AI & Technology
US AI Infrastructure and Capital Markets: Follow the Cash Flow
The investment question is who converts AI spending into durable cash flow, at what price, and with which risks left on the balance sheet.
Research question: How does US AI infrastructure investment flow through the real economy and capital markets, and where can a convincing growth story fail to produce attractive returns?
Short answer: AI spending supports a chain that includes computing hardware, data centres, electrical equipment, cooling, generation and financing. Each participant earns money at a different stage. Capital expenditure, contracted capacity, electricity consumption, revenue and free cash flow are distinct measurements. Our analysis starts by reconciling them.
The electricity outlook has changed: use the current report
Berkeley Lab's 2025 Update, published in June 2026, estimates US data centre electricity consumption at 192 TWh in 2024. Its reference case reaches 649 TWh in 2030, with a compounded uncertainty range of 521–843 TWh. The reference case represents 11.8% of projected US electricity consumption in 2030. These figures cover data centres overall, not AI-only facilities, and the future values are modelled scenarios. 1, 2
The report's name can mislead a quick reader: “2025 Update” is the title, while June 2026 is its publication date. It updates an earlier report. We do not splice an old estimate of the historical baseline into the newer forecast and present the combination as one series.
Our reading of the range is that electricity access is an important underwriting variable, while the exact demand path remains uncertain. A project that works only at the highest consumption scenario is making a much stronger bet than a project supported by an existing customer's committed demand.
A company example: capex is not the same as cash spending
Microsoft's July 29, 2026 earnings call reported approximately $41 billion of capital expenditures for its fiscal fourth quarter, $5.6 billion of finance leases, and $35.8 billion of cash paid for property, plant and equipment. It said roughly two-thirds of quarterly capex was for shorter-lived assets, primarily CPUs and GPUs. These are company-wide disclosures, not a US-only or AI-only spending measure. 3
Its cash-flow statement reports $115.948 billion of additions to property and equipment for the year ended June 30, 2026, versus $64.551 billion for the preceding fiscal year. 4 That cash-flow line should not be added to an overlapping capex measure as though it represented another pool of spending.
| Measure | What it helps explain | A common analytical error |
|---|---|---|
| Capex including finance leases | Asset commitments recognised under the company's presentation | Treating the full amount as a same-period cash outflow |
| Cash paid for property and equipment | Cash used for the specified asset purchases | Assuming it includes every future lease payment |
| Annual or quarterly guidance | Management's expectation for a defined period | Calling it money already spent |
| Remaining lease or purchase commitments | Future contractual obligations | Adding overlapping commitments to current capex without reconciliation |
Microsoft also said that a useful-life change affecting buildings would move more future data centre leases from finance to operating classification, changing reported capex while leaving its underlying calendar-year investment expectations unchanged apart from the classification effect. 3
The implication is analytical rather than promotional: a lower capex headline can reflect a presentation change rather than a reduction in physical investment. Conversely, more reported spending does not by itself demonstrate improved economics. Lease cash payments, depreciation, maintenance needs and customer collections must still be examined.
Map the chain from commitment to cash collection
An infrastructure announcement travels through several gates before it produces operating cash flow:
- Site and permissions: land, construction approval and appropriate operating rights.
- Power and equipment: a credible connection date and delivery of the required electrical, cooling and computing systems.
- Commissioning: a facility that can actually run to the agreed specification.
- Customer usage and billing: delivery under contracts that establish price, availability and remedies.
- Cash collection: payment received after operating costs and required investment.
These stages should not be collapsed into “megawatts announced.” Secured land is not a commissioned data centre, and commissioned capacity is not necessarily occupied capacity. The timing gap matters to lenders and equity investors because interest and fixed costs can start before customer revenue.
Which parts of the capital market carry which risks?
The following is our research framework, not a recommendation of a security or a prediction that a particular segment will outperform.
| Exposure | Route to potential cash flow | Main questions to resolve |
|---|---|---|
| Computing and networking suppliers | Equipment sales, software and support | Order durability, customer concentration and the replacement cycle |
| Electrical and cooling suppliers | Delivery of power-distribution and thermal systems | Backlog conversion, capacity expansion and margins after shortages ease |
| Data centre owners and operators | Leasing, hosting or service revenue | Energisation timing, utilisation, customer credit and expansion funding |
| Utilities and generators | Electricity supply and approved infrastructure investment | Tariffs, who funds upgrades, regulatory approval and operating costs |
| Infrastructure lenders | Interest, fees and principal repayment | Completion risk, collateral value, covenants and refinancing access |
An attractive sector can still contain expensive assets. Scarcity may support supplier pricing today but attract capacity that reduces future margins. A long customer contract may help an operator finance construction but also leave it exposed to technology changes or costly service obligations.
Utilities require particular care. Higher electricity sales do not automatically translate into higher returns on equity. The economics depend on the regulatory arrangement, investment required to serve the load, and allocation of upgrade costs between large customers, shareholders and other users.
A worked example: what a power-price change can do
Assume a facility has an average metered draw of 10 MW, maintained for 8,760 hours per year. This is an illustrative, constant-load calculation; it is not a project forecast or a claim about average industry utilisation.
Annual consumption is 10 × 8,760 = 87,600 MWh, or 87.6 million kWh. At $0.08 per kWh, the energy charge would be $7.008 million. At $0.12, it would be $10.512 million: an increase of $3.504 million before demand charges, taxes, grid fees or other costs.
We use the metered facility load, so a separate power-usage-effectiveness multiplier is not added. If the starting figure were IT load instead, overhead assumptions would need to be specified. This distinction prevents counting cooling and other facility consumption twice.
Who bears the additional $3.504 million? A pass-through arrangement can shift it to a customer, but then customer credit and competitiveness matter. A fixed-price contract can leave it with the operator. The same building and the same growth narrative can therefore create very different equity outcomes.
What we would monitor next
- Timing: the difference between promised and achieved energisation dates, and the cash cost of delay.
- Economics: customer collections, utilisation, pricing and replacement investment, rather than capacity alone.
- Balance sheets: lease obligations, financing maturities and customer concentration.
- Forecast discipline: revised demand scenarios and evidence that efficiency gains change consumption or unit costs.
- Valuation: how much favourable execution is already required by the entry price.
Our thesis would weaken if demand fails to materialise at commercially viable prices, if customer contracts prove less durable than expected, or if financing and replacement costs absorb the operating cash generated. Stronger compute efficiency can also change the distribution of value even while usage grows.
Questions readers ask
Does more AI infrastructure spending guarantee higher stock prices?
No. Spending is an input. Shareholder outcomes depend on future cash flow, funding, dilution, risk and the valuation already paid for growth.
Are the electricity forecasts solely for AI?
No. The Berkeley Lab figures cited here describe total US data centre electricity use. The report models different equipment and operating assumptions; the forecast range is not a guaranteed outcome. 1, 2
What is Lumiana Global's conclusion?
The useful research unit is the contract and balance sheet attached to an infrastructure asset. Follow spending through delivery, power, customer usage and collection. That is how an industrial investment story becomes an assessable capital-markets proposition.
Related research: ADIA's portfolio shifts · Africa emerging markets · AI-native service firms.
Sources and methodology
Public-source desk research, information cut-off October 24, 2026. Forecasts are attributed to their issuers; the cash-flow framework and electricity-cost example are Lumiana Global's analysis. Company reporting periods and definitions are retained. This note is informational and is not personalised investment advice.