State-of-the-art chips, billions to invest, a well-located site: none of it is enough if the power does not arrive. With artificial intelligence, the data center is becoming an industrial facility whose critical resource must be negotiated long before the servers are installed. Behind the race to develop models, another, less spectacular competition is emerging: for grid connections, cooling and construction projects that can meet their deadlines.
Looking ahead to September 2026, this battle is a strategic issue. The following analysis draws on facts documented through June 2024; the developments envisaged for 2026 are forward-looking assessments, not outcomes presented as established facts.
Available power matters as much as chips
A conventional data center already powers equipment around the clock, with backup and cooling systems. AI intensifies this requirement: training a large model involves clusters of accelerators, while its everyday use, known as inference, multiplies the workload. As services become more widespread, consumption no longer depends solely on a few exceptionally demanding training runs, but also on millions of repeated uses.
In its Electricity 2024 report, the International Energy Agency estimated that global electricity consumption by data centers, AI and cryptocurrencies could exceed 1,000 terawatt-hours in 2026, compared with around 460 in 2022. This scope does not cover AI alone, and the estimate was a projection. Nevertheless, it illustrated a growing strain: digital demand that could rise rapidly, set against electricity infrastructure that takes years to transform.
The problem is primarily local. A country may generate enough electricity over the year without having the capacity in a particular area to supply a new campus. Between the power plant and the servers, power lines, substations, transformers and protective equipment are needed. National energy availability therefore guarantees neither power in the right place nor its delivery at the right time.
Grid connections become a competitive advantage
For an investor, land already served by robust electricity infrastructure may be worth more than a cheaper plot far from the grid. The connection timetable determines future revenue: a completed building without sufficient power remains an idle asset. And reserving IT equipment whose installation is delayed risks technological obsolescence before it even enters full operation.
Warning signs were already visible before 2026. In Ireland, the regulator adopted a more restrictive approach to data center connections as early as 2021, taking particular account of their location and constraints on the electricity system. In Singapore, a pause on new projects introduced in 2019 was followed by a selective resumption prioritizing efficiency and sustainability. Two different markets, one shared lesson: the welcome is no longer unconditional.
Competition between regions could therefore shift. Alongside tax incentives and proximity to customers, grid strength, access to low-carbon generation and the administrative capacity to process applications are becoming important factors. But not all computing workloads can move so easily: some training can take place far from users, whereas interactive services require low latency. Sovereignty and data location requirements also limit the options.
Cooling increasingly dense server installations
The second battle is taking place in server rooms. Accelerators concentrate substantial heat in a small space. For some configurations, blowing more air is no longer enough, or becomes inefficient. Liquid cooling, particularly through plates in contact with components, is therefore becoming more attractive. It does, however, require suitable fluid circuits, maintenance procedures and design.
An available building is therefore not automatically ready for AI. Its power distribution, heat removal capacity or floor load capacity may necessitate extensive work. Converting an existing facility requires a detailed technical assessment: available floor space alone says nothing about how many accelerators the site can actually accommodate.
Water makes these trade-offs even more sensitive. Not all systems withdraw or consume the same volumes: evaporative cooling, closed loops and dry cooling solutions involve different compromises. Reducing water use can increase electricity consumption under certain conditions. In a drought-prone watershed, a project must therefore be assessed locally, rather than solely through a flattering annual average.
Building quickly without shifting the costs
The promise of a rapidly delivered campus runs up against a very real industrial supply chain. Transformers, high-voltage switchgear, backup generators and cooling equipment all require manufacturing capacity. Permits, grid works and the availability of qualified contractors add further constraints. Accelerating construction of the building does not necessarily shorten the project’s critical path, which may lie several kilometers away in an electrical substation.
Long-term power purchase agreements can support new renewable capacity and provide price visibility. On their own, however, they create neither an available grid connection nor a carbon-free power supply at every hour. Buying as much renewable electricity as is consumed over the year is not the same as operating without emissions at all times. This distinction is essential when assessing environmental claims.
The convergence of digital infrastructure and electricity generation was already evident in March 2024, when AWS announced the acquisition of a data center campus adjacent to the Susquehanna nuclear power plant in Pennsylvania. This type of deal demonstrates the strategic value of nearby generation. It also raises questions about regulation and cost sharing: who pays for shared infrastructure, and what are the consequences for other users?
For regions, choosing rather than merely attracting
Hosting a data center can bring investment, tax revenue and substantial construction activity. Permanent jobs, however, are generally fewer than the size of the buildings might suggest. A local authority must weigh these benefits against the power committed, competing land uses and the needs of other industries pursuing electrification.
Heat recovery can enhance local benefits, provided there are nearby users, a network and a viable business model. Another potential avenue is to negotiate a degree of electricity demand flexibility for tasks that can be deferred. Not every service can be interrupted, but some computing workloads could be shifted in time. These commitments must be measurable, rather than merely promised.
What next? Looking ahead to 2026, the advantage could lie less with the regions announcing the most projects than with those securing the entire chain: energy, grid connections, cooling, permits and public acceptance. Improvements in chip and model efficiency may moderate demand, without guaranteeing that they will offset the surge in usage. For public authorities and operators alike, the key question is becoming: what value can be created with each available megawatt, and on what terms for society?
