[DigitalToday reporter Jinju Hong (홍진주)] The key variable in the race to build AI infrastructure is shifting from semiconductors and servers to securing power. As AI moves beyond experimentation into large-scale operations, business success is expected to depend not only on how quickly data centres are built but also on whether stable power can be supplied when needed.

Tech outlet TechRadar reported on Sept. 10 that as AI data centres expand to a far larger scale than conventional commercial and industrial facilities, power supply capacity is emerging as a key factor determining site selection, start-up timing and the potential to recover investment.

AI data centres consume massive amounts of electricity concentrated in specific areas, and they can scarcely tolerate power interruptions. This is making it important not just whether overall generation is sufficient, but whether firm, immediately usable power can be secured in the needed place and at the needed time.

Natural gas is also expected to play a bigger role. Unlike solar or wind power, whose output varies with weather, natural gas can supply electricity around the clock and is drawing attention as a power source that can support large-scale AI workloads.

PwC scenario analysis forecasts that AI-related demand for natural gas will rise to 5.2 billion cubic feet per day by 2030 even in the most conservative case. That would be a large increase from the current level of about 1.6 billion cubic feet per day. It also projected demand could expand to 7.6 billion to 11.5 billion cubic feet per day by 2035.

A problem is that having natural gas does not immediately translate into power that can be used at a data centre. To put gas into actual power generation requires multiple stages of infrastructure, from production facilities to pipelines, storage facilities, pressure maintenance equipment and generation facilities. Ultimately, the power issue for AI data centres is shifting from a simple lack of generation to whether electricity can be delivered to end users.

Permits, pipeline rights-of-way, grid connections, generation turbines, water supplies and skilled labour are also cited as major bottlenecks in data-centre development. Even if AI companies and data-centre developers invest tens of billions of dollars, there are limits to adding the needed infrastructure in a short period of time.

As a result, criteria for selecting data-centre sites are changing. In the past, land, telecom networks and access to the power grid were key considerations. Now, analysis says developers need to assess from an early stage whether stable energy can actually be secured within a fixed schedule.

That implies the value of already-built pipelines, permitted corridors, storage facilities and generation capacity that can be brought online immediately could rise. Securing energy infrastructure that can be used right away could become more important for data-centre development speed than securing cheaper power sources over the long term.

New approaches to securing power are also emerging. One example is behind-the-meter generation, where a data centre does not rely only on existing grid connections but builds natural-gas generation facilities on-site or nearby to receive power directly.

PwC estimated that more than 30 percent of AI-related natural gas demand by 2035 could be supplied through this type of approach. Other alternatives cited include building dedicated pipeline laterals together with generation facilities, or connecting natural gas supply, transport, storage and generation with data-centre loads into a single system.

In this way, expanding AI infrastructure is also creating a need for different cooperation structures. It requires hyperscalers and data-centre developers to work together with utility companies, natural gas suppliers and pipeline operators.

It is difficult to meet the speed demanded by AI data centres if each player optimises only its own area. The explanation is that the entire process, from gas supply to transport and generation, power supply and final data-centre start-up, must be coordinated as a single project.

Ultimately, the AI race is not expected to end with securing more powerful chips and servers. Even if land and equipment for data centres are secured, AI computation cannot begin if the required power is not connected on time.

TechRadar pointed out that AI is a technological revolution while also rapidly turning into a physical infrastructure problem. In the AI industry ahead, not only how much computing capacity is secured but also how quickly power can be secured and operations started is expected to be a key variable in determining when profits are generated.