We Found 10GW More Power, Yet the Shortfall Grew by 19GW

Supply Grew, But the Power Shortfall Grew Even Faster

Morgan Stanley Research has updated its outlook for US data center power. It now projects that US data centers will need 97GW¹ of new power from 2026 through 2028. Back in the August report, that figure was 68-69GW.

The chart is shaped like a waterfall. Start with the 97GW needed, subtract the 21GW already secured by data centers currently under construction, then subtract 19GW that can be drawn from the grid. That leaves 57GW. Even after factoring in every alternative—gas turbines, fuel cells, nuclear plant sites—33GW still remains. That’s 34% of the total need.

What stopped me on this chart were the two middle bars. In the August report, the same slots showed about 15GW under construction and about 15GW from the grid—30GW combined. This time, that combined figure rose to 40GW. So in just over a month, the supply side improved by 10GW. But the pre-alternatives shortfall grew by 19GW, from 38GW to 57GW. In other words, the revision to the need side moved far more than the improvement on the supply side.

So why did the need estimate jump by 29GW? It wasn’t because dozens of new data center plans suddenly appeared.

What Changed Wasn’t the Building Count—It Was the Math on a Single Rack

The basis for Morgan Stanley’s revision is the rack² power of Nvidia’s next-generation products. It raised its estimate for a single Vera Rubin rack from 149kW to 234kW, and for the following generation, Rubin Ultra, from 415kW to 600kW—increases of 57% and 45%, respectively. It’s worth noting these are Morgan Stanley’s estimates, not specifications Nvidia has officially announced.

Let me run a quick calculation to give this some scale. A household using 300kWh a month draws roughly 0.42kW on average, continuously. A single 234kW rack pulls as much power, around the clock, as about 560 such households combined. And a single data center building holds hundreds, even thousands, of these racks.

A common objection comes up here: shouldn’t better chips do the same work with less power? Power per computation can indeed fall. But the unit this industry sells isn’t a single chip anymore. The unit is a rack—72 or 144 GPUs bundled with memory, networking, and cooling into what functions as one machine. The more densely you pack chips together, the more power the memory and networking connecting them consume, and the more power the cooling needed to remove that heat consumes. Chip efficiency and rack power can move in opposite directions.

So this upward revision is different in kind from an adjustment reflecting hotter AI demand. It’s the same set of plans, with a rewritten estimate of how much power each machine within those plans will draw. Morgan Stanley sees 2027-2028 as the inflection point where the industry shifts to rack-level design—moving from a competition over building the best single chip to a competition over how to power an entire rack.

Who’s Filling the Gap

The second chart is more interesting. It lays out, in bars, the ways of getting power outside the grid. The industry calls this “time to power”—the problem of how fast you can bring electricity online. New grid connections can take 5-7 years depending on the region, so these are the methods that skip that line.

• Gas turbines and engines installed behind the meter³: 17GW
• Bloom Energy’s fuel cells⁴: 5GW
• Using sites at operating nuclear plants: 2GW

These figures are already multiplied by a probability of success. According to the summary materials, turbines and engines are pegged at a median of 19GW with a 90% success probability, fuel cells at 5GW with 90%, and nuclear sites at 3GW with 75%. Multiply and round, and you get the chart’s 17, 5, and 2. Converting sites that already have power infrastructure—like Bitcoin mining facilities—into data centers was calculated separately at 13GW, and the report notes it was kept out of this chart to avoid double-counting with the under-construction figures.

Two things caught my eye.

First, the biggest bar is gas. The main way the shortfall gets filled isn’t new transmission lines or nuclear power—it’s turbines built right next to the data center. Data centers are turning from customers who buy electricity into facilities that house their own power plants.

Second, a single company’s name occupies an entire bar. One firm’s fuel cell supply capacity is rated higher than all operating nuclear plant sites combined—not because nuclear capacity is small, but because the question here is whether it can actually be plugged in before the 2028 deadline. What matters in this calculation isn’t the cost per unit of power generated, but installation speed.

And even adding all of this together, 33GW still remains.

