Solving AI’s Power Bottleneck

David Arcaro · Morgan Stanley

The AI data center power shortfall has grown to upwards of 80 GW by 2029, and on-site power is becoming long-term.

Morgan Stanley's shortfall estimate has risen over the last several iterations and now stands at upwards of 80 GW by 2029. Data centers are turning to on-site 'bring your own power' generation because waiting years for a utility grid connection is too slow. Fuel cells, engines and turbines are being adopted for fast, reliable power. Hyperscalers and developers have moved from treating these as short bridges to signing 15-, 18- and 20-year contracts, so they now look like permanent supply rather than a stopgap.

“We're now looking at upwards of an 80 gigawatt shortfall by 2029.”at 1:00

David Arcaro · Morgan Stanley

On-site power for data centers is growing fast and is increasingly being contracted on long-term, 15-20 year deals.

Data centers face a grid connection wait of years, so they are bringing their own power. Speed matters most, but community support, air emissions, water use, noise and equipment operating characteristics are increasingly weighed as well. That led the analysts to fuel cells, engines and turbines, which the industry is pursuing in parallel. Hyperscalers and developers now sign 15- to 20-year contracts for these technologies, which were once considered smaller-scale, less attractive bridge solutions.

“...extremely rapidly in popularity among hyperscalers and data center developers, and we're seeing them much more being willing to consider it on a long term basis.”at 1:43

David Arcaro · Morgan Stanley

Utility transmission networks will fall meaningfully short in keeping pace with data center power demand.

Utilities have never grown this fast and are scrambling to meet unprecedented data center electricity demand. Their transmission networks are falling behind, especially long-distance power lines and substation equipment. These are large-scale infrastructure projects that take years to develop and build, so the analysts expect utilities to fall short, and pretty meaningfully. That is why the bottleneck is often in transmission rather than power generation.

“It's often their electric transmission network that is falling behind. It's the long distance power lines and substation equipment that the utilities are really trying to react and to grow alongside the data center industry, but we expect them to fall short and pretty meaningfully short.”at 3:13

Andrew Percoco · Morgan Stanley

Reshoring of US manufacturing adds roughly 10 GW a year of incremental power demand.

The analysts' teams estimate that the trend toward onshoring manufacturing and the reindustrialization of the US adds about 10 gigawatts a year of incremental power demand. This is a source of new demand separate from data centers. The estimate is the analysts' own, with a hedge on the number ('I think').

“Our teams have collaboratively estimated as, I think, 10 gigawatts a year of incremental demand for power just from industrialization or reshoring of The US.”at 5:02

David Arcaro · Morgan Stanley

Data center battery storage demand will exceed grid battery storage demand in new projects by 2030.

Data centers ramp electricity load up and down instantaneously. The grid cannot handle that, and some on-site power equipment is not ready for such fast changes. Battery storage is therefore being added to a growing share of data centers across the country. The analysts expect this to be one of the fastest-growing applications for lithium-ion storage. By 2030, they expect data center demand for battery storage to exceed grid demand for battery storage in the incremental new projects added that year.

“...we would expect the data center demand for battery storage in 2030 to even exceed the grid demand for battery storage in that year in terms of the incremental new projects that are added in that year.”at 5:31

Andrew Percoco · Morgan Stanley

Vehicle-to-grid becomes a major power source only once EVs reach 20-40% of US sales, a 20-30 year conversation.

California already has vehicle-to-grid programs that dispatch 500 to 1,000 megawatts to the grid at peak. This is isolated to California because it has very high EV penetration and residential solar with batteries. Nationally, the analyst says the technology only starts to matter once EVs make up 20 to 40% of annual US vehicle sales. He is candid that this is more of a 20- to 30-plus-year conversation before it moves the needle.

“...over time, we start thinking about, you know, 20, 30, 40% of annual vehicle sales being EVs across The US, that's when that could start to become a much more significant contribution to the overall power supply. So it's probably more of a twenty, thirty plus conversation if I'm going to be honest...”at 9:08

Andrew Percoco · Morgan Stanley

Automakers can use underused EV capacity and repurposed auto tech to win faster, higher-margin AI power business.

Lower-than-expected EV demand has left automakers with underutilized capacity. Companies are exploring energy storage as a new end market to reuse it. Tesla has a large US storage business, and Ford is building one and has started a rather large backlog. Storage offers faster growth and, in some cases, higher margins than EVs, so the analyst calls it the most obvious opportunity for the auto industry in AI power infrastructure. Auto suppliers are also repurposing automotive-grade technology into turbine generators, storage systems and microgrid inverters for data centers, with little incremental capital. This matters because the auto industry has seen little growth over the last five years.

“...you are seeing a number of companies explore energy storage as a new end market to essentially reutilize their low utilized capacity or underutilized capacity as it relates to EVs, just given the lower than expected demand there. ...which is faster growth and in potentially some cases, higher margin than they've seen in the EV industry.”at 6:33