Imagine a restaurant that keeps adding tables. More seats, more reservations, more people queuing out the door. The kitchen is fine. The chefs are fast. The problem is the gas line feeding the stoves, which was installed when the place had a dozen covers a night and now needs to serve four hundred. The food isn’t the bottleneck. The fuel is.
That’s roughly where the AI buildout sits right now, and Morgan Stanley just put a number on the gas line.
A 32-gigawatt gap
The brokerage estimates a 34% net power shortfall for US data centers through 2028, equivalent to about 32 gigawatts. That figure already accounts for the workarounds the industry has been scrambling to put in place, including behind-the-meter generation (power produced on-site, bypassing the public grid) and fuel cells. In other words, this isn’t the number before anyone tried to fix it. It’s the number after.
If you’ve never thought about electricity in gigawatts, the unit itself isn’t the point. The gap is. Roughly a third of the power the industry expects to need simply may not be there when the buildings are ready for it. Concrete poured, racks installed, cooling plumbed, and then a wait for electrons.
For anyone who uses AI tools and agents without thinking much about the machinery underneath, this is the part worth understanding: the constraint on AI has quietly moved. For a few years the story was chips. Can anyone get enough GPUs? Now the story is increasingly about whether there’s power to run them.
Why the giants stay dry
Here’s the part that surprised me. You’d assume a power shortage this size would hit the chip designers hardest. Morgan Stanley argues the opposite for Nvidia and Broadcom, and doesn’t see these bottlenecks threatening either company’s 2027 forecasts.
The reasoning comes down to information and flexibility. Both companies have visibility into where their chips are actually being placed, which means they know which sites are powered and which aren’t. They can expand geographically, shifting volume toward regions where power is available instead of waiting on a single constrained grid. And they coordinate directly across the chain, between data center operators, chip suppliers, and the power supply chain itself.
Think of it as being the chef who gets to see every restaurant’s gas line before deciding where tonight’s deliveries go. If one kitchen can’t cook, the shipment moves to one that can. The order doesn’t get cancelled. It gets redirected.
The companies without that view
Someone does absorb the shock, though. Morgan Stanley flags makers of memory, optical, and other secondary chip components as the ones exposed if AI deployments slip. Power-management and analog suppliers sit in the same category, facing greater inventory risk.
These are the parts that make a finished AI server actually work:
- Memory — the chips that hold the data a model is working with
- Optical components — the parts that move information between machines at high speed, often over light rather than copper
- Power-management and analog chips — the quieter components that regulate voltage and keep everything stable
These suppliers typically build to forecast, shipping into a demand picture they don’t fully control. If a data center’s energization date slides six months, those components are already made, already boxed, and already sitting somewhere. That’s what inventory risk means in practice: the right parts at the wrong time.
What this means if you just use the tools
You’re not going to notice a gigawatt shortfall. You’ll notice its side effects, and they’ll be subtle.
AI agents are power-hungry in a way that chatbots weren’t. A single question gets a single answer. An agent that researches, writes, checks its work, and tries again runs many steps for one request. Multiply that across millions of users and the electricity bill is the real constraint on how generous these products can afford to be.
So when a provider caps your usage, puts the better model behind a higher tier, or routes your request to something smaller and faster, physical limits are often part of the explanation. Not strategy alone. Infrastructure.
Morgan Stanley isn’t alone in flagging this. Goldman Sachs has pointed at mounting constraints on the US data-center buildout too, expecting limited near-term impact from political pushback, while Morgan Stanley points to labor and power as the binding issues.
What I take from this is a reframing of how AI progress actually gets limited. The interesting question stopped being whether the models keep improving. It became whether the grid can keep up with them, and which companies in the chain have enough visibility to route around the gap. Nvidia and Broadcom apparently do. The suppliers one layer down are the ones watching the calendar.
🕒 Published: