Global technology companies are investing hundreds of billions of dollars in artificial intelligence, yet a significant proportion of the planned data centres in the US may never get off the drawing board. According to estimates by Bernstein Research, only around 35% of the announced capacity will actually be built. The greatest imbalance is evident in Texas, where applications for large consumers to connect to the electricity grid have exceeded 474 gigawatts — more than five times the state’s record peak electricity consumption. An analysis by UkrMedia, based on data from Bernstein, ERCOT and the International Energy Agency, shows how the technological boom is simultaneously creating a real infrastructure deficit and leading to exaggerated forecasts of future demand.
- Analysts at Bernstein estimate that only around 35% of the announced US data centre capacity will actually be built.
- In Texas, applications for connection from large consumers have reached 474 GW — 5.2 times the state’s peak demand record.
- Approximately 90% of the declared capacity in the Texas pipeline is associated with data centres.
- Some of the projects do not have a guaranteed electricity supply, permits or confirmed customers.
- The IEA forecasts that global electricity consumption by data centres will rise from 485 to 950 TWh by 2030.
- Experts are warning of the risks of incorrect planning of electricity networks due to inflated demand forecasts.
Where did these «phantom» data centres come from?
The boom in generative artificial intelligence has led to the large-scale development of computing infrastructure. Training and running new models requires powerful servers, specialised processors, cooling systems and an uninterrupted power supply.
This market has seen the emergence not only of technology corporations with many years’ experience, but also of new property developers announcing the construction of huge complexes.
The problem is that simply presenting a project does not guarantee its implementation. Some companies claim to have a capacity of several gigawatts, yet they have no connection to the electricity grid, have not completed the necessary licensing procedures, or have no confirmed customers.
In the industry, an unofficial term has even emerged for such ambitious claims bragawatts — a combination of the English words «brag» and «watts».
Another phenomenon is «phantom» projects. A developer may submit several applications at once to connect future data centres in different locations, even though they will ultimately only go ahead with one of them.
As a result, the electricity network operator sees a significantly higher potential demand than is likely to materialise in reality.
Analysts consider only 35% of the planned capacity to be realistic
Bernstein Research analysed thousands of US data centre projects using data from the specialist platform Aterio. The assessment took into account the readiness of construction, the availability of electricity, the companies’ track record and the likelihood of the announced plans being realised.
According to the analysis, only approximately 35% of the announced data centre capacity in the US has sufficient potential for implementation.
This does not mean that 65% projects have already been cancelled. This is a forecast: according to analysts, a significant proportion of the planned construction projects will not actually be brought into operation.
Bernstein’s findings have already drawn criticism from some industry players. In particular, the management of O’Leary Digital has stated that such rankings may underestimate early-stage projects, as analysts do not always have access to confidential agreements, financial documents and engineering documentation.
Therefore, the 35% rating is an important risk indicator, but it is not a definitive list of data centres that will or will not be built.
In Texas, electricity demand was 5.2 times higher than the previous record
The situation was most telling in the state of Texas — one of the major centres for the development of American artificial intelligence infrastructure.
According to the electricity system operator ERCOT, as of June 2026, the total capacity of connection applications from large consumers had reached approximately 474 GW. Around 90% of these applications were for data centres.
By way of comparison, the record peak load on the ERCOT power system, provisionally recorded on 22 July 2026, stood at 91.134 GW.
Consequently, the volume of requests was approximately 5.2 times higher than the highest recorded peak electricity consumption in the state.
What the UkrMedia comparison reveals
| Indicator. | Power |
|---|---|
| All connection applications from large consumers, June 2026 | 474 GW |
| Of these, data centres account for an estimated 90% | 427 GW |
| ERCOT’s record peak demand, July 2026 | 91.134 GW |
| The ratio of all applications to record demand | 5.2 times |
Calculation by UkrMedia based on ERCOT data. The share of data centres is rounded. Connection applications do not represent actual consumption and may include alternative or duplicate projects. Peak demand is also not equivalent to the state’s total available generation capacity.
It was precisely this difference that was one of the reasons why the Texas authorities launched an investigation into the planned data centres.
Regulators have begun to require developers to provide more detailed information on electricity supply, water sources, cooling and funding.
The aim of the review is to distinguish genuine investment projects from preliminary proposals that may never progress to the construction stage.
Why overestimated forecasts are dangerous for electricity grids
The construction of high-voltage power lines, substations and new generating capacity requires significant capital investment and many years of preparation.
If the electricity network operator plans its infrastructure to accommodate all the proposed data centres, some of the equipment may remain unused once the projects are completed. This creates a risk of inefficient investment and additional costs, which, under certain circumstances, could be passed on to consumers in the form of higher tariffs.
The opposite mistake is also dangerous. If the actual growth in demand is underestimated, the existing infrastructure will be unable to ensure the rapid connection of new data centres.
The International Energy Agency explicitly highlighted this issue in its report dated 14 October 2026. Experts emphasised that queues for grid connection often contain duplicate and speculative applications, which make it more difficult to reliably forecast the needs of electricity grids.
Consequently, what matters most is not the total capacity announced in megawatts, but the readiness of a specific project: confirmed funding, technical specifications, permits and a realistic commissioning schedule.
Despite «phantom» projects, AI’s energy consumption will continue to rise
Problems with the roll-out of data centres do not mean that global demand for electricity for artificial intelligence stopped growing.
According to the International Energy Agency’s updated forecast, data centres worldwide will consume around 485 TWh of electricity in 2025.
By 2030, this figure could reach 950 TWh — almost double the current level. Moreover, energy consumption by data centres specialising specifically in artificial intelligence could increase by approximately threefold.
The MEA also forecasts that capital expenditure by the largest data centre operators will exceed $700 billion in 2026.
At the same time, these figures relate to projected investment and consumption, rather than capacity that has already been built.
The actual pace of development will be influenced by the cost of loans, the availability of electricity, the supply of microchips, planning permission and the cost-effectiveness of the AI services themselves.
A comparison with Ukraine: the scale of global demand
According to data from the NEURC published on the Energy Map portal, total electricity consumption in Ukraine in 2025 stood at 80.2 TWh. If we compare this figure with the IEA’s forecast for global data centres for 2030, the difference is approximately 11.8 times.
In other words, according to the baseline forecast, data centres around the world could, in four years’ time, consume as much electricity in a single year as Ukraine has used in approximately twelve years, based on 2025 consumption levels.
This comparison illustrates the scale of the global infrastructure, rather than a forecast for the Ukrainian power system. In 2025, consumption in Ukraine was significantly lower than pre-war levels due to the consequences of the Russian invasion.
For Ukraine, international experience demonstrates, first and foremost, the importance of verifying the actual electricity requirements of major investors.
In a country where energy infrastructure As it is regularly subjected to Russian attacks, sound planning for new high-power consumers is of particular importance. At the same time, the available international data provide no grounds for linking the current electricity shortage in Ukraine to the development of AI data centres.
Could the data centre market face a crisis?
The main risk lies in the gap between the scale of the announced investments and the projects’ actual readiness for launch.
If some developers fail to secure funding, grid connection or the necessary permits, the announced capacity will remain unrealised. This could affect the expectations of server manufacturers, energy equipment suppliers and investors.
However, a complete halt to investment in AI This is not evident from the data provided. Major technology companies continue to expand their computing capacity, whilst demand for new services remains strong.
When assessing the sector’s actual development, four indicators are becoming more important than high-profile announcements: completed construction, actually connected electrical capacity, installed server equipment and the actual utilisation rate of data centres.
Bernstein’s research and data from Texas suggest that there may be a significant discrepancy between the publicly announced scale of AI infrastructure and its actual roll-out. The extent of this discrepancy in the long term will depend on the implementation of specific projects.
