Can AI-driven profit growth outweigh the pressure from higher interest rates and expensive energy and how much of that success is already priced in?

Suppliers collect first; infrastructure owners must earn their investment back

The transaction explains the tension. A cloud operator buys equipment before selling the services that equipment will support. The supplier recognizes a sale; the buyer commits cash and subsequently bears depreciation, operating costs and financing obligations. Strong supplier economics can therefore coexist with a long wait for the buyer’s investment to pay back.

Nvidia provides the clearest evidence of supplier monetization. Its data-center revenue reached $89 billion in the quarter ended July 26, 2026, up 117% year over year. Its consolidated gross margin was 75%. Those disclosures show substantial demand and company-wide gross-profit capture; the customer’s return depends on what happens after the equipment arrives. [3]

AWS shows that cloud services can also generate substantial operating profits. Its $16.6 billion of June-quarter operating income represented a $6.4 billion increase, or 62.7%, calculated against the prior-year $10.2 billion. Dividing operating income by $42.2 billion of sales gives a 39.3% operating margin. These whole-segment results include established non-AI workloads. [1]

Our view is that this earning power makes continued profit growth credible enough to test against cost shocks. But the investment question requires the costs and profits of the same businesses over the same period. Whole-cloud earnings cannot supply an AI attribution, and a supplier’s margin cannot supply its customer’s payback.

The next test is cash conversion as the new assets enter service. Amazon’s negative trailing free cash flow and AWS’s rising quarterly operating income illustrate different parts of that process. Their differing company and segment scopes matter: the cash measure captures Amazon’s broader investment demands, while the operating result describes AWS’s current profitability. [1]

Meta makes the cash demand particularly visible. In the June quarter, cash property-and-equipment purchases rose to $30.1 billion from $16.5 billion, while free cash flow fell to $784 million from $8.55 billion. Calculated from those corresponding quarters, spending increased 82% and free cash flow declined 91%. The immediate cash burden is clear; lifetime investment returns depend on the earnings those assets subsequently produce. [4]

Accounting presentation also deserves attention. Microsoft’s fiscal fourth-quarter call explained that finance leases enter its capital-expenditure measure while operating leases do not. Shifting the classification adjusted its calendar-2026 capex expectation to approximately $175 billion while leaving the underlying investment expectation otherwise unchanged. Our view is that investors should follow the economic commitments through the presentation change. [5]

Higher rates reach borrowers gradually and valuations immediately

Financing adds a second claim on the cash generated by the infrastructure. Its importance depends on who owns the assets, how they were funded and when the debt reprices. An existing fixed-rate bond and a new construction loan respond differently to the same Treasury market.

CoreWeave shows how large that claim can become. In the June quarter, it reported $2.575 billion of revenue, $640 million of net interest expense and a $626 million GAAP net loss. Interest expense divided by revenue was 24.9%. We give financing risk greater weight for externally financed builders than for businesses with established recurring cash generation. [6]

The market benchmarks are demanding. The ten-year nominal Treasury yield was 5.31% on October 5 and the ten-year real yield was 2.95%; the BBB corporate effective yield was 6.23% on October 2. These dated market observations describe the terms available to capital today. Existing borrowers encounter them as funding needs arise and debt matures. [7] [8] [9]

The maturity schedule puts the timing into perspective. Reuters’ analysis of LSEG data identified approximately $572 billion of nonfinancial corporate bonds issued in US markets maturing in 2027. Refinancing that entire cohort at an additional one percentage point would add $5.72 billion of annual pretax expense once the refinancing was complete: $572 billion multiplied by 1%. The cohort extends beyond the S&P 500, and the expense would emerge over time. [10]

Our view is that borrowing-cost analysis should remain issuer-specific. It needs debt maturities, floating-rate exposure, cash interest income and funding requirements. Applying the latest yield to all outstanding debt would obscure those differences.

Share prices can respond much faster. At unchanged forward earnings, a move from 19× to 17× reduces price by 10.5%; a move to 16× reduces it by 15.8%. Those are direct consequences of paying less for each dollar of earnings. We use them as valuation stresses, without assigning a mechanical P/E response to any particular change in Treasury yields. [2]

Electricity prices affect margins; power connections determine when revenue can begin

Electricity introduces two separate economic questions. An operating facility pays for each kilowatt-hour it consumes. A facility waiting for a connection may postpone the output and revenue expected from equipment already purchased. A tariff model answers the first question; a commissioning schedule helps answer the second.

