The 32 GW Warning: Why Power, Not Chips, Is the Bottleneck

For three years, the data center industry’s capacity story has been a chip story: who can get GPUs, how many, how fast. Morgan Stanley’s infrastructure analysts just told buyers to stop looking there. In an October 2026 report, the bank estimates that U.S. data centers will face a net power shortfall of roughly 32 gigawatts through 2028 — about 34% of the additional power the industry’s own buildout plans call for — even after accounting for on-site generation and fuel cells that operators are already deploying to close the gap. The chips, in other words, may show up before the power does. For anyone negotiating a colocation contract, a build-to-suit lease, or a multi-year capacity reservation right now, that reordering of constraints changes which risks actually deserve a clause in the contract.

The Shift: Power, Not Silicon, Is the Binding Constraint

The conventional worry list for data center buyers has run: GPU allocation, then construction labor, then land, then power. Morgan Stanley’s analysis inverts the top of that list. Nvidia and Broadcom, the report argues, are relatively insulated from near-term delay risk because their products ship early in a project’s timeline and get installed as soon as a shell is powered. The components at real risk of stranded capital are the ones that go in last — memory, networking gear, and optical interconnects — because they sit in inventory, unpaid-for and unprofitable, for every month a site waits on a grid interconnection that isn’t ready. Reuters’ coverage of the same report flagged memory and optical component makers specifically as exposed for this reason. A GPU cluster that can’t get energized isn’t a chip problem. It’s 32 GW of commitments sitting idle because the wires, substations, and generation didn’t arrive on the same schedule as the silicon.

This tracks with a trend TechInfraHub covered earlier this year on the GPU and memory supply crunch: the hardware bottleneck has been real, but it has also been the easier of the two problems to plan around, because chip lead times are at least quoted in months. Grid interconnection and transmission upgrades, Morgan Stanley notes, typically take four to seven years — a timeline that simply does not bend to match a hyperscaler’s 18-to-24-month construction schedule, no matter how much capital is thrown at it.

What the Numbers Actually Say

Context matters here, because the 32 GW figure is a gap, not a ceiling. Morgan Stanley’s own math implies the U.S. data center sector will need roughly 94 GW of total incremental power supply to meet its build-out ambitions through 2028, and the bank projects it will fall about a third short of that even after crediting the behind-the-meter gas turbines, fuel cells, and microgrids operators are racing to install. Separately, Goldman Sachs has estimated U.S. data center capacity will reach roughly 64 GW by the end of 2026 and climb to about 90 GW by the end of 2027 — numbers that, read alongside Morgan Stanley’s shortfall estimate, suggest the industry isn’t short on ambition or capital. It’s short on megawatts that can actually be delivered to a meter on schedule.

Morgan Stanley first ran this calculation in September 2026 and has stood by the shortfall estimate since, even as some outlets have cited a slightly higher 38 GW figure in later write-ups of the same research — a reminder that these are modeled estimates with real uncertainty bands, not a measured fact. The direction of the finding, though, has been consistent across every version: the gap is large, it is dollar-denominated in the tens of billions once you price out the generation and transmission assets needed to close it, and it does not resolve on its own by 2028.

Why Self-Generation Hasn’t Closed the Gap

Operators know this is coming, which is why gas turbines, fuel cells, and on-site generation deals have become standard parts of hyperscale announcements over the past 18 months. Morgan Stanley’s 32 GW figure already nets out that self-generation buildout, which is the part of the report that should worry buyers most: even with every behind-the-meter project currently announced or under construction factored in, the shortfall still holds. Turbine manufacturers have multi-year order backlogs of their own, air permits for new on-site gas generation face their own review timelines, and large fuel cell deployments at gigawatt scale are still more pilot than proven supply chain. Self-generation is real and it is growing, but it is supplementing grid power, not replacing enough of it fast enough to erase a structural, multi-year gap between buildout speed and interconnection speed.

Who Actually Bears the Risk

This is where the buyer calculus gets specific rather than theoretical. If you are reserving capacity in a colocation facility or signing a build-to-suit lease, the exposure isn’t abstract industry risk — it’s a specific question about whose interconnection queue position, whose utility service agreement, and whose on-site generation permit stands between a signed contract and an energized rack. A facility with a firm, already-executed utility interconnection agreement and a documented substation upgrade timeline is a fundamentally different risk than one marketed on projected capacity that assumes a queue position clears on schedule. The gap Morgan Stanley describes doesn’t hit every site equally; it concentrates in markets where transmission is already congested and where multiple hyperscale projects are competing for the same substation capacity, which in 2026 increasingly means Texas, parts of Virginia, and sections of the PJM footprint.

It’s also worth being honest about the counterargument: not every forecast of a power crunch has aged well, and bank research arms have incentives of their own — Morgan Stanley’s report is partly a stock-picking exercise, framed around which semiconductor and utility names are protected versus exposed. Utilities are accelerating large-load tariff structures and grid-enhancing technologies that could pull some of that 2028 timeline forward, and several gigawatts of nuclear and SMR capacity currently in planning could land earlier than historical nuclear build timelines would suggest. The shortfall is a base-case estimate, not a certainty, and buyers who treat it as settled fact risk over-paying for “guaranteed power” premiums that may not be necessary everywhere.

What This Means If You’re Buying Capacity Now

The practical response isn’t panic, it’s diligence that matches the actual constraint. Before signing, buyers should ask providers for the specific interconnection study status and queue position for the site in question, not just a marketed in-service date. They should ask what share of a facility’s power is firm grid supply under an executed agreement versus projected or contingent on a future substation upgrade, and what the contractual remedy is if energization slips. They should weight site-selection decisions toward regions with demonstrated utility capacity rather than announced capacity, and treat any colocation or build-to-suit timeline that assumes a four-to-seven-year grid upgrade compresses into eighteen months with real skepticism. None of this means avoiding growth markets like Texas or Northern Virginia; it means pricing the power-delivery risk into the contract the same way the industry already prices construction and labor risk, instead of assuming the chips arriving on time means the capacity will too.

The headline risk in data center capacity planning has quietly flipped. For most of this AI buildout cycle, the question buyers asked was “can I get the GPUs.” Heading into 2028, the harder question is “can my provider actually get the power to run them, on the schedule they’re promising.” Buyers who build that question into their diligence and their contracts now will be the ones who aren’t caught waiting on a meter next to a warehouse full of unused memory and networking gear. For a structured way to evaluate power delivery risk, interconnection status, and contract terms before you sign, see the TechInfraHub Data Center Buyer’s Toolkit.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →