure—and the expectations embedded in it—still determine the trade.