AI | Crisis Research

Vertiv Holdings Co VRT

Data-center power and cooling

AI: demand is real, but the return on capacity is unproven

AI is an investment chain: users and businesses pay for applications; model companies buy compute; cloud providers build capacity; chip, networking, cooling, and power suppliers sell the equipment. The central question is whether revenue and useful output will grow enough to earn a return on that capacity. A contract announced upstream is not the same as a profitable, fully utilized data center downstream.

Our prediction

Through 2028: AI use and revenue will keep growing, but installed compute and power commitments will outrun profitable demand for many projects. The strongest distribution platforms and low-cost infrastructure operators will earn returns; weaker projects will be repriced, delayed, or consolidated. This is a split market, not a prediction that AI demand disappears.
What would confirm it: Rising paid usage together with uneven data-center utilization, renegotiated capacity commitments, and widening cash-return gaps between suppliers and operators.
What would prove us wrong: Broadly high utilization and durable positive cash returns across new capacity, with disclosed AI revenue covering depreciation, power, and financing costs.

What is observable

Evidence	Dated fact	What it cannot prove
Electricity demand	A 2024 DOE/Lawrence Berkeley study estimated data centers used 176 TWh, or 4.4% of U.S. electricity, in 2023 and projected 325–580 TWh in 2028.	This is a scenario range, not a measured 2026 load or a guarantee every planned campus will connect to the grid.
Private AI funding	OpenAI said it closed a $122 billion financing round in March 2026 and described broad cloud and chip partnerships. Anthropic announced a $65 billion financing in May.	Funding provides capacity to spend; it does not establish audited operating cash flow or future public-market returns.
Commercial traction	OpenAI's March statement said it was generating $2 billion in revenue per month and more than 40% of revenue came from enterprise.	These are company-reported run-rate and mix claims, not a complete public filing with margins, retention, or cash-flow notes.

The bottleneck can move. A chip shortage can shift toward power and interconnection; abundant capacity can shift the question toward utilization and pricing. Model efficiency is double-edged: lower cost per task could improve margins and expand demand, but it could also make earlier hardware forecasts too high. The correct unit of analysis is paying usage relative to total cost, including power, depreciation, networking, and model training.

Where listed stocks fit

NVDA supplies accelerators; MSFT, AMZN, and GOOG sell cloud capacity and AI services while also funding infrastructure; VRT supplies power and cooling equipment; ETN has electrical-infrastructure exposure; CEG has power-supply exposure. These are not identical bets. A new model-company contract might help a chip vendor immediately while creating a long payback period for a cloud operator. Data-center orders can also lead realized end-customer demand by years.

What could change the view

The positive case is rising paid usage and enterprise retention, higher utilization, and falling cost per useful task. The negative case is that equipment and power commitments become fixed before customers generate sufficient cash return. To distinguish them, we would compare hyperscaler capex with disclosed AI revenue, operating cash flow, depreciation, capacity delivery, and grid connection dates. We would also check whether supplier backlogs convert to revenue without margin erosion.

Current view: The demand signal and funding are large, but the evidence available here does not establish that the entire AI buildout earns an attractive return. Published company results and physical power delivery should take priority over announced capacity totals.