What if the thing that finally slows down the AI buildout isn’t a model plateau, a regulation, or a funding freeze, but an electrical shortfall nobody can code their way out of?
Morgan Stanley just put a number on it. The brokerage estimates a 34% net power shortfall for US data centers through 2028 — roughly 32 GW — and that figure already accounts for workarounds like behind-the-meter generation and fuel cells. In other words, that’s the gap after the clever fixes. Thirty-two gigawatts of demand with nowhere to plug in.
And yet the firm’s read on Nvidia and Broadcom is a shrug. Neither company’s 2027 forecasts look at risk, according to the note. The reasoning: both have visibility into where their chips are actually going, they’re expanding geographically, and they coordinate directly with data centers, semiconductor suppliers, and the power supply chain. They see the bottleneck forming before it forms, and they route around it.
Why This Matters to Anyone Reviewing AI Tools
I spend my days testing agents and AI products, and the single most consistent complaint I hear from builders isn’t about model quality. It’s about capacity. Rate limits. Queue times. Regions that are “temporarily unavailable.” Enterprise tiers that quietly cap you at a fraction of what the marketing page promised.
Every one of those frustrations traces back through a chain that ends at a substation. When Morgan Stanley says delays in AI deployments could hit makers of memory, optical, and other secondary chip components, that’s a supply-chain observation. But the downstream version of that sentence is the one you’ll actually feel: the agent platform you’re evaluating can’t get the capacity it promised, so your workloads sit in a queue.
The vendors won’t phrase it that way. They’ll call it “optimizing for efficiency” or “intelligent routing.” What it often means is that compute is scarce and you’re being rationed.
Two Tiers, Not One Ecosystem
The more interesting detail in this note is the asymmetry. The AI chip stack is usually discussed as a single rising tide — Nvidia goes up, everything connected to Nvidia goes up. Morgan Stanley is describing something closer to a hierarchy.
- Top tier: Nvidia and Broadcom, insulated by direct coordination with the people building the power infrastructure and by knowing where every chip lands.
- Everyone else: memory, optical, and secondary component suppliers who ship into that chain and absorb the timing risk when a data center slips a quarter.
That distinction is a useful lens for reading vendor claims. A company whose product depends on a tight relationship with the accelerator makers is in a very different position from one that depends on broad component availability. Both might tell you their roadmap is solid. Only one of them has the visibility to know.
The Part That Should Make You Skeptical
I’m not going to pretend a bank note is prophecy. Morgan Stanley is modeling a shortfall three years out, and three-year models in this space have a mediocre track record. The firm has also flagged labor and permitting constraints alongside power, and Goldman Sachs has raised its own concerns about the data-center buildout while expecting limited near-term impact from political pushback. These are estimates, not measurements.
What I’d treat as more reliable is the shape of the problem rather than the magnitude. Power plants take years. GPUs take months. That mismatch doesn’t require a forecast to believe — it’s just how construction timelines work compared to semiconductor fabs. The 32 GW figure might be wrong by a wide margin in either direction. The structural asymmetry won’t be.
What I’d Actually Do With This
If you’re buying or building on AI infrastructure right now, a few practical moves follow from this:
- Ask vendors about capacity guarantees in writing, not about model benchmarks. Benchmarks are easy to publish. Capacity commitments cost something.
- Assume your provider’s roadmap slips. Build fallbacks across providers rather than optimizing hard for one API surface.
- Treat “unlimited” tiers as aspirational. Scarce compute eventually shows up as a limit somewhere in your bill or your latency.
- Watch where new regions open. Geographic expansion is the tell for who actually has power secured.
The AI tools market has spent two years selling intelligence as if it were software — infinitely copyable, instantly scalable. Morgan Stanley’s note is a reminder that it’s closer to heavy industry. It needs turbines and transformers and transmission lines, and those things don’t ship overnight.
Nvidia and Broadcom appear to have figured that out early enough to plan around it. The question worth asking your vendors is whether they did too, or whether they’re just hoping the lights stay on.
🕒 Published: