The Future Power Bottleneck in the Development and Deployment of AI – Another Catalyst for Market Woes?
The pressure on ROIs (return on investments) in AI is acute and large by nature. When CoreWeave’s $9.9 billion AI data centre out in Texas faced a 2 month construction delay in mid 2025, around $14 billion was wiped from its market cap in five weeks.
For context this site – although consuming enough continuous power to run a small city (260MW capacity) – was one in a portfolio of 41 sites.
There is much debate around whether the sell-off was also driven by belated reporting by CoreWeave – having not admitted the delay earlier – or potentially concerns around honesty in scale of the problem and cause. Thunderstorms were in part to blame according to CoreWeave – “delay in concrete pouring” - but weather reports show that there were only three days of thunderstorms at the site between June and September.
Perhaps the cause of delay was something more sinister for investors like technical issues or money problems.
Either way, the CoreWeave incident shows the main implication for the potential power problems I will discuss – that the market tightly prices “capacity comes online exactly on schedule and to the intended scale.” With large sums of investment being added at rapid rates, the AI market is very sensitive to any potential divergences in expected output.
Bottom-line of article:
Power supply could struggle to meet AI data centre demand in the U.S., which is a risk for a market that seems very sensitive to disruption.
Note:
The Context section discusses what prompted the idea for this blog post.
It also goes on a tangent to address the implications of the original FT article that prompted my idea.
In short, feel free to skip to the next section (after Context) if you don’t want to read about the implications of renewable purchase power agreements losing government subsidies in the U.S.
Although it does have slight relevance to the main idea of the blog post (potential causes of future power supply constraints for AI).
*******************************************************************************************************************************
Context
https://giftarticle.ft.com/giftarticle/actions/redeem/05ec077c-01ae-4f08-8825-8d38185dde56
Or
https://www.ft.com/content/911ad6e1-154d-4d10-a7dc-2bd0fd6fbdc2?syn-25a6b1a6=1
(Please contact me directly if you need a functioning and free-to-view link)
The thinking behind this blog post was originally prompted by reading an FT article about how subsidies are being removed from renewable purchase power agreements (PPAs), potentially suggesting demand is already sufficient not to merit subsidy - although may be a second order effect of America discounting environmentalism and/or rising cost of infrastructure (e.g. transformers).
Since subsidies for renewable PPAs are being removed, and such continuous demand for renewable PPAs is coming from the AI industry, other users (e.g. hospitals, manufacturers) are being priced out.
Renewable power projects can sell to the highest bidder, and there shouldn’t be deflationary concerns from a smaller customer group having greater bargaining power because according to projections, the AI industry’s need for power is not slowing down.
Data:
- According to Goldman Sachs, data centres consume about 6.6% of U.S. power generation, expected to double by 2027. Importantly, they could also account for nearly half of the overall growth in U.S. electricity demand between now and 2030.
The problem of renewable PPAs becoming more expensive because of subsidy removal and AI demand is that it forces other large grid users to look for other sources as a form of hedge – potentially hydrocarbon PPAs – in order to have a secure supply at a more locked-in price. The reason PPAs act as a hedge is because by providing initial investment, the producer guarantees you a fixed rate per megawatt-hour.
This is particularly relevant since U.S. power prices have gone from $135 MWh to $195 MWh in the period 2020-2026 – the generative AI boom. Price volatility has also increased over the same period because of the growing load (e.g. data centres, electric cars) creating less spare grid capacity in times of need (e.g. extreme weather).
Average Price of Electricity per Kilowatt-Hour in U.S. Cities
However, with the proposed changes to the Greenhouse Gas Protocol (GGP) for 2027, companies may have to match their energy consumption with renewable production on an hourly basis rather than as an end-of-year total. For example, if a company is consuming 100 MW 24/7, then 100 MW from renewables has to be added to the grid every hour of the day.
The problem is that a lot of renewables are by nature intermittent, and with battery technology still in its infancy, this poses a problem to the AI industry as it requires huge amounts of power at a constant rate.
It also poses a problem to other users who are potentially being pushed out of renewable PPAs because of price, since they would still have to undergo hourly matching albeit on a smaller scale (and not necessarily every hour of the day unlike data centres).
The FT says that many users don’t want to sign 10-year renewable PPAs currently at such high prices, particularly as the GGP will not be finalised until 2027. I think with the pressure put on the grid by AI, the suggested protocol amendments won’t be introduced in the U.S.
But if the amendments are made then I) one contributing reason the AI industry may not be able to keep up with planned scales of development/deployment and II) other industries will struggle to afford the power bill and consequently offload the expense onto consumers, or try to borrow more, or scale down operations, or rebel. None of which is positive for an economy.
*******************************************************************************************************************************
Returning to the main theme of the article: What are the main reasons for power supply concerns for the AI industry? And what does this mean for markets more broadly?
Despite AI companies seemingly being pretty price inelastic – i.e. they pay whatever is necessary – power supply is also relatively inelastic and questions regarding the grid capacity in the U.S. are coming to light under the current demand outlook. This could pose a serious problem for future development/deployment of AI.
U.S. grid capacity and projections:
- The Energy Information Administration (EIA) predicts the U.S. will need about 200 GW added capacity by 2030, and planned additions look set to reach that on paper. But roughly 80% of what’s planned is solar and wind, which only generate power part of the time – about 23% yearly availability for solar, 34% for wind, versus 90+% for gas. Once that’s accounted for (provided there isn’t a sudden acceleration in grid batteries), dependable capacity being added covers only about half of what’s needed (*National Centre for Energy Analytics). And the EIA predictions is probably understated, since it’s based on data only through 2025, before the sharpest jump yet in AI-Driven demand.
- Regional data from the North American Electric Reliability Corporation (NERC) suggests certain parts of the grid already lag projected demand. Such regional grids flagged at risk of not having enough power capacity within the next decade include: Texas, the Midwest, the mid-Atlantic corridor stretching from Washington DC to Chicago, and the Pacific Northwest. Note: the mid-Atlantic corridor includes Virginia a.k.a “Data Centre Alley”, the world’s most concentrated area of data centres.
Equipment and project build times:
- Transformers - which are used to properly connect any new source of power to the grid - now take 3-5 years to be built, roughly double the 24-30 months it took prior to 2020. This is because manufacturing hasn’t been scaled up to match the demand (they are specialist processes, requiring specialist materials). Switchgear, in effect a switch to a transformer, is also reportedly sold out until 2028. Transformer costs themselves have risen 60-80% since the start of 2020 – in part due to the rising price of metals.
(References: American Society of Civil Engineers; Sightline Climate, 2026; Wood Mackenzie, 2026; ChargedUp, 2026)
- Solar and wind projects go up quite fast once approved, but getting permission to connect to the grid now takes about 5 years on average, up from under 2 years in 2008. This is because far more projects are applying than grids can process. Also each project's proposal assumes everyone ahead of it in the line will connect - so when one project drops out or changes, everyone behind it often has to be reassessed from a power engineering perspective. Most projects that enter this queue never make it through at all.
(References: LBNL/PV Tech, 2025; Alliance for Competitive Power, 2026; FERC, 2026)
- Gas turbines, like transformers, have massive lead times because of demand. Manufacturer backlog now means new orders take 3-8 years to arrive, and prices have roughly quadrupled.
(References: GE Vernova, 2026; Power Engineering, 2026)
- Nuclear is the slowest by far — the last major US nuclear plant took around 15 years to build and cost more than double its original budget.
(References: Utility Dive, 2026; Nuclear Costs, 2025)
- Geothermal is currently the fastest non-intermittent option - the leading US project is aiming for large-scale output within about 5 years. However, geology restricts the expansion and access to geothermal.
(References: Fervo Energy, 2026; Carbon Credits, 2026)
Overall, supply constraints are a potential problem in addition to already quickly rising marginal power costs in the U.S. – this physical bottleneck could be the first disruption in widescale AI development/deployment even before the ceiling for power prices is tested.
Another reason for capped supply beyond infrastructure is: social – consumer power costs become elevated enough because of data centre demand to force backlash / tariffs / policy.
