Goldman Sachs Research now expects global data center power demand in 2030 to run about 170% above 2025 levels. Four months earlier the same team's number was 117%.
The efficiency gains everyone predicted from better models and faster chips are arriving on schedule. Demand has absorbed all of them and grown past them, and the binding constraint has moved from generating electricity to the turbines, transformers, transmission lines and electricians needed to deliver it.
"It's adding the power equivalent of the whole country of Japan, which is the number five power consuming country."
Brian Singer is global head of GS Sustain at Goldman Sachs Research and covers the innovation cycle and the sustainability trade-offs it creates. Carly Davenport is the senior analyst on the firm's Americas utility team, and hers are the numbers on how much power the United States will need and where it comes from.
I listened to the full episode so you can skip it. 32 minutes of audio, 16 minutes of reading.
Here are the 12 takeaways that matter.
👤 Guests: Brian Singer, global head of GS Sustain at Goldman Sachs Research, who covers the innovation cycle and the sustainability trade-offs of the AI build-out; and Carly Davenport, senior analyst on the firm's Americas utility team, who produces its US power demand forecasts
🎙️ Hosts: Allison Nathan, a senior strategist in Goldman Sachs Research who hosts Exchanges, and George Lee, co-head of the Goldman Sachs Global Institute
📰 Published: 1 September 2026 on the Exchanges feed (Goldman Sachs) · recorded 12 August 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 32 min | ✅ Time saved: 16 min
Key Takeaways
The US power demand forecast was raised again, to a 3.5% compound annual growth rate through 2030 Data center demand in 2030 goes to about 108 gigawatts, from about 83 gigawatts four months ago
Data center vacancy in the US has fallen to 1% to 2%, against 2% to 7% in recent years
Token efficiency is real, and it raised power demand rather than lowering it Cheaper tokens widen the set of enterprise tasks worth running, so total consumption keeps climbing
The Midwest, not Texas, is the second-largest data center market by 2030 Regulated utilities in Iowa, Wisconsin and Louisiana are building the generation themselves
The constraint is no longer megawatts; it is turbines, transformers, transmission and tradespeople One large turbine maker expects to be halfway sold out for 2031 by the end of this year
More than half of data centers sit where heat, humidity or drought forces a choice between water and power New water costs more to find than new power, so US operators will spend the power
Waiting for a grid connection can take seven years, and that is why on-site power plants are being built
Public opinion, not physics, is the steepest hill A plurality of Americans told a survey they would rather live near a nuclear plant than a data center
1. The Forecast Went Up Again
George Lee opened by putting the change in one line: the fight is no longer about whether enough electricity can be generated.
Lee said the shortage has moved downstream, into the equipment and the labor that deliver power. "The binding constraint has migrated away from megawatts and towards what I think of as all the Ts." He listed them: "We have turbines, transformers, transmission, and tradespeople." He added that the politics have arrived too, with "regulators, ratepayers, and neighbors" now at the table
Davenport's team raised its US power demand forecast. "So we recently raised our forecast up to a 3.5% CAGR through 2030. And that was up from our prior forecast of about 3.2% CAGR through 2030." Actual growth is running ahead of even that: "As we think about where we sit today in power demand growth, year to date we're up over 4%."
The revision came from data center projects, measured off a 451 Research data set. "And that has taken our data center power demand in 2030 up to about 108 gigawatts, up from about 83 gigawatts before."
The second driver is that existing buildings are filling up. Vacancy has fallen across every major US data center market: "They're sitting today around 1% to 2%. Over the last few years, they were ranging from about 2% to 7%." Her team forecasts a 3% US average vacancy rate in 2030
2. Capex Up, Demand Up 170%
Nathan noted that Singer had last raised his data center power estimates in March, three or four months earlier. He started with what the bank's analysts now expect the hyperscalers to spend.
The spending forecasts moved by hundreds of billions of dollars in a single quarter. "If we go back to that period at the beginning of March, we were projecting and our analysts were projecting hyperscaler spending was at about $1.2 trillion for 2027. That's $1.7 trillion now." The 2029 figure went the same way: "Hyperscaler spending for 2029 was projected at $1.5 trillion. That's today at $2.1 trillion."
