Why solid-state transformers became an AI infrastructure story
AI training loads swing fast enough to destabilise the grid. Iron-core transformers cannot regulate that swing. Solid-state transformers can, because they replace the windings with SiC and GaN transistors operating as a controllable converter.
The binding constraint between the grid and the data centre is no longer just substation capacity. It is whether the substation can absorb 1 GW load swings without back-propagating instability onto the wider grid.
The grid-to-DC step is the new bottleneck
Every gigawatt of AI capacity has to step down from transmission voltage (hundreds of kilovolts) to the 800 V or 400 V DC the racks consume. Traditionally that conversion happens in a string of passive transformers and switchgear: iron cores, copper windings, time-tested, slow, and absolutely passive.
The new problem is that AI loads are not behaving like older industrial loads. A 100,000-GPU training cluster can swing hundreds of megawatts up and down on minute timescales as the run cycles between compute-heavy and communication-heavy phases. Passive transformers carry that swing straight back into the grid.
"Power Plant No Blow Up" is a software kludge
The clearest evidence the problem is real is a PyTorch feature flag that was nicknamed Power Plant No Blow Up. When the flag is on, GPUs that have finished their math step are forced to keep drawing power at full rate by running junk calculations on disposed-of outputs. The point is to flatten the load curve so the upstream substation does not see a sudden drop.
That is software paying for a hardware problem. The chip is wasting energy. The data centre is paying for that energy. The grid operator is being kept happy. None of those is a stable equilibrium.
Solid-state transformers regulate load actively
A solid-state transformer replaces the iron-core windings with a circuit built from high-voltage transistors. The transistors switch the input AC at high frequency, transform the voltage through a much smaller magnetic component, and rebuild a clean output. Because the switching is active, the converter can be controlled.
With a solid-state transformer in the loop, the data centre operator can dampen its own load swings instead of asking PyTorch to. The grid operator sees a clean draw. The hyperscaler stops eating waste energy. The permit process stops getting blocked on grid-stability objections.
Inside the data centre versus outside the data centre
There are two places solid-state transformers can show up. Inside the data centre, they regulate the rack-side power architecture, which is where the NVIDIA reference design and OCP working groups have focused. Outside the data centre, they sit between the high-voltage transmission link and the campus, performing the grid-interface job that iron-core gear performs today.
Irrational Analysis argues the more interesting opportunity is outside the data centre, because that is where load regulation actually solves the grid operator's complaint. Inside the DC is a more crowded market with established players. Outside the DC is the structurally new role.
The semiconductor question: SiC and GaN
A solid-state transformer is a power transistor problem. Two wide-bandgap materials beat silicon at the relevant scale.
- Silicon carbide. Tolerates much higher voltage than silicon. Used in the 1 kV to 10 kV range that grid-side and rack-side conversion both need. Wolfspeed is the only US pure-play vendor, currently emerging from Chapter 11 with very low factory utilisation but a 10 kV power-transistor part nobody else has.
- Gallium nitride. Faster switching, lower voltage range. Used in board-level conversion, server power supplies, and similar tighter integrations. TI and Navitas are tied for the leader position per Irrational Analysis, Infineon a distant second, onsemi a question mark.
The assembler side
Power semiconductors are only half the supply story. Someone has to assemble the transistors into a box that handles thousands of amps at thousands of volts without exploding. That is its own engineering problem, with its own players.
- Delta Electronics, Vertiv, Eaton. The established power-conversion assemblers. They already do this kind of work for data centres at scale. Stocks have repriced accordingly.
- SolarEdge and Enphase. Originally solar micro-inverter vendors with the right protection-circuitry expertise. After the residential solar tax-credit downturn they are pivoting into solid-state transformers. Irrational Analysis describes this as a degen but credible 5x setup.
Strategic read
Solid-state transformers are not a 2026 ramp. They are a 2027-second-half story with the lead time of any new grid-side hardware. The interesting part here is the timing: the power-semi side has spare capacity today because the EV downturn left fabs underutilised, so the bottleneck will be assembly and qualification, not silicon.
The broader frame is that AI infrastructure is silently rewriting the rules of grid-side equipment. Iron-core transformers have been good enough for a century. They are not good enough for a load profile that swings half a gigawatt on a minute clock. The chip industry crossed into the power industry's territory, and the power industry is going to have to follow.