AI's New Bottleneck Is Electricity: 32GW Short by 2028, Says Morgan Stanley
October 7, 2026 · 5 min read
The AI buildout has a new bottleneck, and it isn't silicon. Morgan Stanley now estimates that even after behind-the-meter generation and fuel cells are counted, US data centers will still be short 32 gigawatts of power by 2028 — a 34% gap against projected demand. For two years the industry's question was "can we get enough chips?" Now the question is "can we plug them in?" (via Reuters)
Sit with that number for a second: 32GW, 34% of what the country is expected to need. This is the supply line nobody priced in. You can tape out accelerators and pre-order racks all day, but if the grid can't deliver the electrons, the project schedule slips. And a slipped schedule in this business means deferred revenue, renegotiated contracts, and capacity payments burning cash while the meters spin at zero.
From card-constrained to power-constrained
This shift rewrites the whole playbook. The 2023–2025 story was allocation — who gets the GPUs, at what markup, with what lead time. The next chapter is permitting and interconnection queues: substations, transformers, and utilities that approve projects on timescales measured in years, not quarters. The industry spent a fortune solving the chip shortage. The power shortage doesn't respond to purchase orders. You can't bid your way out of a transformer lead time, and there's no secondary market for a signed interconnection agreement.
Why Nvidia and Broadcom get a pass
Morgan Stanley's report draws a line inside the supply chain. Nvidia and Broadcom, it says, are relatively shielded from the crunch — the analysts left their 2027 expectations for both companies unchanged. The logic is simple enough: the constraint bites hardest where orders can be pushed out, not where demand is already locked in. If you sell the flagship accelerator every hyperscaler has already committed to, a delayed substation moves your revenue between quarters. It doesn't cancel it.
The second tier takes the hit
The pain lands on memory and optical modules — the second-tier components, in Morgan Stanley's framing. When a data center project waits on power, the first thing that happens is the order book gets re-sequenced: memory, cables, and optics get deferred while the big-ticket items keep their place in line. These are exactly the parts with the least pricing power and the most competition. A six-month delay at the top of the chain becomes an inventory problem at the bottom of it. The headline names survive; the line items below them get repriced.
What to watch, and a note of caution
Three signals. First, utility interconnection timelines — if queue backlogs keep stretching, 32GW starts looking optimistic. Second, behind-the-meter projects: on-site generation and fuel cells are supposed to close part of the gap, so their delivery schedules are now as important as any chip roadmap. Third, order patterns for memory and optics — watch for pushouts, because that's the canary. And a cold note to end on: this is one investment bank's report, not a forecast from the grid. Power demand models move around, so treat 32GW and 34% as a scenario with error bars, not a carved-in-stone number. And to be explicit: a Wall Street report saying two chip stocks look insulated is not a reason to buy anything. That's the bank's view, not investment advice.