The power forecast followed the capex. "If we look globally, so not just the U.S., but globally, we see data center power demand in 2030 up about 170% from 2025 levels. We were at about 117% back at the beginning of March when we last spoke."
The scale comparison Singer reaches for is a country. Counting AI's demand beyond data centers over the seven-year period that starts at the beginning of 2024: "It's adding the power equivalent of the whole country of Japan, which is the number five power consuming country."
Nathan pointed out that the revision was not Singer's alone. "And Carly referenced the 25% increase in 2030 capacity buildout." The bank's technology, media and telecom teams also raised their forecasts for AI-related server shipments
3. Cheaper Tokens, More Demand
The question Nathan put was whether efficiency gains would eventually let buyers spend less. Singer's answer is that it has not happened yet, and Lee's is that efficiency is making the demand curve steeper rather than flatter.
Nobody is cutting technology budgets to bank the savings. "Thus far, to your question on productivity, we haven't seen the productivity gains impact and negatively impact tech R&D budgets." Singer framed the open question as the moment a buyer says it will spend less money for the same tokens rather than the same money for more: "I don't think we're there yet"
Lee said the efficiency gains are genuinely happening, and demand has swallowed them whole. "So, the Hill you describe is so steep on its face, but it's actually almost greater than it would seem because there is token efficiency taking hold in all of these systems."
Davenport sees the first sign of restraint, and says it is about intensity rather than volume. Customers are looking harder at how and where they consume tokens. "And that I think is the beginning of some level of trying to reduce consumption intensity without necessarily negatively impacting output."
Lee's counter is that cheaper tokens enlarge the market rather than shrink it. Using tokens more judiciously widens the set of enterprise tasks it is economic to run. "Tokens get cheaper. You're more efficient with their usage. And that expands the demand or the addressable market for this commodity."
4. MISO Overtakes Texas
Davenport walked through where the load actually sits today and where her team expects it to move by 2030.
Three regional grids carry most of it now. "The top three data center power demand markets in the US are PJM, which is the mid-Atlantic market that houses about 35% of data center demand today." ERCOT in Texas has about 15%, and the Southeast and the Pacific Northwest tie for third at about 13% each
PJM stays on top through 2030, and Davenport gave three reasons. A large amount of merchant nuclear capacity that data center companies want, a high-voltage transmission backbone, and existing clusters that pull further development toward them
The surprise in the work was the middle of the country. "The second takeaway we've had from our analysis is, what's surprising to us was that the MISO market, which is the mid-continent of the US, so think from Minnesota down to Louisiana, is going to overtake ERCOT, the Southeast, and the Northwest and become the second largest data center market in the U.S. by 2030." She put MISO at about 16% of the US data center market by 2030, driven by regulated utilities in states such as Iowa, Wisconsin and Louisiana
Texas slips to third, at about 14%, and carries a policy overhang. Davenport pointed to the state governor's moves to audit data center projects as a source of uncertainty in the growth outlook, while still calling ERCOT an important market
5. Why Regulated Utilities Win
Nathan asked why regulated utilities are taking the lead in MISO. Davenport gave two reasons, one commercial and one political.
The first is that a regulated utility is a single counterparty. "I think one part of it is it's a one-stop shop, right? You come to the utility. They have the relationships with the regulators." It supplies the plant, the transmission and the distribution In a deregulated market the customer negotiates with a power developer for generation and separately with the utility for the grid connection, which she said is a more complex negotiation
The second is that a regulator, not a market, sets the price. "You don't have supply-demand dynamics necessarily driving pricing. It's set by a regulator." That, she said, makes it easier to manage the affordability problem for existing customers
6. The Seven Ps of Constraint
Singer keeps a framework of constraints that all begin with the same letter. In March there were six; he has since added a seventh.
Pervasiveness of AI — still expanding in size and scope, and still a driver of growth rather than a limit on it
Productivity of servers, chips and models — real, but not yet pushing budgets down
Price of power — Singer said this is less of a concern for the hyperscalers than it is for households and communities
Policy, which he splits into three: permitting, pushback and populism. He noted that some governments are trying to write the rules now precisely so the issue does not become a long-term constraint
Parts — the transformers and turbines. "One of the major turbine manufacturers on their most recent conference call said that they expect that by the end of this year, they'll be halfway sold out for 2031." This shapes the fuel mix rather than the total: Singer expects simple-cycle natural gas and renewables plus battery storage in the near term, combined-cycle gas in the medium term, and nuclear later
People — "We are still absolutely in need of more electricians, welders that can handle high voltage connectivity." A decade of flat US power demand gave nobody a reason to enter those trades, and certification takes at least four years
Physical environment, the new one, which is about cooling. "More than half of them are built in areas with elevated physical risk, either extreme temperatures, humidity, drought."
7. Saving Water Costs Power
The seventh P sets up a trade-off with no clean answer: a data center in a hot or dry place can use more water and less power, or more power and less water.
In the US, Singer expects water to be the thing that gets minimized, which means deploying cooling technologies that use more electricity per unit of computing
He said the priority is not universal. Goldman's analysts covering China expect the opposite there, because regulation pushes operators to minimize power consumption instead, accepting higher water use
The reason is a price difference between two inputs. "It's a lot more expensive to find new sources of water than it is to find new sources of power."
Nathan raised West Texas as the case where this bites, since ERCOT is a growth market and parts of the state have little water to spare Singer allowed that cooling technology keeps improving and that "it's always hard to bet against the engineers," with a great deal of capital going into it, but said the trade-off is still forcing choices today
8. Gas 60%, Renewables 40%
Asked whether the country is heading for a gap between what is demanded and what can be supplied, Davenport declined the dramatic version of the answer.
There is no national shortage in her forecast, but there are tight regions. "We don't forecast a nationwide power shortage by any means. But there are certain markets that we think are getting critically tight." PJM tops that list ERCOT is not tight today; she expects it to become tight as demand ramps around 2028
The build is deliberately mixed. "About 60% of data center demand we think will be met with natural gas. 40% we think will be met with renewables." A handful of one-off nuclear restarts help bridge the gap, and she described the effort as all hands on deck
Behind the meter as a bridge
On-site generation that never touches the grid is now a real share of the forecast. "So, we think you'll see about 30 gigawatts of behind the meter gas capacity by 2030, which is about 20, just over 20 gigawatts of power delivery." That is roughly 20% of her 2030 data center demand estimate She said the pushback and the physical constraints in markets such as Texas are what strengthened the case for it
It is a bridge, not a destination. "That would absolutely be our view, that behind the meter is more of a bridge solution and it is not a long-term solution." The largest customers keep telling her team they would rather be on the grid, because at multi-gigawatt scale the grid is cheaper and more reliable A project due in 2028 that cannot get a connection can start behind the meter and switch later
Nathan's reading, which Davenport agreed with, is that the capacity being built off-grid may be reconnected to the grid in time, which would leave regulated customers with more supply than they had. What is driving it, she said, is the extraordinary length of the interconnect queue
9. A Seven-Year Grid Queue
Davenport called connection policy the biggest live question in her sector, and it is a queue problem before it is a generation problem.
This is the single largest uncertainty for the companies she covers. "I think the regulatory uncertainty on how data center connections will be handled is probably the single most important issue facing power and utilities companies today."
The wait is measured in years. "The queues today are lengthy. You could be sitting in the queue from anywhere - two, three, five, seven years depending on which regional grid in the U.S. you're trying to connect to." Her phrase for the customers is that they want to be up and running yesterday
Two workarounds are being used. Going behind the meter is one. The other is a special tariff structure in regulated markets: a standardized, transparent formula with the regulator's approval already attached, setting timelines, financial requirements and customer protections in advance
Utilities are pairing that with promises to existing customers, either quantifying the bill savings data center load produces or agreeing not to raise rates for a set period. Davenport's argument is that this is how the growth gets accommodated without putting reliability or affordability at risk
10. Populism Is the Steep Hill
Lee pushed back that the economics may be sound and still lose, and put the question to both guests.
Singer said the opposition will not fade quickly. "It is hard to see how this goes away quickly because there is just an understandable long history of nimbyism, even before we delve into the intricate details of the potential solutions."
