Powering the AI Boom: Can Energy Supply Keep Up?

In terms of build-out of data centers and data center demand, as discussed, that's a little bit challenging to speculate on, but I think we all know that we're gonna continue to see that build-out happen.

The issue will be: Do we have enough power generation and a grid in place with which to sustain that build-out? I do believe that we will.

I think a lot of this will actually become on, on a topic we didn't touch on much, which is through behind-the-meter or bring-your-own-power.

That may be the solution here, which bypasses a lot of the issues that we've been talking about.

So at the end of the day, when hyperscalers and ot- neo clouds and others are worth billions or trillions of dollars, they will find a way to get their power.

It's just a matter of where, whether they need to build it themselves or find ways to access the grid.

I think the constraints are likely around generation, transmission, and interconnection milestones for power.

This is likely to be a major constraint, and I also think that for behind-the-meter projects, equipment deliveries are gonna be a major constraint.

So those are gonna contribute to longer lead times on delivery of data centers.

In both cases, both grid-connected and behind the meter, there's tons of investor demand to provide capital to these projects, but for them to reach the point where they're bankable, they're gonna need to get to the front of the interconnection queue for grid-connected data centers, and for the behind-the-meter stuff, they're gonna need to actually obtain equipment and build their power plants.

There's a few items that could change the outlook that I'll just touch on briefly that I think could make things a little, a little quicker to come to market.

The first one being demand.

There are more and more utilities and grid operators that are doing things and taking efforts to cut high-probability load forecasts and shrink the interconnection pipeline so that the speculative developers that are in the interconnection queue can kind of get pushed out and clear the way for real demand.

That will likely drop overall demand, and this is a pattern that we've started to see in places like Northern Virginia and Ohio, where additional large load interconnection requests have been delayed through 2030 in Northern Virginia and been paused outright in Ohio.

So this is a place to watch.

And then for technology, there's been a real improvement in technology over the last few years with new generations of GPUs outperforming old ones.

But a further step change in compute efficiency with more inference or training output per watt could significantly ease grid constraints without requiring any new generation and could ease the way to speeding up the queue.

In terms of build-out of data centers and data center demand, as discussed, that's a little bit challenging to speculate on, but I think we all know that we're gonna continue to see that build-out happen.

The issue will be: Do we have enough power generation and a grid in place with which to sustain that build-out? I do believe that we will.

I think a lot of this will actually become on, on a topic we didn't touch on much, which is through behind-the-meter or bring-your-own-power.

That may be the solution here, which bypasses a lot of the issues that we've been talking about.

So at the end of the day, when hyperscalers and ot- neo clouds and others are worth billions or trillions of dollars, they will find a way to get their power.

It's just a matter of where, whether they need to build it themselves or find ways to access the grid.

I think the constraints are likely around generation, transmission, and interconnection milestones for power.

This is likely to be a major constraint, and I also think that for behind-the-meter projects, equipment deliveries are gonna be a major constraint.

So those are gonna contribute to longer lead times on delivery of data centers.

In both cases, both grid-connected and behind the meter, there's tons of investor demand to provide capital to these projects, but for them to reach the point where they're bankable, they're gonna need to get to the front of the interconnection queue for grid-connected data centers, and for the behind-the-meter stuff, they're gonna need to actually obtain equipment and build their power plants.

There's a few items that could change the outlook that I'll just touch on briefly that I think could make things a little, a little quicker to come to market.

The first one being demand.

There are more and more utilities and grid operators that are doing things and taking efforts to cut high-probability load forecasts and shrink the interconnection pipeline so that the speculative developers that are in the interconnection queue can kind of get pushed out and clear the way for real demand.

That will likely drop overall demand, and this is a pattern that we've started to see in places like Northern Virginia and Ohio, where additional large load interconnection requests have been delayed through 2030 in Northern Virginia and been paused outright in Ohio.

So this is a place to watch.

And then for technology, there's been a real improvement in technology over the last few years with new generations of GPUs outperforming old ones.

But a further step change in compute efficiency with more inference or training output per watt could significantly ease grid constraints without requiring any new generation and could ease the way to speeding up the queue.