Morgan Stanley: $1.6 Trillion AI Infrastructure Investment Faces Triple Downside Risks
nashnova research
Morgan Stanley warns that America's $1.6 trillion AI infrastructure plan faces a triple squeeze — politics, power, and pricing cycles — and if all three compound, the entire capex cycle could reverse at what the bank calls a "catastrophic" pace.
Where exactly is the $1.6 trillion AI buildout getting stuck?
Strategist Ariana Salvatore boils the constraints down to three words: people, power, and politics.
The bank still holds a constructive base case on AI infrastructure, but flags that it now comes with a growing list of preconditions.
This means → Morgan Stanley is not bearish on AI itself — it is saying the uncertainty around *whether this gets built* is far larger than the headline number suggests.
Washington can't stop it — so who can?
A federal blanket ban is unlikely. National-security framing and presidential veto power shield the buildout at the top.
The real chokepoint is state governments and utility commissions: grid interconnection approvals, large-load tariff design, siting permits, and tax breaks are all local decisions.
The debate has shifted from "whether to restrict" to "how to restrict." In plain terms = it is no longer about building or not building — it is about who writes the rules and who bears the cost.
Arizona has already frozen data-center tax exemptions through June 2029. States flagged by Morgan Stanley account for nearly 60 GW of data centers under construction or planned — Texas alone holds 24 GW.
"Volume and price fall together" — the risk the market is ignoring?
Morgan Stanley calls the pricing cycle the least-watched of the three risks. The current phase is a classic scarcity loop: buyers expect higher prices tomorrow, so they pay a premium today; suppliers stack margins on top, reinforcing the spiral.
Once supply catches demand, the loop reverses fast. The bank's analogy: the late-2022 memory downturn, when shipments fell only about 5% but prices crashed 20–30% — margin collapse far outran the volume decline.
This means → Amazon has said its 2026 capex increase is driven mainly by rising memory costs, not expanded construction. A meaningful share of today's capex "boom" is price-inflated — and memory is the component most likely to turn first.
Not enough power — how big is the gap?
Morgan Stanley's proprietary "AI power-supply model" shows that even after accounting for nuclear, gas turbines, engines, and fuel cells, the U.S. faces a 32 GW net shortfall by 2028 — roughly 34% of the 97 GW needed. By 2029 the gap widens to about 72 GW.
This reflects a timing mismatch, not a demand collapse. Hyperscalers can keep buying GPUs and stockpiling them in unpowered shells, but "continually purchasing power-dependent equipment that cannot be turned on is ultimately unsustainable."
Risk varies by supply-chain segment: memory, networking, and optics ship with the rack and carry the highest exposure; electrical and cooling gear faces slower backlog conversion from permitting delays; the relative winners are behind-the-meter generation (gas turbines, modular nuclear) and already-energized shell providers.
How do the three risks feed into each other?
The transmission chain is clear: political resistance slows permitting → the power gap widens → chips bought at peak prices sit idle → volume-and-price decline is triggered.
In plain terms = a buildout that must pass the ballot box, the grid, and the bond market all at once will stall if any single link breaks — and it drags the other two down with it.
Industry is not standing still: several hyperscalers have signed "ratepayer-protection pledges," self-funding generation and grid upgrades; Google has committed to replenishing more water than its data centers consume by 2030; Meta has set up a $1 billion fund for host-community infrastructure.
市场有风险，内容仅供研究参考，不构成投资建议。