He counted five specific community objections and a technical answer to each: an outage, a price surge, water use, noise, and waste heat On outages, agreeing to be disconnected when the grid is about to fail is becoming a license to operate for data centers On price, the take-or-pay contracts Davenport described are the mitigant On water, closed-loop cooling that does not consume water continuously, and dry or air cooling — some of which raise power use On noise, abatement technology inside the building; on heat, capture and reuse, which Singer said places such as Finland are doing and which needs a cold enough climate to have a market
Davenport said the pushback is real and that customers are not trying to force their way in. For every community objecting loudly, "there are other communities that are still actively courting data centers because they want the increase in the tax base. They want the construction jobs." Her expectation is that this pushes data centers into greater geographic concentration
Singer's caveat on the welcoming communities is short. "But those communities want it until they get it. And then they are living through these issues and problems."
Lee has watched this film before. "As a child of the 1970s, I watched the nuclear power industry fall afoul of populist and political opposition." He said the association with AI itself is part of the problem, and cited a survey: "I read a survey recently that said that a plurality of Americans would prefer to have a nuclear power plant in their neighborhood versus a data center." Of everything discussed, he said, "this dimension of popular sentiment might be the steepest hill we have to climb in getting to the right balance of supply and demand here"
11. Utilities Can Fund It
Nathan asked whether the utilities have the balance sheets to keep up, since the capital question is usually asked only of the hyperscalers.
The spending step-up is large. "We see capital investment for our regulated utilities coverage up 60% in the next five years versus the prior five-year period."
The credit metrics still have room. "We're still seeing about 100 basis points of cushion there," measured against the leverage ratio that would trigger a downgrade
Equity is doing a lot of the work. "We would expect incremental capital to be funded about 30 to 50% with equity." Davenport said the companies have issued ahead of the need, which is what keeps the cushion in place
12. What They Are Watching
Nathan closed by asking both guests for the indicator that would tell them whether the constraints or the demand are winning.
Singer watches state-level moratoriums, not local ones. He means a New York or a Texas, the kind that could delay projects or make a developer question whether to operate in a market at all "I think that would probably be the piece that we would be most focused on, especially as we're leading into the elections in November and potential changeover post that."
Davenport agreed, and added a distinction. Some moratoriums are a deliberate attempt to work out what a state needs before development continues, rather than a turn against it; the question is which kind they turn out to be
Her own metric is the forecast itself. "If I'm looking at metrics to watch, does this 2030 data center capacity get revised up or down from here?" Her name for the risk is borrowed: "Call it the field of dreams data centers."
Singer's off-the-beaten-path metric is nuclear. "And then if I'm looking for an off the beaten path metric to watch, it would be nuclear contracting by hyperscalers." Because a nuclear contract does almost nothing for the next five years, signing one says what a buyer believes about the decade after that, and what it thinks about emissions
Bonus Insights
Lee's closing point was a messaging one, and it carried the episode's largest count. "And Carly talked about there's a single state-level data center moratorium today. But there are more than 300 regional or local moratoria going on around the country." He said the AI and power industries need to coordinate on making the positive case in communities or the rest of the build-out is at risk
Lee also described the supply chain as stretching back from the models to the physical goods and "all the way back to the base elements", which he called one of the most complex, intricate and global supply chains in the world
Nathan took the opposite mood from the same conversation. "But I, for one, took a little bit of comfort from the takeaway that we are not seeing tremendous constraints in terms of power demand." Lee's answer was that getting this far from a standing start five or six years ago is a testament to the free enterprise system
The episode is part of the AI Exchanges series that Nathan and Lee co-host, and both guests had been on the same series in March, which is where the earlier forecasts being revised here came from
The bottom line from both guests is that the demand side of the AI power story is not in doubt and the supply side is solvable, but the part nobody can engineer — whether communities will accept the buildings — is now the constraint most likely to bind.
Products, Companies & Tools Mentioned
PJM (The mid-Atlantic grid that houses about 35% of US data center demand today and stays the largest market through 2030, on merchant nuclear, transmission and existing clusters)
MISO (The mid-continent grid Davenport expects to become the second-largest data center market in the US by 2030, at about 16%, built by regulated utilities rather than developers)
ERCOT (Texas, about 15% of data center demand today and about 14% by 2030, with the state's audit of data center projects hanging over the outlook)
451 Research Datacenter KnowledgeBase (The data set behind the revision that took 2030 data center power demand to about 108 gigawatts from about 83)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: