AI is different because the constraint has changed.
For years, the data center industry has talked about scale in terms of megawatts: bigger campuses, bigger substations, bigger utility connections.
That still matters. But AI adds another layer. It asks how much useful compute can be delivered from each megawatt.
Jensen Huang’s language of the “AI factory” is useful here. The facility is no longer just a building that houses servers. It is becoming an industrial production system where chips, memory, networking, racks, cooling, electrical distribution, software and operations all affect the final output.
That is why NVIDIA talks about extreme co-design. The chip cannot be designed in isolation from the rack. The rack cannot be designed in isolation from the cooling system. The cooling system cannot be designed in isolation from the electrical architecture. And the electrical architecture cannot be designed as if grid power is infinite.
Without that level of co-design, we will not turn constrained watts into the maximum amount of useful compute. We also will not deliver the full economic value that customers expect from these facilities.
Grid connections take time. Transmission projects take years. Communities are increasingly aware of the power and water demands associated with large digital infrastructure projects. Meanwhile, utilities are being pulled in every direction by homes, industry, electric vehicles, heat pumps and now AI.
So the real question is not “AC or DC?”
The better question is: what architecture helps us turn constrained power into the maximum amount of useful compute?
This is where higher-voltage DC becomes interesting, and where the conversation shifts from abstract electrical theory to very real constraints driven by AI system design.
NVIDIA has been clear that cabinet densities are going to continue rising sharply. This is not density for its own sake. It is driven by the mathematics of modern AI.
Both training and inference still rely heavily on dense matrix operations that benefit from massive parallelism. Over time, algorithms will undoubtedly become more efficient. Models may become more selective, software will improve, and the industry will find ways to waste less compute. But for now, the practical reality is that AI infrastructure remains highly dependent on GPU systems designed for massive parallel processing.
That parallelism requires tightly coupled GPUs, and those GPUs must communicate with each other at extremely high speeds.
Within NVIDIA’s architecture, NVLink provides that high-speed interconnect. Staying within the most tightly integrated compute domain allows GPUs to operate more like a single coherent system. Moving outside that domain can introduce a material performance penalty, and that changes the economics of AI workloads.
The result is that systems are being designed to keep as much compute as possible within a tightly integrated domain. That drives cabinets to become denser, not because operators want higher numbers on a spec sheet, but because the algorithms and interconnects are pushing us in that direction.
At GTC 2025, Jensen Huang referenced the move toward 600 kW racks. That is not a theoretical curiosity; it is a direction of travel. At the same time, organisations such as the Open Compute Project are already exploring architectures that support AI racks up to 1 MW.
Once you accept that trajectory, the electrical implications become unavoidable. Delivering hundreds of kilowatts into a single cabinet forces a rethink of how power is distributed, converted and managed. That is precisely why higher-voltage DC is back in the conversation.
A 600 kW AI rack at 48 VDC draws around 12,500 amps. At 54 VDC, it is still above 11,000 amps. At 800 VDC, the same rack draws around 750 amps.
The power has not changed. The voltage has.
But the physical implications are enormous.
Lower current means less copper, more manageable busbars, smaller conductors, reduced heat, more practical connectors and more feasible high-density rack design. At some point, current becomes the limiting factor. You can only make the copper so large before the architecture starts to fight against itself.
That does not mean every AI rack will become 800 VDC. It simply explains why the industry is looking seriously at higher-voltage DC again.
Solar PV and battery energy storage systems have already moved into much higher DC voltage classes, commonly around 1,500 VDC in utility-scale applications. That industry has had to learn hard lessons about DC protection, isolation, maintenance, safety, power conversion and controls.
The data center industry should not pretend it is the first sector to face these issues.
We should be humble enough to learn from those who are already living with them.
One of the most practical developments is the Open Compute Project’s Diablo 400 architecture.
The concept is not to rip out every AC system in the data center and replace it with a pure DC world. It is much more pragmatic than that.
A dedicated power rack, or sidecar, sits next to the high-density AI rack. Conventional AC can be brought into the white space and converted into higher-voltage DC close to the IT load. The IT rack can then focus on compute, cooling and networking, while the sidecar handles a specialised part of the power architecture.
That may sound like an incremental step, but incremental steps matter.
Liquid cooling is a useful comparison. Liquid-cooled computing is not new. I trained years ago on Cray systems, and the supercomputing world had already been working with liquid cooling for decades. The Cray-2, introduced in the mid-1980s, used Fluorinert immersion cooling, with its distinctive cooling “waterfall” becoming one of the visual icons of early supercomputing. What changed the mainstream data center conversation was not the invention of liquid cooling. It was the arrival of more modular, repeatable and understandable ways to deploy it.
The Cooling Distribution Unit or simply CDU made liquid cooling approachable.
The sidecar could do something similar for higher-voltage DC.
It gives the industry a bridge. It allows DC to be applied where it solves a real problem, without forcing every data center to become a science project.
This is where I part company with anyone trying to make this a religious argument.
The future will not be all AC. It will not be all DC.
It will be hybrid.
Even the most advanced AI data center still has chillers, pumps, fans, compressors, building systems and other mechanical infrastructure that will remain AC for a long time. The global cooling industry is far larger than the data center market. It is not going to redesign itself around 800 VDC just because AI infrastructure would find that convenient.
At the same time, higher-voltage DC may make very good sense in the places where density, conversion stages, distribution losses and rack integration matter most.
That points towards a more sophisticated architecture: AC where it remains mature and practical, DC where it solves a specific engineering problem, and power electronics acting as the bridge between the two.
Solid-state transformers are one example of where this could become interesting. Companies such as DG Matrix and Polyax are exploring architectures that are less like a conventional transformer and more like a power-routing platform. They could potentially manage AC and DC flows, batteries, renewable inputs, grid interfaces and high-density compute loads through a more flexible common system.
That is promising. It is not magic.
The questions will be the same ones that always matter in critical infrastructure: reliability, protection, maintainability, certification, operator training, spare parts, cost and field experience.
Higher-voltage DC must also be discussed honestly.
DC is not simply AC with a different label.
AC naturally crosses zero volts many times per second, which helps interrupt arcs. DC does not. A sustained DC arc can be harder to interrupt, and that changes the protection philosophy.
This does not mean DC cannot be engineered safely. It absolutely can be. We see that every day in telecoms, electric vehicles, solar PV, battery energy storage systems and industrial applications.
But high-voltage DC in AI data centers will require serious thinking about breakers, disconnects, grounding, insulation monitoring, connectors, fault detection, lockout/tagout procedures, selective coordination and maintenance practices.
Safety cannot be the paragraph at the end of the design narrative. It has to be one of the starting assumptions.
If we want the industry to adopt higher-voltage DC, we need to make it boringly safe, inspectable, maintainable and repeatable.
That is not a marketing challenge. It is an engineering challenge.
The growing interest in higher-voltage DC does not mean the industry has already chosen its future.
Many of the most important questions remain unanswered. Standards are still evolving. Safety frameworks must mature. Operators, manufacturers and regulators need confidence that new architectures can be deployed reliably at scale.
What is becoming clear, however, is that AI is pushing power architecture from a background consideration into a strategic design decision.
The real opportunity now is to ensure this transition happens in a way that strengthens both Europe's AI ambitions and its energy goals.
In Part 3, I'll explore why Europe could be uniquely positioned to lead this next phase of data center power innovation, and what the industry needs to do next.
Read up on Part 1.