This social element is the least considered angle as to why future AI development/deployment could be impeded. It is entirely feasible considering historical behaviours:
Plattsburgh, New York (2018): Cryptocurrency miners flocked to the town for its cheap hydroelectric power, and once demand exceeded the town's hydroelectric power allocation, utility companies had to buy costlier power on the open market - pushing residents' electric bills up by as much as 50% and hitting local businesses with tens of thousands of dollars in extra monthly costs. Public anger forced the city to pass an 18-month moratorium on new crypto mining - the first such ban in the US - a direct precedent for energy-hungry computing infrastructure getting capped once it raises costs for other people.
The California Water Wars (1920s): Los Angeles diverted water from the Owens Valley via a new aqueduct to fuel its growth, drying up local farmland in the process. Local ranchers and farmers, cut off from a resource they depended on, escalated to sabotage - seizing and dynamiting the aqueduct multiple times in 1924. The most extreme historical example of the same mechanism: a large and powerful demand centre capturing a shared resource at a local population's expense - provoking a forceful response.
Finally, there is also the general concept in power of grid operators naturally curtailing supply away from data centres - despite being the highest bidder - because of blackout mechanisms / balancing the grid.
Mechanism: The grid must stay balanced in real time - if demand outpaces supply even briefly, the system risks cascading blackouts. Paying more can't summon more electricity into existence in that instant, since supply is fixed by what's physically being generated at that moment. So operators pre-arrange the right to cut off certain loads during emergencies, and data centres are often first in line since they can be curtailed quickly and ease a large amount of load at once. Being willing to pay the most buys market priority, not immunity from the physics of keeping the grid from failing.
In order to try and counter the physical bottleneck concern as well as the need for reliable energy around the clock, the industry could and is building out their own private supply lines.
- Note: There is a distinction between purely private and collared private power infrastructure. Purely private meaning it is totally off grid – and therefore avoids GGP rules and also curtailment during balancing episodes. Collared private meaning the infrastructure is funded/owned by a company, but still connects to the public grid, meaning it is subject to potential environmental rules and the data centres are still at risk of curtailment.
Examples:
- Oracle’s Project Jupiter (New Mexico) – purely private. Specialised fuel cells with no connection to the public grid. Oracle is the operator of the physical site, holding a long-term power agreement with OpenAI.
- Microsoft’s Three Mile Island (Pennsylvania) – collared private. Microsoft funded the restart of the nuclear reactor there and receives the output via a PPA, but the plant still connects to the main grid. So Microsoft benefits from locking in a fixed price for power, plus a strong claim toward its clean energy accounting, since nuclear also generates reliably around the clock. But is still subject to the usual curtailment risks from the grid.
- Bloom Energy’s deal with American Electric Power (AEP) – purely private. AEP, a major US utility, has partnered with Bloom Energy to implement up to 1GW of Bloom Energy’s specialised fuel cells to power data centres directly on-site - bypassing the grid entirely. This shows that utilities are starting to see the opportunity in helping create private power supply for data centres. Bloom Energy’s own industry survey (take with pinch of salt) found that developers expect roughly a third of all US data centres to run entirely on onsite power by 2030, a figure that rose 22% from just six months earlier.
Investment ideas (examples of what to invest in to capture this opportunity for retail investors):
- Bloom Energy (BE) – the fuel cell maker helping to enable the purely-private/off grid model.
- Vertiv (VRT), Eaton (ETN), GE Vernova (GEV) – power management, switchgear, and turbine suppliers that benefit regardless of which model (on-grid or off-grid) wins. Note: many of these are at or near to all-time highs as a consequence of the past two years’ AI boom.
- ETFs for diversified retail exposure:
o The ALPS Electrification Infrastructure ETF (ELFY) – most generalised to the buildout of America’s power infrastructure as a whole (other drivers will include EVs and manufacturing, not just data centres).
o The First Trust NASDAQ Clean Energy Smart Grid Infrastructure ETF (GRID) – focused on grid modernisation related stocks e.g. meters, transmission, distribution equipment, storage, software. More of a proxy for money being spent upgrading the grid itself.
o The Defiance AI & Power Infrastructure ETF (AIPO) – at least 90% of stocks must be companies tied to AI and power infrastructure together e.g. data centre operators, private energy, AI hardware. Its top 10 holdings make up more than half the fund, so it’s the most direct pure-play on the AI power theme specifically.
