Why AI Data Center Power Should be Built Like AI Compute
August 10, 2026
By: Raghu Belur, Co-Founder and Chief Product Officer, Enphase Energy
For most of the industry’s history, a data center’s limit was compute. Today, the limit is power. We’ve started describing capacity in megawatts and gigawatts instead of in racks or FLOPS, and that tells you where the constraint now sits. The electrical path between the utility feed and the GPU was built for steadier, smaller loads, and AI has outgrown it faster than almost anyone planned for.
The numbers are easy to state and hard to engineer around. A top-end rack in mid-2026 draws a couple hundred kilowatts off a 50 VDC bus. The next generation, built around an 800 VDC rack bus, is expected to push past 600 kW. The one after that, due around 2028, is being designed for more than a megawatt in a single rack. NVIDIA’s roadmap and the Open Compute Project’s Mt. Diablo work point the same way: higher rack voltage, higher density, and a power chain that has to deliver it.
The most important decision in this transition is architectural, and it gets skipped over.
The Two Things That Break First
Raise rack power far enough and two problems show up first. The first is space. Front-end AC-to-DC conversion used to live inside the compute rack and take up less than a tenth of it. As density climbs, that hardware grows until it needs its own sidecar rack beside the compute. At a megawatt, the sidecar fills with power supplies, battery backup units, and capacitor backup units. You end up spending the most expensive real estate in the building, the white space next to the GPUs, on power conversion instead of compute.
The second is speed. AI loads are not steady. During training, thousands of GPUs move in lockstep, and the load can swing between ten percent and full power several times a second. A power system that reacts slowly papers over those swings with local energy storage, which is why the capacitor and battery units sit in the sidecar in the first place. The slower the system, the more buffering you bolt on.
The Solid-state Transformer
The solid-state transformer (SST) is the piece of this that interests me most. The idea is more than a century old, but it only becomes practical with modern semiconductors. Instead of changing voltage with a passive core of iron and copper at grid frequency, an SST uses semiconductor switches to do it actively, at a synthesized frequency hundreds to thousands of times higher. Higher frequency means smaller magnetics. Active control means you can regulate the output, produce DC directly, and collapse several conversion stages into fewer.
For an AI data center, that’s the payoff: convert medium-voltage AC straight to regulated DC in a single stage, hold it stable under a load that won’t sit still, and give back the white space the old conversion chain was eating. Done right, an SST also lets you move bulk energy storage out of the building, into a battery energy storage system (BESS) parked in the black space, instead of scattering it across every rack.
That’s conditional, and the condition is the whole game: you can only pull the local storage out if the transformer rides through the load swings on its own. Response time is the hinge everything hangs on.
How Many Modules?
No single semiconductor switch can stand off tens of thousands of volts, so every SST is modular by necessity. You build the high-voltage device from many smaller power modules, wired in series on the medium-voltage side and in parallel on the output. That much is settled physics.
What isn’t settled is how many modules you use, and how big each one is. That’s the real fork.
The common instinct is to reach for the biggest switch available, usually high-voltage silicon carbide, and build the transformer from a few large blocks. Fewer parts and a simpler bill of materials; easy to reason about. The trouble is that large, high-voltage modules switch at low frequencies, which narrows control bandwidth and slows the system. A slow system takes several AC line cycles to react to a load change, so you are back to needing super-capacitor buffering on the DC bus to cover the gap. The architecture forces the very batteries you were trying to remove.
We made the opposite bet: build the transformer from a large number of small, lower-power modules. Each one uses smaller inductances switched at much higher frequency, which gives a wide control loop and a response well under a millisecond, roughly a thousand times faster than the conventional approach. When the system reacts that fast, the local buffering it needs drops by about the same factor. That’s what lets the storage move out to a remote BESS and gives the white space back to compute.
I keep coming back to one analogy, because it actually shaped the design. AI compute is not built around a single enormous processor. It is built from many smaller processors working together, coordinated in software. The power for that compute should be built the same way: many small, intelligent, semiconductor-defined modules acting as one. We didn’t invent it for AI; we’ve been building distributed power conversion for twenty years, first in solar and then in batteries. It’s the same principle here.
Why GaN, and Why Single Stage
Small modules only work if the switches inside them are good. Wide-bandgap semiconductors, silicon carbide and gallium nitride (GaN), have displaced the slow silicon thyristors and insulated-gate bipolar transistors (IGBTs) that older SST applications like rail traction still run on. For low-voltage, high-frequency modules, GaN is the right tool. A GaN high electron mobility transistor (HEMT) switches fast and clean, and the newer bi-directional switch, a single GaN device that replaces a back-to-back pair of one-directional switches, carries a cost advantage of up to four times and makes soft-switching AC converters practical.
GaN also makes a single-stage topology workable, and single-stage is where the efficiency and the low electromagnetic interference (EMI) signature come from. A clean, low-EMI module has a quiet benefit that is easy to miss: you can put it in a low-cost plastic enclosure, which makes it far simpler to grade the electrical field across the boundary between the medium-voltage and low-voltage sides of the module. The semiconductor choice and the mechanical design are connected all the way down.
The Hard Part is Control
Single-stage converters are efficient, but they are hard to control. This is why we’ve designed our own control application-specific integrated circuits (ASICs) since the company started, rather than buying generic parts. The current generation, which we call Kestrel, is a custom 22-nanometer chip built to run a single-stage converter, with the security and functional-safety features and the multi-core software that always-on infrastructure needs.
An SST adds a problem a single microinverter never had: the controller has to work across an insulation boundary of tens of thousands of volts. Kestrel handles it with a tandem function. Two chips, one on each side of that gap, talk over a fiber-optic link and behave as one controller. The power module crosses the medium-to-low-voltage boundary at only two points, the fiber link and the high-frequency transformer itself. Keep those crossings few and well defined and the insulation design becomes tractable. That transformer, with its custom cores and windings, is the other piece we build in-house, because it must meet the insulation and lifetime requirements the rest of the module depends on.
Reliability Comes from the Same Choice
Telecom has a number for availability that data centers adopted: five nines, 99.999 percent, about five minutes of downtime a year. You get there with parallel redundancy and parts you can swap without taking the system down. A high module count helps. Build in roughly ten percent module-level redundancy and the system runs through individual failures, and because the modules are small, a failed one can be hot-swapped, increasingly by robots, which is what you want for gear sitting on a 34.5 kV line.
Distributed energy resources already provide advanced grid functions under IEEE 1547 and UL 1741, so a transformer built on that lineage starts out compliant with what regulators will likely require of grid-connected SSTs.
The Choice the Industry is Making Now
The voltage and density jumps of the next few years will lock in power architectures that the industry lives with for a decade. The temptation is to treat the transformer as one big monolithic box and size it up. For a load this dynamic, that’s the wrong reflex. What decides whether you can drop the local batteries, hit five nines, and build at volume is module count, and the physics points to many small modules over a few large ones. Build the power the way the compute is built.