To make the operating exposure tangible, consider one gigawatt of capacity measured at the facility meter. We assume a 50–70% average load and electricity prices of 10–15 cents per kilowatt-hour. These endpoints provide a transparent sensitivity to utilization and price, rather than an estimate for a particular operator.

The calculation produces annual electricity bills of approximately $438–920 million. A further two-cent tariff increase adds $88–123 million. Its use is to translate an exposed volume into dollars. Assessing a cloud company requires its actual consumption, contracts, hedges, repricing terms and ability to pass costs to customers.

A delayed connection works differently: it can extend the period during which invested capital earns no operating return. Our view is that commissioning schedules deserve scrutiny alongside tariffs. The available evidence leaves the portfolio-wide value of delayed revenue unresolved, so we cannot rank power delivery as the universal constraint on AI profits.

The multiple looks ordinary because the earnings expectations are strong

That brings the question back to price. On September 30, the S&P 500 closed at 7,651.54 and traded at 19× next-twelve-month earnings, according to FactSet’s October 2 report. The ten-year average forward multiple was 19.1×. By that comparison, the market’s forward valuation looks ordinary. [2]

The earnings denominator carries considerable optimism. FactSet projected 32.4% calendar-2026 earnings growth and another 15.8% in 2027. The first figure combines already-reported quarters with the remaining forecast; the second extends the expected expansion. Our view is that strong earnings expectations accompany the price even though a distinct AI premium cannot be isolated. [2]

Dividing the matched index level by the rounded forward multiple implies approximately $403 of forward EPS and a 5.26% earnings yield. That yield is roughly level with the nearby 5.31% nominal Treasury observation. The comparison makes growth important to the equity proposition. An equity earnings yield represents risky profits, some retained for investment, so the yield gap alone cannot price the risk of ownership. [2] [7]

The quality of earnings also matters. FactSet reported 118.5% second-quarter earnings growth for the Magnificent Seven, boosted by investment gains. It identified $98 billion of Alphabet other-income gains and $53.4 billion at Amazon. A recurring AI operating-profit thesis needs a different foundation from investment revaluations. [11]

We therefore frame “how much is priced in?” as an earnings requirement. Start from the observed price and forward earnings, specify the return sought, and ask what future earnings would support it at different multiples. This makes the dependence on valuation explicit.

At 17×, the return objective requires approximately $550 of terminal forward EPS; at 16×, approximately $585. Those outcomes require 11.0% and 13.2% annual earnings growth respectively. With 10% annual EPS growth and the same dividend assumption, modeled annual returns fall to approximately 7.1% at 17× and 4.9% at 16×. Successful earnings growth can coexist with modest shareholder returns. [2]

The counterargument deserves equal weight. Multiple compression is optional, not inevitable. If 19× persists, the same return objective requires approximately $492 of terminal forward EPS, or 6.9% annual growth. Nvidia’s realized demand, AWS’s operating profits and the historically ordinary index multiple give the bullish case substance. Our caution rests on cash conversion and the price paid for future earnings, rather than a belief that the technology must disappoint. [2] [3] [1]

The next evidence must connect adoption to cash

Productivity gains could broaden the earnings base beyond suppliers and infrastructure owners. Bank of America offers a concrete operational example: its April 2025 disclosure said more than 90% of employees used Erica for Employees and that the assistant had reduced IT service-desk calls by more than 50%. Converting that improvement into profit depends on staffing, redeployment and the cost of running the service; the disclosure supplies the operational result. [12]

Our view would become more constructive with several quarters of improving cash generation after cash capex and lease payments as capacity enters service, supported by paid customer demand. Persistently weak cash conversion while investment rises would push us the other way. The distinction favors examining each owner’s funding and utilization rather than treating every AI beneficiary alike.

The unresolved economic test is precise: incremental AI contribution for a defined group of companies, after ordinary operating costs, compared with additional financing and energy expenses outside that profit baseline. It requires the same geography, period and corporate perimeter on both sides. Until that comparison is available, a sector-wide offset remains plausible rather than measured.

The valuation test is already visible. A credible trajectory toward $550–585 of terminal forward index EPS would meet the illustrated return requirement at 17–16×. Conversely, a 10% reduction in the starting forward EPS estimate combined with a reset to 17× would imply approximately 19.5% price downside before dividends, calculated as 0.90 × 17 / 19 − 1. These are concrete conditions against which to monitor the investment case. [2]

The next proof must arrive as cash after the build—and earnings sufficient for the price paid.