Here’s a trio of estimates from Bloomberg, Bloom Energy, and S&P.
How Power Bottlenecks Are Reshaping Data Center Strategies
Bloom Energy reports the about US Had About 25 GW of Operating Data Centers in 2024. Where to from here?
Demand for power in the US is growing at an unprecedented rate After 20 years of flat demand,2 US power needs are predicted to rise by 83 terawatt-hours (TWh) in 20253 – equivalent to powering an additional 7.7 million homes.4 Data centers are the largest driver of this growth The US is expected to see the highest share of new data centers outside China. Since 2020, the US colocation data center market alone has doubled, driven by digitization, cloud, and AI. To power this, data centers by 2030 could require 8-12% of the total US demand compared to 3-4% today.
The US grid has not been able to keep pace with this demand While utilities can likely generate sufficient power to meet data center needs, they face bottlenecks with transporting that power via transmission and distribution infrastructure. As a result, grid interconnection takes longer, there is more congestion on the network, and capacity is increasingly expensive. If the US continues to build high-voltage transmission infrastructure at its current rate, it will take at least 80 years to deliver what we need over the next decade. New data center projects will struggle to get timely access to power In the US, 55 GW of data center IT capacity is expected to come online in the next five years. For comparison, this is 10 times the average power capacity used by New York City – and does not include the additional power needed for cooling systems.
Summary of findings
- Data center leaders expect power availability to get worse
Demand for power in the US is growing at an unprecedented rate, with
data centers being the largest driver of this growth. The US grid has not
been able to keep pace with this demand, and new data center projects
will struggle to get timely access to power. - Leaders are increasingly turning to onsite power as a solution
Data center leaders expect approximately 30% of all data center sites to
use some onsite power as a primary energy source supplemental to the
grid by 2030, 2.3 times more than just seven months prior. We find that
new data center announcements corroborate this expectation. - Leaders see value differently as they balance priorities
New data centers are balancing more priorities, and time to power is
playing an increasingly important role in the value equation. Our surveys
and interviews with data center leaders have surfaced seven key criteria
for choosing onsite power technology. - Looking ahead, leaders expect a paradigm shift in 2025
Based on what we have heard from data center leaders, we have identified
three key trends we expect to continue in 2025.
That report notes the trends but is stale. It’s from January 2025.
Here are some current estimates.
AI and the Power Grid
On December 1, 2025, Bloomberg commented on AI and the Power Grid: Where the Rubber Meets the Road
Data-center power demand hits 106 gigawatts (GW) by 2035 in BloombergNEF’s newest forecast – a 36% jump from the previous outlook, published just seven months ago.
The massive growth rate in data center power demand reflects more than a surge in the number of data centers in the pipeline; it also highlights the new centers’ size. Of the nearly 150 new data center projects BNEF added to its tracker in the last year, nearly a quarter exceed 500 megawatts. That’s more than double last year’s share.
This boom in data center demand is colliding with grid realities. In PJM, BNEF forecasts data center capacity could 31GW by 2030, nearly matching the 28.7GW of new generation the Energy Information Administration expects over the same period. In the Electric Reliability Council of Texas, reserve margins could fall into risky territory after 2028, a sign that short-term growth can be absorbed, but longer-term supply will lag.
These pressures point to an inflection moment for US grids: the desire to accommodate AI-driven load without undermining reliability or driving up power costs.
At the same time, the geography of US data centers is shifting. The once-dominant market in northern Virginia market is nearing saturation, sending new projects south and west into central and southern Virginia. Georgia is seeing expansion beyond the metropolitan Atlanta area as land and power constraints tighten. Texas is an exception: Developers there are transitioning former crypto-mining sites into AI data centers closer to population centers and fiber routes.
That’s 106 gigawatts (GW) in new demand by 2035 according to Bloomberg.
Data center grid-power demand nearly triple by 2030
On October 14, 2025, the S&P reported Data center grid-power demand nearly triple by 2030
Data-center power demand will rise to 134.4 GW in 2030. Virginia’s data-center power demand up 30% year over year.
Data centers proliferating across the US will require 22% more grid power by the end of 2025 than they did one year earlier and will need nearly three times as much in 2030, according to the most recent forecast from 451 Research, part of S&P Global.
Utility power provided to hyperscale, leased and crypto-mining data centers will rise by roughly 11.3 GW in 2025 to 61.8 GW, 451 Research said in its updated Datacenter Services & Infrastructure Market Monitor & Forecast, released in September.
In 2026, US data center demand will rise to 75.8 GW for IT equipment, cooling, lighting and other uses, and further expand to 108 GW in 2028 and 134.4 GW in 2030, according to the outlook. That does not include enterprise-owned data centers outside of hyperscale tech giants such as Amazon.com, Apple, Google, Meta Platform and Microsoft Corp.
The data center boom is driving the robust load-growth estimates of many electric utilities, largely fueled by the emergence of AI. But much uncertainty surrounds the precise pace of power demand from data centers, the grid resources needed to cover that demand and the capabilities of large-scale onsite power alternatives.
Related Posts
October 16, 2025: AI Data Centers Build Their Own Power Plants Because the Grid Isn’t Ready
Tech companies are creating an energy Wild West grabbing land and turbines.
November 4, 2025: If Solar and Wind Are Now Cheaper than Fossil Fuels, Why Don’t We Have More?
The answer is they aren’t cheaper.
December 3, 2025: Electricity Costs Surge 6.7 Percent from Year Ago, Residential Consumers Hammered
Expect a backlash. Residential customers pay double the industrial rate.
October 27, 2025: Why China Is On a Pace to Win the AI Race
China has three big advantages over the US: cheap electricity, an open source model, and fewer capital needs.
A couple things to be aware of here: there’s lots of double counting in this data. How much? We don’t know, but we know it’s there because the same data center developers go to different utilities in PJM, in SPP and in ERCOT and tout a new data center and potential hook ups. The same data center to multiple utilities in multiple wholesale districts. The utilities are required to make some provision for these new data centers in the forecasts they send to PJM, SPP, ERCOT, etc and then those go into new power demand forecasts for the various regions.
S&P has a good data set, but there’s a differentiation in the data that is actual operating data centers, data centers under construction, and then various stages of “development.” Some of these don’t even have any land purchased but they get included in these forecasts of power demand anyway.
Also, transmission and distribution systems are a bottleneck and the big issue there is that much of the infrastructure is aging and needs to be replaced. That will send rates rising more than data centers will. Some data center developers for the large hyperscaler operations are planning on behind-the-meter energy sourcing: if I need X MW of firm fixed capacity, I’ll build my own power plant (or plants) to supply my electricity. Most of that will be from natural gas, but that runs into other delays: GE Vernova has about a 4 year queue for CCGTs, Mitsubishi, Siemens et al have tariff barriers and their own queues. Moreover, if I need to build a substation, the components for those are in backlog, and if I need to build a pipeline to get gas to my power plants, some of the needed infrastructure there is also on a wait list. And this is assuming that the permitting process is smooth and without legal interruptions a la NIMBYism.
The issue isn’t data centers, or other large loads (if you want more industry you need more power), but an aging grid infrastructure, manufacturing backlogs for critical components up and down the grid and pipeline network, and uncertainties with permitting and siting regulations from counties and local governments. In other words, data centers are only one symptom of a grid in crisis, they are not the grid crisis itself.
How much more electricity?
Very possibly -50Gw when the American AI scam collapses.
“Tether is blocking discounted share sales while promoting a $500 billion valuation”
“OpenAI is seeking government support primarily through expanded tax credits”
OpenAI is projecting massive capital investment needs, primarily for compute power and data centers, that could total as much as $1.4 trillion over the next eight years. The company also projects a significant short-term cash burn of $115 billion through 2029
Oracle is a central part of a web of “circular” deals within the AI industry, primarily involving OpenAI and Nvidia. These arrangements have raised concerns about the true nature of the AI boom and whether it is an organic market expansion or an investment bubble
Do people know the average commercial cost per monthly “token” usage is $3700 for US A.I.s vs $70 for Chinese AIs for the 5% of customers who pay and the average cost is “free”? Is the West going to punk itself again? Is Sam Bankman-Fried orchestrating all this from his jail cell?
China is planning a 1 kilometer-wide solar array in space which will create as much energy in a year as the total amount of oil that can be extracted from the Earth.
(Solar collection is ten times more efficient in space, it generates 24hrs/day, and has free cooling available. The power is then transmitted by microwave back to earth.)
https://sustainabilitymag.com/articles/chinas-1km-solar-array-the-manhattan-project-of-energy
An Australian company plans to have a solar power satellite in orbit in two years….
https://www.solarspacetechnologies.com.au/
134.4 GW in 2030… That’s 400W per citizen average? Insane numbers… that’s like 3 desktops working day and night, and that with CPU power more efficient every year.
The question is how much of their capacity is spent on stupid social media postings and private entertainment…
A crash in AI (which to me seems likely) could dramatically change such estimates. Consider the recent reports that Microsoft has been cutting sales targets for its AI products by 50% or more, simply because people don’t find the products useful.
