The Future of AI Infrastructure isn’t just More Power – It’s Smarter Power Distribution

If you ask ten people in the data center industry what the biggest AI infrastructure challenge is, nine will likely have an answer centered around power.

Power availability, higher rack densities, and speed to market are all critical parameters to the industry … but from an engineering perspective, I believe we may be asking the wrong questions. 

The question isn’t simply how to generate more power – it’s how we deliver it more efficiently. How do we deliver more compute with the same total load?

Redesigning the electrical architecture to have lower losses (namely fewer transformations and lower ampacities), while future-proofing advancements is a task we can’t ignore.

Consider the scale of what’s ahead: Gartner projects that worldwide data center power demand will grow 27% in 2026 alone, from 104 GW to 132 GW, and to nearly 290 GW by 2030. In the U.S., Goldman Sachs Research expects demand to nearly double, from 31 GW in 2025 to 66 GW in 2027. We cannot simply maintain pace to keep up with that demand; we must get ahead of it. 

The Path Power Takes Today

When most people think about power infrastructure, they may picture just a transformer, but that’s just one piece of a much larger system.

Power typically enters a data center campus at medium voltage (generally 13.8kV – 34.5kV). It passes through transformers, switchgear, UPS systems, and power distribution units before arriving at the rack. 

Along that journey, power is converted several times. Whether voltage magnitude (a step down transformer), rectification (AC → DC), or via an inverter (DC → AC), losses occur at every step. Every additional piece of equipment also consumes space, introduces complexity, and requires maintenance over the life of the facility. 

For years, those tradeoffs made sense, but today’s AI and HPC environments are changing that equation. 

AI is Changing the Design Rules

Traditional enterprise workloads were relatively predictable by comparing historical trends. Around holidays, large events (think Super Bowl or Prime Day), the ramping of equipment could be predicted with high confidence down to the hour. 

AI is different, and workload shifts can cause equipment and operators to see dramatic changes in power, surpassing 20% swings in seconds. This isn’t a hypothetical – Meta’s The Llama 3 Herd of Models paper mentions challenges with power fluctuations, and that is “only” a 24,000 H100 Cluster (30MW of IT capacity).

As rack densities continue to increase, power demands are growing rapidly. Customers want to deploy more compute within the same footprint while maintaining reliability and uptime AFCOM found that average rack density more than doubled from 7 kW in 2021 to 16 kW in 2025 … and at Data Centre World London 2026, consultants said they are already designing racks for U.S. clients expected to draw as much as 2.2 MW per rack within the next five years. 

From an engineering standpoint, these changes force us to rethink the electrical system – not because the existing architecture is broken, but because it was designed for a different generation of equipment. 

Think in Systems, Not Products

One of the biggest mistakes we make in our industry is evaluating equipment for one component at a time. Power systems don’t operate that way; one equipment selection can have an impact three modules up or downstream.