Morgan Stanley: U.S. Data Center Power Gap Reaches 34%, Divergent Impact Across Chip Supply Chain
nashnova research
Morgan Stanley estimates a net 32 GW power shortfall for US data centers from 2026 to 2028 — 34% of total demand — making electricity, not chip supply, the top bottleneck for AI compute expansion, with sharply divergent effects up and down the semiconductor chain.
How big is a 34% power gap?
Total US data center power demand gap for 2026–2028 is 57 GW. After subtracting behind-the-meter generation (BTM — small power plants built on-site) and fuel cells, the net shortfall is still 32 GW.
This means → for every three watts AI compute needs, one watt has no reliable source today.
The gap widens each year: 9 GW in 2026 → 15 GW in 2027 → 32 GW in 2028 → 68 GW in 2029. In plain terms = the problem is not narrowing — it is accelerating.
Who is consuming all that power?
Next-generation high-power GPU racks — such as Nvidia's NVL72 — deploying at scale are the core driver of surging demand.
Hyperscalers — Amazon, Google, Microsoft, Meta — account for 60% to 95% of new data center capacity, with Google adding the most operational capacity.
This means → top players leverage credit quality and purchasing scale to dominate the race for power; smaller cloud operators and less-efficient chip makers risk being squeezed out.
Why are Nvidia and Broadcom safe — for now?
Nvidia and Broadcom together hold roughly 90% of the AI XPU market. Their combined 12-month AI revenue guidance is about $800 billion, and that guidance faces no material threat from the power shortfall so far.
Morgan Stanley cites four supports: management guidance already prices in power risk; full visibility on chip deployment sites; active build-out of overseas compute nodes; and Nvidia's compute output per gigawatt is significantly higher than peers.
In plain terms = given the same watt of electricity, an Nvidia GPU produces more results — the tighter power gets, the more that advantage is worth.
Who gets hit hardest by the power crunch?
ASIC chips — custom silicon designed for specific tasks — produce fewer tokens per watt, so their market share risks being squeezed by GPUs in a power-constrained world.
Lower-value segments — memory, optical modules, power management, analog components — are most exposed to the bullwhip effect (demand swings amplified down the supply chain).
This means → if compute deployment slows, customers cut these "accessory" orders first, leaving those suppliers with the highest revenue volatility.
Can building overseas close the gap?
Morgan Stanley's verdict: overseas expansion cannot fill the domestic US shortfall. Europe is constrained by power and permitting; the Middle East carries geopolitical risk; Asia is projected to absorb only 14 GW of US overflow demand by 2030.
The bank accordingly cut its US share of global compute from 60% to 55%.
This reflects a forced redistribution of the compute map driven by power constraints — but the pace of dispersal lags far behind the widening gap.
What can actually narrow the shortfall?
Behind-the-meter generation (BTM) is the leading incremental source: in the base case, gas turbines and engines can contribute 19 GW from 2026 to 2028; in the bull case, up to 49 GW.
Delivery is uncertain, however — skilled labor shortages, engineering complexity, and local permitting all constrain actual build-out. Morgan Stanley models three scenarios: bull case narrows the gap to 15 GW; bear case widens it to 42 GW.
In plain terms = the ability to secure power and maximize efficiency is now the decisive variable in the AI race — it is not about who has the best chip, but who can get the electricity.
市场有风险，内容仅供研究参考，不构成投资建议。