The Potential AI Choke Point

The conflict in the Middle East has put in quite sharp perspective the importance of energy, specifically oil and natural gas. Most of the focus of this concern has, understandably, been around household consumer sensitivities like the price at the pump and the reduction of flights due to rationing of air fuel. We though would like to explore another energy issue; the rapid expansion of electricity use to power AI. Whilst AI inevitably is also subject to geopolitical fallout in international energy markets, this is certainly not the central concern when it comes to how the AI surge is to be powered.

AI and Datacenters

For the purposes of examining the use of electricity for AI, datacenters provide the best way of measuring it. While of course a portion of datacenter usage does not directly pertain to AI applications, in recent times the overwhelming bulk of datacenter capacity is devoted directly to serving AI and in many cases the data from classical, pre-AI, datacenter storage is also used in conjunction with AI models. So for practical purposes datacenters are a solid, “bricks and mortar” type economic representation of the AI economy and very specifically its energy consumption.

Hyper Scaling: Hyper Variable Future Projections

Look at the following two projections for the increase in datacenter electricity consumption in the US from 2024 to 2030. The IEA (International Energy Agency) in an April 2025 report, “Energy and AI” estimated a base case increase of 243 TWh (Terawatt-hours) and an upper range estimate increase of 352 TWh. A McKinsey report in August 2025, “The data center balance: How US states can navigate the opportunities and challenges” had its lower case estimate increase over the same period by 334 TWh and its upper range estimate by 582 TWh. It is unusual to see such variation across industry estimates, which is a reflection, of the amount of guess work involved in estimating how much impact AI adoption will have on society. The term hyper scalers has become fashionable to describe companies able to invest massive capital in their AI infrastructure build out. The varied estimates quoted above reflect that no one really has a definitive model to project what hyper scaling means in practice. It doesn’t distill nicely to some simple calculation derived from Moore’s Law or the number of GPUs.

Accepting the vagueness in the estimated datacenter electricity requirement, these figures have to be viewed in the broader context of power generation. The US total electricity used in 2024 was 4,110 TWh, 4.3% of which was consumed by AI. Electricity used in the US over the last 10 years has increased with a compound annual growth of 0.73%. Assuming equivalent growth would give a projected total use of 4,350 TWh by 2030. On this basis AI consumption by 2030 could represent between 9.8% to 17.5% based on the lower to upper range estimates.

The Political Dance With Public Perception

Politicians have successfully played the AI economic growth story. Securing the location of datacenters in their districts, brings with it much needed jobs to their area. In other words, securing datacenter locations, can help secure political office. That said, these same politicians and in many cases their challengers, are now becoming sensitive to the potential implications of a developing negative public perception and anger that datacenters are causing problems. From water consumption, to heat effects, to noise, to rising household electricity bills. This feeling of resentment is particularly potent among those, who that have not been the direct beneficiaries of having a datacenter that has brought jobs and improved the local economy.

This is a PR problem for datacenters. The image of stealing the general populaces’ electricity, along with operational concerns regarding peak draw insecurities from taking electricity from the grid, is encouraging new datacenter projects to build electricity generation as part of the datacenter development. The datacenter has its own supply and therefore does not draw on the grid and in so doing helps mitigate negativity. This is becoming so important that the concept has got its own anacronym BYOP (Bring Your Own Power).

When, however one takes a look at the numbers, it would seem that BYOP is more spin than substance. For example, the McKinsey report referenced earlier predicts a whopping $6.7Tn datacenter spend from 2024-2030. The largest portion of this spend is $3.5Tn on servers (GPUs and CPUs). Interestingly though, only $400Bn of the spend goes towards electricity generation. To put this in context, according to Lazard, the average cost of producing electricity in the US (across all generation types from solar, wind, geothermal, hydroelectric, gas, coal and nuclear) is $113/MWh. Hence the $400Bn spend by 2030 for BYOP datacenters would only pay for 3.5 TWh of the required increased capacity. A veritable drop in the ocean, when compared to the estimated requirement (ranging from 243 – 582 TWh).

If Power Can’t Be Brought Then It Will be Bought

It is important to remember that 2030 is just around the corner. We are looking at a very short-term horizon, one in which the AI juggernauts are likely to remain dominant corporate power-houses with vastly more capital to spend than their counterparts in other industries. Thus it is unlikely that the AI surge will be deprived of the electricity it requires because it can be bought and the large AI companies have plenty of money to buy it. This though will come at the expense of other industry unable to compete for increasingly expensive and less available electricity. The impact of an electricity squeeze could significantly impact businesses directly or the economy more broadly through higher inflation. There is a question if this impact would extend into a reduction of potential AI business customers.

Electricity Generation: Fertile Ground for New Investment

There is real potential for the US to significantly improve the amount of electricity it produces per capita. In 2025 for example this figure was 13,100 kWh/Capita, the equivalent figure for Iceland, the world’s largest electricity producer per capita stood at 55,000 kWh/Capita. Interestingly Iceland’s electricity production is almost 100% renewable with 73% coming from geothermal and 27% from hydroelectric and furthermore it is nationally self-reliant making it immune to geopolitical energy market disruptions, a major plus from a national security perspective. It is not a coincidence that Iceland is also a leading global hub for datacenters.

For investors trying to plan out investments with a five-year horizon, perhaps the best new AI investment opportunities actually now lie beyond AI directly and more in the field of electricity production and improved electrical efficiency (according to the EIA transmission and distribution losses accounted for almost 6% of all electricity produced in the US in 2025).

To make this kind of investment, we would recommend thinking beyond the traditional investment sleeve for electricity generation as utilities. This slot will likely already have been filled in many investor portfolios, as a steady dividend provider. The developing AI electricity requirement is different in that it is more urgent and consequently has got the potential to spring innovative more dynamic solutions to both power generation and efficiency. In a way this offers investors to approach the electricity market from two different angles, the traditional conservative utilities sector one, and a more adventurous, speculative Tech sector angle. The question is, given the pressures and short time horizon involved if the market players can step up sufficiently quickly to avoid throttling the expansion of the AI economy. ■

Image Credit: Analogicus at Pixabay.