Artificial intelligence already has chips, computing power, and a global investment race behind it. But the next bottleneck won't necessarily be within the servers themselves, but rather in the physical infrastructure capable of powering them.
“This data center race will no longer be won or lost solely based on computing power or the technology of NVIDIA or any other manufacturer, but rather on energy,” said Christian Soler , Director of Product and Applications at Power Electronics , during the III Energy Storage and Value Chain Conference .
“We have the chips and the technology to train artificial intelligence. However, energy is not guaranteed in terms of quantity or stability, nor does it seem likely to be available in time,” he added.
The projected growth necessitates looking beyond software. According to the executive, global data center consumption could double before the end of the decade: what the industry built in 65 years could be replicated in just 5 .
The scale of the new projects confirms the leap. GPU clusters are already operating in the 100-300 MW range, while Power Electronics is beginning to see developments in the gigawatt range.
In the United States, data centers currently account for between 3% and 4% of electricity demand . By the end of the decade, that share could reach between 8% and 9%.
Spain is also entering this race. Spain DC anticipates that demand associated with these assets will increase from around 200 MW in 2024 to approximately 730 MW in 2026. Furthermore, its latest projections raise the potential capacity to 2,537 MW by 2030, in a scenario of strong investment expansion.
The question is no longer just how many data centers can be built, but how many will be able to connect, operate stably, and absorb their peak consumption without compromising the network.
From backup to data center enabler
An interesting highlight of the day was the shift in storage's role. Until now, its most widespread function within a data center was related to ensuring continuity of supply. In this new scenario, it becomes an integral part of the critical operational architecture.
“The logic is that the PCS charges the battery when there is a surplus and discharges it when demand requires it, providing that energy that the turbine or the grid cannot provide as quickly,” Soler explained.
Power Electronics' approach is part of a broader shift in the Spanish market: batteries are no longer seen merely as a complement to self-consumption or as insurance against power outages, but as a strategic infrastructure for companies with sensitive processes, high consumption, and a need for operational continuity.
The shift began to accelerate after the blackout on April 28, 2025. Since then, the conversation has changed scale, from a tool designed to optimize consumption to occupying a central place in energy security, efficiency and flexibility strategies.
According to APPA's 2025 Annual Report , Spain added 339 MWh of behind-the-meter batteries , compared to 155 MWh in 2024. The year-on-year growth was 119%, with a strong presence from the commercial and industrial segment, which reached 181 MWh and surpassed the residential segment.
AI doesn't consume: it hits the network
In traditional data centers, the typical power of a rack ranged from 5 to 15 kW, with relatively stable profiles and simpler cooling systems. But with AI , the leap is abrupt, and new racks can reach 130 kW and require liquid cooling due to the high power density they concentrate.
“Artificial intelligence does not consume energy linearly, but rather constantly impacts the grid,” Soler summarized.
During the training phase of a model, GPUs work synchronously for weeks or even months. In this process, power consumption can jump from zero to the full capacity of the campus in extremely short periods. This behavior generates transients, resonances, and disturbances that can extend beyond the perimeter of the data center itself and spread to the distribution network or other connected assets.
And it's no small matter that today digital deployment is advancing faster than electrical infrastructure. According to Soler, it's now possible to build a data center in less than two years, but the challenge comes later: how to supply the necessary power.
Nuclear power is usually excluded from the equation due to its development time, which can take around a decade. Conventional thermal power plants also involve lengthy timeframes, of five years or more.
Connecting the campus directly to the grid seems like the natural alternative, but capacity for 300 MW or higher isn't always available . That's why many developers are turning to gas turbines installed near the campus. They are modular, quick to deploy, and can provide large blocks of power.
“These turbines alone will not be enough to power a data center,” Soler said.
This gap is where storage systems and static converters come in, capable of reacting in the order of milliseconds and dampening the peaks that the grid or turbines cannot absorb at the same speed.
It is also worth noting that Power Electronics is already working on different architectures to respond to this new demand.
The first involves measuring the load consumption and injecting or absorbing power, and its effectiveness depends on the storage system, but also on the short-circuit power, the distance and possible communication delays.
The second alternative incorporates a relevant impedance towards the campus and a grid-forming converter connected in parallel to the load. Instead of operating with power references, it works with voltage and frequency, allowing it to maintain stable electrical values.
The third way is double conversion , where The data center is virtually decoupled from the grid using converters and batteries. Meanwhile, the fourth architecture is particularly prevalent in the United States : massive campuses without grid connection capabilities, powered by turbines and grid-forming converters that create a stable internal network.
According to the presentation, some solutions can reduce these impacts by around 60%. With in-line inductors, the reduction can reach between 80% and 90%. Double conversion, meanwhile, allows for almost complete decoupling of the installation.