I love history, especially historical dramas like the 2017 film The Current War, which depicts the 19^th century competition between Thomas Edison and George Westinghouse, assisted by Nikola Tesla, to decide which electrical power delivery system would be used in the United States.

Westinghouse won out with AC.

But as they say, it is a marathon, not a sprint,  and Edison may yet have the last laugh.

Many conventional AC based data center designs start to become increasingly difficult to scale as rack densities move through the 150–200 kW range, and the challenge becomes unavoidable as designs head toward 400 kW, 600 kW and eventually megawatt class racks.

A single NVIDIA GB300 NVL72 rack, with 72 Blackwell Ultra GPUs, is already in the 132–140 kW range depending on configuration. That puts today’s AI infrastructure uncomfortably close to the point where traditional power distribution models become difficult to scale.

Roadmaps are already discussing racks approaching 600 kW, with some vendors publicly discussing eventual megawatt class AI racks just a few years away.

The problem isn’t that AC is suddenly “bad.” The problem is that AI has changed the rules.

Ten years ago, a 10 kW rack was considered dense.

Today we’re deploying 140 kW AI racks.

Tomorrow we’re talking about 600 kW and beyond. The electrical architecture that powered enterprise IT for decades simply wasn’t designed for this.

Depending on the number of conversion stages and where the measurement boundary is drawn, traditional AC-based delivery can lose meaningful energy before power reaches the silicon. That loss becomes increasingly painful at scale, as every GPU, CPU and SSD ultimately runs on DC power. In today’s data centers, electricity arrives as AC, is converted to DC inside every server power supply, and then stepped down again into multiple voltage rails for the components themselves. Every conversion introduces losses in the form of heat. As organisations are chasing ever lower PUE values, native DC power distribution could become another important tool in improving overall data center efficiency.

So what is the answer? Rather than performing repeated AC/DC conversions across thousands of server power supplies, the aim is to move more of that conversion into fewer, higher efficiency stages and distribute high voltage DC closer to the load. The rack still needs DC/DC conversion, but the overall architecture becomes simpler, denser and more efficient.

Much like IPv6, the industry has known the likely destination for years. But organisations rarely replace working infrastructure until the economics or necessity make the decision for them.

Just like we have seen the gradual move from air to liquid cooling due to necessity, so will we see the transition from AC to DC for the same reasons.

But this isn’t going to happen overnight but in a series of stages.

Stage 1: Hybrid Deployments

Existing AC powered data centers won’t suddenly disappear, so vendors are introducing “sidecar” power cabinets that convert facility AC into 800V DC immediately adjacent to the AI rack. This allows next generation DC powered AI systems to coexist within today’s AC facilities. Some early roadmaps including NVIDIA’s future Rubin Ultra NVL576 / Kyber rack architecture (Expected H2 2027) already show 800VDC sidecar power cabinets supporting AI racks up to 600 kW.

Stage 2: DC Distribution

Stage 2 is likely to be tighter integration between DC distribution and energy storage. Batteries already store energy as DC, yet many conventional designs convert through multiple AC/DC and DC/AC stages before power reaches the IT load. A DC oriented architecture gives operators the option to simplify that chain.

Stage 3: Native DC Power Delivery

Stage 3 is native DC power delivery. AC still arrives from the grid, but it is converted into DC far earlier in the facility power chain, potentially through high efficiency rectification or future solid-state transformer technologies. The goal is not magic; it is fewer conversion stages, less copper, less heat and better end to end efficiency.

So nearly 140 years after losing the current war, Thomas Edison may finally win his rematch.

Not in our homes, not in the National Grid, but inside the AI data centers powering the next generation of artificial intelligence, now that’s memorable!

This isn’t really a story about AC versus DC. It’s a story about AI changing assumptions that have remained largely untouched for decades.

Cooling has changed.

Networking has changed.

Storage has changed.

Power is simply the next domino to fall.