AI data centers prepare to switch from alternating to direct current

AI workloads are changing not only servers but the entire power setup of data centers. Site operators and equipment suppliers are seriously discussing a return to direct current, because when powering accelerators, every extra conversion stage means losses, heat and higher costs.

Why the design is changing

Today, electricity in most data centers passes through a long chain of conversions. The grid supplies alternating current, which is then converted to direct current several times: for batteries, uninterruptible power supplies, server power supplies and, finally, on the boards themselves. That was tolerable for classic servers, but the AI era has sharply raised power density per rack. When a single GPU cabinet draws tens of kilowatts, and soon hundreds, even a small percentage of losses turns into serious money and extra cooling requirements.

Historically, alternating current won not because it is ideal everywhere, but because it was easier and cheaper to transmit over long distances. That remains true for city grids. Inside a modern data center, however, conditions are different: distances are short, loads are predictable, and almost all computing hardware ends up running on direct current anyway. So the industry is asking a pragmatic question: if the end consumer is electronics, why keep so many intermediate stages between the power inlet and the chips?

What is changing in data centers

Interest in direct current is driven by economics, not fashion. Operators are trying to squeeze more compute out of the same megawatts without endlessly building new power hubs. If some conversions are removed, the power system can become more efficient, which means the same facility could serve more AI accelerators without immediately expanding external infrastructure. For hyperscalers and large colocation sites this matters especially: power, rather than racks or servers, is increasingly the main constraint on growth.

The potential advantages of such a shift look like this:

• fewer energy losses at each conversion stage
• lower heat output in power circuits and less load on cooling
• simpler rack design for powerful GPU and AI systems
• a better chance of fitting within existing power limits without an expensive rebuild of the site

But this shift will not be fast. Power distribution architecture has to change, standards have to be agreed, and compatibility has to be tested with existing UPS units, batteries and server platforms. Moreover, data centers are not built for one or two years: many sites have invested huge sums in current AC schemes and will not throw them away for the sake of an elegant idea. In practice, the market will most likely move through hybrid models, pilot zones and new AI clusters designed for DC from the start.

Why this is not a step into the past

At first glance the news sounds like an energy rollback of a hundred years, but the comparison is misleading. This is not about cities again transmitting direct current to homes and offices on a large scale. The discussion concerns local architecture inside data centers, where conditions differ greatly from the citywide grid. In such systems, direct current has a different profile of advantages: it works over a short run, next to the load, in an environment where every percentage point of efficiency matters.

In a sense, the industry is not returning to the past but shedding a legacy of an era when server power was far more modest. Batteries already store energy as direct current, solar panels and some power electronics are also closer to DC logic, and modern AI racks demand ever more direct, dense power delivery. So the shift looks less like romance about Edison and more like engineering optimization for a new reality: accelerators are getting more expensive, power is scarce, and any extra conversion becomes too costly.

What this means

If the trend takes hold, competition in AI infrastructure will shift from chips alone to power engineering. The winners will be the operators and vendors that can rebuild data center electrical systems for dense computing faster and lower the cost of each megawatt for AI workloads.