Every megawatt that reaches a GPU passes through a transformer, and for a century that transformer has worked the same way: a steel core, copper windings, and 50 or 60 hertz doing the work. It is cheap, it is enormously reliable, and almost nothing about it has changed.
AI density is the first load in a long time that makes the alternative worth discussing. A solid-state transformer replaces the magnetics with power electronics, and in doing so changes what the building's electrical spine can do. It is not yet the default, and for most deployments it should not be. But the reasons it is being specified at all are worth understanding before someone sells you one.
What a solid-state transformer actually is
A conventional transformer converts voltage by magnetic coupling at line frequency. The physics sets the size: at 50 hertz, the core has to be large, and a large core is heavy.
A solid-state transformer takes a different route. It rectifies the incoming medium-voltage AC, switches it at high frequency through a much smaller transformer, then converts it back to whatever the load wants. Because the magnetics now operate in the kilohertz range rather than at 50 hertz, the core shrinks dramatically for the same power.
That size reduction is the headline, but it is not the interesting part. The interesting part is that every one of those conversion stages is controllable.
What becomes possible when the transformer is active
A conventional transformer is a passive device. Whatever the grid does arrives at the load, attenuated and transformed but not managed. A solid-state transformer sits in the path as an active element, which brings several capabilities that currently require separate equipment:
- Native DC output. The architecture already produces a DC link internally, so DC distribution comes out of it rather than requiring another conversion stage.
- Power factor correction at the device, without separate correction equipment downstream.
- Harmonic isolation. Load-side distortion does not propagate upstream in the way it does through magnetics.
- Voltage regulation that responds in milliseconds rather than through tap changers.
- Fault current limiting, because the electronics can stop conducting rather than relying on downstream protection to clear.
Several of those map directly onto problems covered elsewhere in our material on power factor and power quality. In a conventional design each one is a separate box with its own footprint, maintenance interval and failure mode.
Why AI loads make the case stronger
Three things about GPU density change the arithmetic.
Density itself is the obvious one. When a single rack draws what an entire row used to, floor area given to electrical rooms becomes floor area not given to revenue. A smaller transformer footprint has direct commercial value in a way it did not when racks drew 5 kW.
The second is DC. The move toward 800V DC distribution to megawatt racks exists to cut conversion losses and copper mass between the utility connection and the chip. A solid-state transformer fits that architecture natively, because the DC link is already inside it. Bolting a rectifier onto a conventional transformer achieves the same thing with one more stage and one more set of losses.
The third is volatility. AI training produces load swings that conventional electrical design did not anticipate, and an active device that can regulate through them is worth more than a passive one that simply passes them along.
The honest case against
It would be easy to stop there, and the result would be marketing rather than analysis.
Cost is the first obstacle and it is not close. Conventional transformers are a mature, commoditized product with a century of manufacturing scale behind them. Solid-state units are neither, and the premium is substantial enough that it has to be justified by something other than efficiency.
Reliability is the second. A conventional transformer has no moving parts and no semiconductors, and service lives measured in decades are ordinary. A solid-state transformer contains a large number of switching devices, control electronics and cooling, each of which is a failure mode that the magnetics version simply does not have. The field history is thin by comparison, which matters more in infrastructure that is expected to outlive several hardware refreshes.
Serviceability follows from that. A utility or contractor anywhere in the world can service a conventional transformer. A solid-state unit is far closer to a specialist product, and the supporting ecosystem, spares and trained technicians are correspondingly narrower. For a site in a region where lead times already shape project schedules, that is a real exposure rather than a footnote.
And the efficiency argument is more subtle than it looks. Modern conventional transformers run extremely efficiently at rated load. Solid-state units can beat them across a wider load range, particularly at partial load, but the headline comparison at full load is not the walkover it is sometimes presented as.
The efficiency question, done properly
Efficiency comparisons between the two technologies are frequently made badly, usually by comparing a solid-state unit's best operating point against a conventional transformer's nameplate figure.
A fairer comparison asks three things. What is the efficiency across the load range the site will actually run at, which for a facility that ramps is rarely a single point. What downstream equipment does each option require, and what does that equipment cost in losses of its own. And what is the cooling load created by the losses, because waste heat in an electrical room is a cost that appears on the mechanical side rather than the electrical one.
A conventional transformer running at 30 percent load is meaningfully less efficient than the same unit at 80 percent. A solid-state unit holds its curve flatter. Whether that matters depends entirely on how the site loads its electrical infrastructure, and sites that build ahead of demand spend years at partial load.
None of this makes the comparison automatic. It makes it a modeling exercise against a specific load profile rather than a datasheet comparison, and the answer will differ between two sites with identical equipment and different ramp plans.
How this interacts with redundancy
Redundancy design assumes certain failure behavior, and solid-state units do not fail the way magnetics do.
A conventional transformer degrades slowly and predictably. Insulation ages, oil degrades, and condition monitoring gives years of warning. Catastrophic failure is rare and usually preceded by signs that a maintenance program is designed to catch.
Power electronics fail differently. Semiconductors tend to fail suddenly rather than gradually, and the useful warning signs are different ones: thermal cycling history, capacitor condition, cooling system performance. A maintenance program written for magnetics will not catch them, which means adopting the technology also means adopting a different monitoring regime.
The counterweight is that solid-state units can often be built with internal redundancy at module level, so a single device failure degrades capacity rather than removing it. That is genuinely attractive, and it is also a claim worth testing against the specific product rather than accepting as a property of the category.
Where it fits today
The deployments that make sense now share characteristics, and most data centers do not have them.
They are space constrained in a way that gives the footprint reduction genuine value. They are building DC distribution anyway, so the native DC link removes a stage rather than adding a novelty. They have power quality requirements that would otherwise require a stack of separate equipment. And they have the engineering depth in house to own a less mature technology, rather than depending on a contractor who has never seen one.
If a site has none of those, a conventional transformer with conventional downstream correction remains the right answer, and will be for some time. The relevant question when planning capacity for AI workloads is not whether solid-state is better in the abstract. It is whether the specific constraints on this specific site are ones it actually relieves.
What to ask if someone proposes one
- What is the installed base of this exact model, and how many unit-years of field operation sit behind it?
- Who services it, how far away are they, and what is the response commitment?
- What spares are held in region, and what is the lead time on the switching assemblies?
- What is the efficiency curve across the load range, not just at nameplate?
- What is the failure behavior, and does the site stay up through it?
- What separate equipment does it genuinely eliminate, costed honestly?
The last one usually decides it. A solid-state transformer that replaces a transformer plus correction equipment plus a rectifier stage is a different proposition from one that replaces only the transformer.
The summary
Solid-state transformers are a real technology aimed at real constraints, and AI density is the first load that makes those constraints common enough to matter. They are also expensive, less proven, and harder to service than the thing they replace.
Treat them as a tool for specific situations rather than a general upgrade. Where DC distribution, space pressure and power quality requirements all land on the same site, the case can be genuinely strong. Where they do not, the conventional answer is still the correct one, and there is no shame in that.