AI's Power Shortage Is Reinventing the Transformer

Transformers are supposed to be infrastructure: heavy, durable and largely invisible. AI has turned them into schedule risk. Enphase’s answer is to make power conversion behave less like a piece of iron and more like a computing system.

The rack is starting to reach back toward the grid

AI has made the electrical path inside a data center worth redesigning. GPU racks are climbing toward power densities that make every conversion stage larger, hotter and more consequential. At the same time, conventional transformers and switchgear have become procurement constraints for projects already waiting on utility interconnections.

Enphase Energy is betting that the answer is not simply to manufacture more of the old stack. The company has begun building 4-kilowatt power modules in Texas for what it calls the IQ Solid-State Transformer, a modular platform designed to convert medium-voltage AC directly to 800-volt DC for AI data centers.

Hundreds of modules operate together. Enphase says full systems can scale to five megawatts and respond to changing AI loads in less than a millisecond. Full-scale demonstrations are planned for November, followed by customer pilots.

The device is still a development program, not a proven replacement for conventional data-center power systems. But its architecture points toward a larger change.

AI is forcing the data center to treat electricity as a software-defined system rather than a sequence of fixed boxes.

A transformer becomes hundreds of coordinated switches

A conventional transformer changes voltage through electromagnetic induction in a large magnetic core. A solid-state transformer moves more of that job into high-frequency power electronics, controls and smaller magnetic components. In Enphase’s proposed design, gallium-nitride-based modules are coordinated in series and parallel rather than relying on one centralized conversion block.

The first architecture announced in April used 342 intelligent modules in a 1.25-megawatt rack. The newer manufacturing update says the platform is being expanded toward racks as large as five megawatts.

Why bother? Because AI loads are unusually dynamic. Thousands of accelerators can change their power draw quickly as workloads synchronize. Conventional systems often compensate with layers of conversion and battery buffering. Enphase argues that faster power electronics can absorb more of that volatility upstream while reducing the amount of local battery capacity required near compute racks.

The commercial argument is equally physical. A modular power architecture can be manufactured as repeated smaller units, tested incrementally and scaled by adding modules. That resembles the manufacturing logic that transformed computing itself: replace one bespoke machine with many standardized components coordinated by control software.

The transformer shortage is becoming an architecture opportunity

None of this means the grid’s transformer problem disappears. Data centers will still need substations, utility equipment and electrical isolation. Solid-state systems introduce their own problems, including semiconductor cost, thermal management, reliability and the burden of proving that thousands of active components can survive infrastructure duty.

That is why the next year matters more than the announcement. Enphase plans full-system demonstrations and customer pilots, while its co-founder Raghu Belur is presenting the concept at the AI Infrastructure Summit under a telling title: “Solid-State Transformers: Speed to Power.”

Speed has become the commodity. A GPU that arrives six months early has little value if the building cannot be energized.

The AI boom has repeatedly moved its bottleneck outward: from accelerators to memory, packaging, cooling, transformers, generation and grid access. Solid-state transformers represent a different response to that migration. Instead of waiting for one constrained component to become abundant, redesign the system so less of the bottleneck survives in its old form.

If the approach works, the most important feature will not be that the transformer became digital. It will be that a century-old piece of electrical infrastructure became modular enough to keep up with the machines on the other side of it.