The video processors used to power AI are much more power hungry than simple data storage systems.
Date center expansion may be restricted by raw material supply for the build out, namely copper and silver. Both metals are experiencing supply shortages and rapid price increases
Just one more of the countless problems created by unfettered capitalism. Strict laws and harsh penalties are called for.
SOUNDS GREAT. who shall decide? pedotus trump? SCOTUS ? mike johnson and hakeem ?
The data supports the reality that big trouble is brewing and now we will get another example of can kicking to a point of no return. NOTHING will be done.
The discussion we need to have is what benefit to society do these data centres bring because when the energy is not enough to meet the needs of residential homes and data centres, it will be the big corporation data centres that will get serviced and the residential people will get nothing. Same goes for water. Time to stop the madness….make all corporations using the grid financially responsible for the costs of generating their power and water needs.
I agree. And serious issues are starting to pop up. Saw a couple of articles in Memphis, TN and Ohio as well
There is going to be some serious pushback coming soon.
Who knows if it will matter. Big money wins out most of the time
You mean time to make consumers pay directly? Generative AI has exploded in usage. Many people use generative AI on a daily basis for several hours each day. Business usage (usually subscription plans) is also exploding Investment in data centers meets consumer demand for generative AI features. Perhaps consumer/business power rates should be metered for generative AI usage. Generative AI requires lots of computational, storage, and communication resources.
That graph of a wild guess at US Data Center Power Demand shows a jump from 2015:10gw to 2035:100gw. That’s a 10x growth over four 5 year periods.
Something to keep in mind: Moore’s Law on the back of an envelope is a 10x change in one single 5 year period – a 10,000x change in that same 2015-2035 period!
Now, is Moore’s Law dead? Well, yes, from some angles, but currently it’s quite alive in the AI world. In fact, in the AI world, Moore’s Law is too conservative.
What does this mean? Well, given the ease with which light pipes can be laid compared to pipelines and power-towers, it appears that data centers belong where electricity is really, really cheap. Kiss off that PJM (rust belt and US gov) category, for instance. And consider how solar powered data centers from Alberta-Texas pair with data centers in India. Or triplet with Perth and the Sahara. For a mere 3x cost, you get 24-by-7 solar. No batteries needed.
Anyway, no one knows the balance between AI training needs and AI inference needs in 2035. And that balance has a large affect on whether AI data centers are the big deal or whether on-device “brains” are the big deal.
Much more than the hype driven bulls on wall street are estimating. However, in a trader/speculative driven market, who looks out past the next quarter or two?
We are talking about grid investments that normally require 40-50 years of amortization. Even if it were possible to build out the generation capacity, shrinking this window to 3 years would make the power extremely expensive.
Capital costs are one reason that utilities are hesitant to expand the grid to what is needed.
Same for gas turbine companies.
Which is why some data centers will be built with “some” co-generation. And some won’t be built at all.
And some data centers will be built in other countries with a greater capacity to provide cheap electricity.
investing in energy companies seems better than AI. picks and shovels. my all time favorite investment is precious metals, with energy and banking close behind. banks have implicit government backstop and free access to printed money via the discount window………..sometimes they even get paid to borrow.
Electricity will indeed be an issue but the problem needs to be expanded out to the whole ecosystem.
Electricity isn’t magic, first there is power generation usually done through natural gas although solar and wind now play a bigger role.
Let’s focus on natural gas, it isn’t just drilling a hole in the ground and pumping out gas, there is an entire ecosystem of well drilling, water pumping, gas extraction, gas refinement (no natural gas doesn’t come out clean ready to use), then gas transmission. How many people does it take to do all of these jobs? While you think about that let’s get back to electricity.
The second step to “electricity” after it is being generated is to transmit said electricity from point A to point B. This requires transmission lines, transmission towers, transformers, distribution plants, and other pieces. Most of this stuff requires copper and other metals.
Let’s focus on mining metals, it isn’t just drilling a hole in the ground and presto clean refined copper spun in wire form ready to use and deploy. Like natural gas, there is an entire ecosystem to this and lots of people to do it.
I’ll stop here but you can extrapolate all the other pieces to the puzzle and they all have one common denominator: PEOPLE!
Now take a look at the BLS median age tables at the link below and tell me what you see. I’ll give you a hint, the median age for most jobs is about 45 as of 2024/2025 which means in 10 years it’ll be 55 or so. How many people are going to start retiring around that age or won’t be able to do the hard work anymore due to age and injury?
