Morgan Stanley: Power Supply Bottleneck Has Become the Biggest Constraint on AI Infrastructure Expansion
N.R. Finch
Morgan Stanley's latest research report warns that power supply has escalated from a minor friction to the core bottleneck for AI infrastructure — transformer lead times have ballooned from 12 weeks pre-COVID to 128 weeks, meaning real-world AI buildout will lag far behind the pace that capex forecasts imply.
How severe is the power gap?
Power transformer lead times now average 128 weeks; generator step-up transformers stretch to 144 weeks. Pre-COVID, the norm was 12 to 16 weeks. This means → just waiting for the hardware takes nearly two and a half years — far longer than building the data center itself.
In plain terms = you can erect a data center in months, but the equipment to power it sits in a queue for over two years. Power is now the slowest link in the entire chain.
Grid interconnection is equally jammed: the U.S. backlog of energy projects awaiting grid connection already exceeds twice the country's total installed capacity — and is still growing.
How is the power bottleneck reshaping AI financing?
Power availability has overtaken land and connectivity as the top criterion for data-center siting. Developers are co-locating generation and compute; "off-grid" solutions — fuel cells, gas turbines, battery storage — and converting Bitcoin mining farms into data centers for "fast power-on" have all entered real construction pipelines.
AI companies are increasingly acquiring or contracting power assets directly, rather than leaving them to utilities. This means → the traditional boundary between AI-infrastructure financing and energy financing is dissolving; two formerly separate capital pools are merging into one.
Credit markets already reflect this: recent investment-grade and high-yield deals are pricing energy assets and compute assets within the same framework.
What other bottlenecks are stacking up beyond power?
Labor shortage: the U.S. is projected to be short roughly 300,000 electricians over the next decade, and more than one-fifth of the current workforce is already 55 or older. In plain terms = even when equipment arrives, there may not be enough hands to install it.
Water stress: 43% of the world's data centers sit in high-water-stress regions, raising sustainability questions about further expansion at major hubs.
Policy headwinds tightening on multiple fronts: New York is moving to impose a moratorium on new data-center projects; Texas Governor Abbott has ordered regulators to ensure data-center power demand does not raise bills for other users; 14 state legislatures across the U.S. are currently considering some form of data-center restrictions.
What is Morgan Stanley's bottom line?
These overlapping constraints are deepening a structural imbalance between compute supply and demand. This reflects a signal beneath the surface: the AI story has no shortage of demand — what it lacks is the physical world's ability to deliver.
In this scarcity regime, companies with scaled, reliable compute — what Morgan Stanley calls "compute merchants" — will command sustained pricing power.
Put simply = AI capex numbers can be enormous, but the capacity that actually gets built and energized will likely fall well short of market expectations. Money isn't the bottleneck — power is.
Content is for reference only, not financial advice.