What a 33GW Shortfall Actually Means

This doesn’t mean blackouts are coming. It’s closer to saying that a substantial number of planned racks won’t get power by 2028 and will be delayed. Even with money and GPUs in hand, if there’s nowhere to plug them in, those GPUs sit waiting in a warehouse.

Morgan Stanley doesn’t see this shortfall as a power problem alone. It points to three bottlenecks together: people, power, and politics. There’s a shortage of skilled labor to build power plants and data centers, and local opposition and regulation slow down site acquisition. Extend the timeframe to 2029, and the firm expects the shortfall to grow to 72GW.

But there’s something worth remembering when reading these figures. The fact that the need estimate moved by 29GW within a matter of weeks also shows how sensitive this forecast is to a single assumption about rack power. If Nvidia’s actual specifications or the pace at which customers deploy their systems change, these numbers will shift again. What I take from this report isn’t the number 33 itself, but the fact that the variable driving the shortfall wider was the power draw of a single machine, not the number of buildings.

Oswarld’s Lens

There’s something I say often: not every person and every task needs superintelligence. There’s a model suited to each job and each setting.

When I’ve made that argument, the reasons I’ve usually cited were cost, security, and the user’s own environment. This chart adds one more reason: electricity. Once power becomes the scarcest resource, deciding which tasks get routed to the biggest model stops being a matter of individual company preference and becomes a question of power allocation. It forces you to ask whether summarizing meeting notes with a model running on a 600kW rack is really the right allocation.

So I don’t read this report only as a signal that “AI investment is hitting a wall.” I read it as a signal that the line separating tasks that need the biggest model from tasks that don’t is shifting from a technology problem to a resource problem. Organizations that draw that line first will pass through the coming 2-3 years of power scarcity with less turbulence.

Closing

The power needed rose to 97GW, and supply grew—but not fast enough to keep up. What widened the gap wasn’t the number of data centers but the estimate of how much power a single rack draws, and the main way the shortfall gets filled is gas turbines built next to data centers. Even so, 33GW remains.

---

💬 Reader, is there a task your organization currently runs on the biggest model that could just as well run on a smaller one? Share the task that comes to mind and why in the comments.

📨 If you have a colleague tracking AI infrastructure investment or data center siting, pass this along.

---

References & Further Reading

• Morgan Stanley Research, “Potential Shortfall in Power for US Data Centers, 2026-28” and “Potential US Power Shortfall for Data Centers, 2026-28” (Exhibit 1, Exhibit 4). ··· The two charts that this piece starts from. The 97GW, 57GW, and 33GW figures, along with the contribution of each alternative, come from here.
• “‘Capital alone no longer clears a site’: Morgan Stanley says data centers’ big money era is over”, Fortune, August 20, 2026. ··· Covers the August report’s 38GW shortfall forecast—the one just before this upward revision—and the list of alternatives: gas turbines, fuel cells, nuclear sites, and mining-facility conversions.
• “Morgan Stanley sees up to 20% shortage of US power for data centers through 2028”, Investing.com, November 2025. ··· A year ago, the shortfall stood at 44GW before alternatives and 13GW after. Useful for comparing how the same bank’s forecast has shifted over time.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Gigawatt (GW): A unit of power. 1GW equals 1 million kW. Since a single large nuclear reactor produces roughly 1GW, data center power figures are sometimes framed in terms of how many reactors’ worth of output they represent. ↩

2. Rack: A shelf-like frame that holds servers stacked in layers. In today’s AI data centers, dozens of GPUs are bundled with memory, networking, and cooling into a single rack, designed and sold as one large computer. ↩

3. Behind-the-Meter (BTM): Generating electricity on-site, within a facility’s own grounds and on the facility’s side of the power meter, rather than buying it from the grid. The biggest advantage is not having to wait for a grid connection. ↩

4. Fuel Cell: A device that generates electricity through a chemical reaction—rather than combustion—using fuel such as hydrogen or natural gas. Because it’s built by stacking modules, it can be installed faster than a large power plant. ↩