And if you don’t have the people then it means there will be shortages or explosive costs (pun intended). Sure you can ask 70 year old boomers to go back to the field and start mining for copper, drill wells or run transmission lines but don’t count on it happening.
https://www.bls.gov/cps/cpsaat18b.htm
I hope you see the problem but if you don’t that’s OK, if you’re under 70 now, you’ll experience it first hand over the next decade and if you’re over 70 well it won’t matter to you.
one last hint: for those contemplating an exit strategy, energy generation will be a key component you will need for a high quality of life. Add it to your research playbook. 😉
Increase of 80 GW is expected between 2025 and 2030 to feed datacenters. Let’s assume that DCs run 24/7. This gives 700 000 000 MWh of electricity demand. Gemini suggests that 320 to 415 millions tons of natural gas would be needed for that. USA currently exports around 100 million tons of LNG. 1/ Will there be enough fuel to make electricity for datacenters at all? 2/ Will that lead to ban on US LNG exports? 3/ If yes what happens? China doing fine, Europe dead and the rest of the world scrambling to get an LNG cargo from Russia? Is it realistic to expect that the US will be able to keep its global dominant position when it exchanges LNG exports for AI services exports ? Here USA will compete against China. In LNG business USA doesn’t have a competitor because Qatar and Australia are vassals and Russia is sanctioned. Sanctioning Chinese AI will be difficult because China will pull the rare earth trigger immediately.
That’s great big picture thinking.
Natural gas/LNG is not the bottleneck for electricity generation in this case. Gas turbine production is the bottleneck with a 3-5 year wait list.
Which forces AI to turn to solar, and some wind.
China built 277 GW of solar in 2024 vs 33 GW in the us.
Similar numbers this year.
Which is why China has an edge in AI development.
The sooner the AI bubble pops, the better.
The lack of return on investment by appliers of AI reflects that (to a large extent) AI is not beneficial enough for customers to want to pay for it…even at its presently under-estimated cost. Furthermore, the harms AI is likely to generate are incalculable.
We dont have any choice : China will be the winner of the AI race, with incalculable catastrophic consequences. We must go “pedal to the metal” even if it is not really cost-effective. This is why all that is crazy and we are doomed. The world as we new it has ended.
We do have a choice, Cesar. All manner of choices in how we proceed. I’m not saying ignore that threat. Or that we don’t continue to develop computer capabilities.
But when the AI bubble pops, the last thing that should happen is the taxpayer bailing them out. AI is like nuclear power; all decisions regarding it need to be informed by the fact that AI can be a potential catastrophic danger.
There is an irony with our names that we are conversing. But a parallel that Caesar took the authoritarian road and that Brutus attempted to save the Republic.
Money isn’t everything.
Candles, invest now and get rich.
We will need them, for the blackouts.
Here’s the solution until commercial fusion power becomes a reality and is widely deployed.
Makes no sense. Easy to shot down a satellite, communication speed much lower than fiber, cost is prohibitive. Better build these data center in northern Canada ith freezing temperature.
Better had you read the website info before posting.
AI doesn’t require instant response and low latency unless it is involved in a war. Otherwise, people can wait a few seconds for the answer to their questions or for a piece of code or whatever. Sheese.
Radiation tolerant chipsets for a single computer cost +$50K, and they run much slower because their silicon paths are wider to accept a few pot holes high speed particles leave behind. A cluster of compute nodes is cost prohibitive.
Did you read the white paper on the website I referenced? If not, WHY exactly are you making a comment?
you have to compare this to total power capacity/demand in the US, otherwise it’s meaningless – 106 GW represents a 8% increase, that puts things into perspective
Yes agreed – but 8 percent is pretty big in light of soaring electricity costs
8% is especially high in a regulated-to-the-eyeballs economy.
After decades of sub 2% growth in electricity demand, a surge to 8% growth has caught utilities off guard and unprepared.
Another of my favorite topics. Energy and particularly, electrical energy.
AI/Data Centers currently use around 40 GW of power. This is expected to grow to over 100 GW by 2030, or 2035, according to the 2 different projections presented above.
It will be difficult, but not impossible, to satisfy that additional 60 GW of demand based on how quickly the US is adding more power generation each year.
In 2024, we added 30 GW of solar, 7 GW of wind, 4GW of natural gas, 1 GW of nuclear.
In 2025 we have added 33 GW of solar, 7 GW of wind, 7 GW of natural gas, and 0 GW of nuclear.
2026 looks to be about the same as 2025.
Key points:
We are not adding much nuclear. Why not? Too expensive, and takes too long to build.
We could add a bit more gas generation, but we cannot get the gas turbines. There is a 3-5 year waiting list for turbines.