You may ask – why doesn’t every data centre develop fully private supply lines to avoid GGP rules and curtailment risk? Simply because it is damn expensive to bear the cost. One must build their own backup power too, instead of relying on the grid’s redundancy for free. Hence, purely private power infrastructure is currently a strategy only the biggest players can implement. And to add… one cannot erect private power supplies instantaneously either. Planning permission, acquiring in-demand specialised equipment/materials, manufacturing lead times etc. still stand in the way.
*******************************************************************************************************************************
Why is the power issue another catalyst for market woes?
Power supply troubles could be a separate catalyst for a market wobble. The CoreWeave case illustrates that the market is very sensitive to disruptions in expected output. Since this power problem could be applicable to more than one individual data centre or company, struggling to access power and falling short of expected output could really hurt investor sentiment on a wide scale and the future development/deployment of AI.
This comes alongside other potential downside catalysts such as performance based ROI concerns (e.g. productivity gains) and more broadly, circular financing concerns.
Supply side risk just adds to a basket of potential AI pitfalls.
*******************************************************************************************************************************
Conclusion
In conclusion… American AI projects need a reliable and growing supply of energy relative to what the U.S. grid can provide now, which means all energy sources will probably be drawn upon because of this pressure.
Hydrocarbons (i.e. gas) will continue to be critical.
However, if AI consumption pressures the grid too much or starts to seriously impact other users, I imagine AI takes the hit.
I think the ceiling for the price of marginal power is a lot lower for other grid users, so this will create backlash towards the AI industry even before their own price ceiling could be tested.
Even if the industry can afford to start building their own private energy projects, will these projects be able to deliver on schedule? This is another part of the physical bottleneck.
If such power disruptions start to occur – I don’t think this will necessarily be limited to one off cases like we have seen with CoreWeave – as this power problem could become universally applicable.
And this is a real risk for the anticipated (and invested) development/deployment of AI, which consequentially is a market risk.
*******************************************************************************************************************************
Question for audience
If anyone reading this knows more about data modelling, I’d love to learn from you, as I’d like to discover how I’d approach applying real data to these hypothesises eventualities.
e.g. at what power price point could there be backlash from the general public? What could the ceiling for marginal power costs be for data centres? How could we calculate clearly if the U.S. grid is on course to keep up with AI data centre demand or not, and when and where could this struggle for power supply first appear?
*******************************************************************************************************************************
References
https://fred.stlouisfed.org/tags/series?t=electricity
https://www.goldmansachs.com/insights/articles/us-data-center-power-demand-projected-to-double-by-2027
https://www.nerc.com/globalassets/our-work/assessments/nerc_ltra_2025.pdf
https://energyanalytics.org/research/renewable-power-capacity
https://www.pv-magazine.com/2026/05/12/u-s-transformer-market-faces-severe-supply-constraints/
https://finance.yahoo.com/sectors/technology/articles/half-planned-us-data-center-150928890.html
https://www.industrialsage.com/power-transformer-lead-times-us-grid-shortage/
https://www.nrucfc.coop/content/solutions/en/stories/energy-tech/transformers-are-facing-major-cost--supply-chain-pressures.html
https://www.utilitydive.com/news/energy-infrastructure-transmission-transformers-civil-engineers/743698/
https://chargeduppro.com/post/data-center-transformer-shortage-power-bottleneck-industrial-property-2026
https://www.pv-tech.org/80-of-energy-projects-withdraw-from-inefficient-us-grid-queues/
https://emp.lbl.gov/publications/queued-2025-edition-characteristics
https://www.allianceforcompetitivepower.org/blog/interconnection-queues-explained-reliability-and-prices
https://www.ferc.gov/explainer-interconnection-final-rule
https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000063/gevpressrelease1q26.htm
https://www.power-eng.com/gas/turbines/data-centers-drive-record-surge-in-ge-vernova-power-equipment-orders-as-turbine-slots-tighten-through-2030/
https://www.utilitydive.com/news/after-2-years-ratepayer-pain-political-fallout-georgia-nuclear-vogtle/817792/
https://nuclearcosts.org/track-record/
https://fervoenergy.com/fervo-energy-breaks-ground-on-the-worlds-largest-next-gen-geothermal-project/
https://carboncredits.com/u-s-geothermal-boom-fervo-energy-leads-with-462m-funding-for-cape-station-project/
Comments
Post a Comment