The only sure way to expand generation is with more solar and wind. Particularly when you add batteries as well. However, Trump is doing his best to stop wind and solar.
US companies are also busy setting up AI/Data Centers in countries that can provide cheap and reliable electricity generation.
The problem with wind and solar is that you also are going to require a massive amount of new transmission lines to transmit the power from where it’s generated to where the AI / Data centers are. That’s going to require a massive investment of it’s own.
AI / Data centers are going to require co-located power generation (gas or nuclear) OR AI / Data centers are going to have to be built in the middle of nowhere if they are going to use solar / wind.
Solar Panels are also not available enough domestically, to have Solar as an option at this point. It is in the mix down the line, and we have some of it available, but not at that level just yet, I don’t believe.
Solar is supplying over 70% of new generation in the US each year. See my earlier post.
Solar and wind remain our best option for new generation. Upgrading and expanding the grid is required no matter what form of generation we use.
AI Data Centers can be co-located with solar and wind as well.
Not for data centers. Data centers need dispatchable, uninterrupted power. Forget wind/solar with with massive batteries in the dead of winter. The cost of wind/solar with massive battery capacity, transmission capacity, and grid intelligence is overwhelming.
Is ge the main gas turbine supplier
The main suppliers of gas turbines in the U.S. include GE Vernova, Siemens Energy, Solar Turbines (a Caterpillar company), and Mitsubishi Power. These firms dominate utility-scale, industrial, and aero-derivative turbine markets.
Why cant we accelerate gas turbine building ? Nat gas is the best option, by far. Can be built near on-site, no grid problem
A reasonable question. Things are rarely as easy as we think.
From copilot:
We can’t rapidly expand gas turbine construction due to a mix of manufacturing bottlenecks, high capital costs, uncertain demand signals, and policy headwinds. Even with rising power needs, especially from AI and electrification, the buildout is constrained by supply chain and market dynamics.
Here’s a breakdown of the key barriers:
🏗️ 1. Manufacturing Capacity Is Already Near Maxed Out
• Global gas turbine manufacturing is running at ~90% utilization as of 2025, leaving little slack for new orders A.
• U.S. developers are facing delivery timelines stretching into 2030 or later for new combined-cycle plants A.
• OEMs like GE Vernova, Siemens Energy, and Mitsubishi Power are cautious about expanding capacity due to past boom-bust cycles B.
💸 2. Capital Costs Have Skyrocketed
• Construction costs for new gas plants have surged, driven by inflation, labor shortages, and materials volatility A.
• In many U.S. markets, power prices are too low to justify new gas builds, especially without capacity payments or long-term contracts A.
⚖️ 3. Policy and Regulatory Uncertainty
• Decarbonization mandates in Canada, the U.S., and Europe are pushing utilities to prioritize renewables and storage.
• Unabated gas (without carbon capture) is increasingly marginalized in long-term planning, especially in jurisdictions with net-zero targets A.
• Permitting for new gas infrastructure is slower and more contested, especially in regions with strong environmental opposition.
🔋 4. Competition from Renewables and Storage
• Solar, wind, and batteries continue to fall in cost, making them more attractive for peaking and mid-merit roles.
• In some markets, renewables + storage now undercut new gas on a levelized cost basis, especially with subsidies.
🧠 5. Demand Signals Are Unclear Despite AI Boom
• While AI data centers are driving power demand, many projects are being delayed or canceled due to siting, cooling, and cost issues A.
• Utilities are hesitant to commit to long-lived gas assets without firm, long-term demand signals.
🔧 6. Midstream Infrastructure Gaps
• Pipeline constraints and lack of firm gas supply contracts limit where new turbines can be sited.
Sources:
AAmerican Public Power Association
CGas Processing & LNG
BInstitute for Energy Research
“hyperscale tech giants such as Amazon.com, Apple, Google, Meta Platform and Microsoft”
I do notice that these corporations have stopped pontificating about ‘carbon emissions’ ever since AI has become a thing. AI will soon become the biggest carbon emitter of all.
Very interesting, you are absolutely right. Suddenly, all the climate alarmists are silent or at least stay silent on the huge massive increase of energy production. Their true motive has become apparent.
Are these data enters avoiding the hotter climate southern states in order to reduce cooling(i.e. higher power needs) requirements?
Buy land in West Virginia. Build data centers with on-site generators there. People relocating for data center jobs will just have to learn to live in a less cosmopolitan environment.
A data center require very few workers on site.
Maintenance if any is almost remote.
Like substations – unmanned.
Whatever happened to GHG CO2 ESG DEI?
They found a better control method!
DEI is still dominant , specially in the whole education system where freedom of speech has died.