The rise of grid power electronics with Drew Baglino

Start with why power semiconductors got interesting. The familiar Moore’s Law story is about shrinking logic transistors; power transistors improved over the same five decades along different axes, including how much voltage they can block, current density, thermal conductivity and, above all here, switching speed. Speed matters because of how electricity behaves on a wire: everything connected to a contiguous circuit affects everything else instantaneously unless something in the middle decouples the flow. Power electronics are that something, and modern silicon carbide and gallium nitride devices start and stop current millions of times per second. Applications arrived in the order the devices allowed, from variable frequency drives on industrial motors around 1980, to direct-current links letting grid regions at different frequencies trade power, to consumer switching power supplies, to the parts that made solar inverters and electric vehicle drive inverters possible in the early 1990s.

Now contrast the grid, where none of that happened. At the branch points the equipment is still mechanical switches and passive transformers. The switches actuate in hundreds of milliseconds, can operate maybe once every couple of minutes, and are not built for more than a couple of thousand operations in their whole life. The transformers are fixed-ratio voltage dividers with no control over power flow; power simply takes the path of least resistance. Baglino says this was true in 1970 and is broadly still true in the 2020s, while the inverter in a battery or an EV charger controls voltage and current hundreds of thousands of times a second. What changed is device voltage. Silicon carbide was a 600-volt or 1.2-kilovolt technology ten years ago and now reaches 4.6 kilovolts, close enough to US distribution voltages that you no longer need an impractical stack of devices in series to work at grid voltage.

The commercial opening comes from a different direction: the legacy transformer market cannot keep up. Baglino declines to rank the causes and simply lists them. Demand is up broadly, from data centers interconnecting at transmission, from new generation, and from electric vehicles and home electrification, and part of it is replacement of 1970s equipment rather than growth at all. Regulatory uncertainty made it worse: the Department of Energy floated changing transformer efficiency rules, which would touch the grain-oriented electrical steel that is the largest material input by mass, so nobody expanded steel supply while the question was open. Tariff volatility makes it hard to know which country to build a factory in. Kann adds a cause Baglino does not, which is that incumbent manufacturers have lived through boom-bust cycles and expand by 2x rather than 5x, by which time the data center forecast has doubled again. He defends that as reasonable rather than negligent, since every manufacturer would rather sell into an undersupplied market. Baglino has a stake in one of his own listed causes: part of the hesitation he describes is people betting that solid state transformers will displace the steel, which is, as he says, one of the reasons he started Heron.

Kann then raises the objection that matters most in this industry. Better does not necessarily win, so what is the net outcome a buyer actually pays for? Baglino leads with reliability rather than capability. Inverters, not modules, are the largest source of underperformance on utility-scale solar, with central inverter availability averaging 97.5% to 98%. The transformers fail too, at roughly 1% to 1.4% a year, because desert plants run them at nameplate for eight or nine hours a day, which is not what they were designed for, and they carry essentially no instrumentation, so owners send people out to check oil and bushings. Against that, the solid state approach moves the 60-hertz transformer to 100 kilohertz, making the magnetics 50 to 100 times more power dense and letting the unit be built in 100-to-200-kilowatt modules rather than megawatt blocks, so a fault costs 100 kilowatts instead of a megawatt, with no crane and no months-long wait for a replacement. Components disappear with it: the legacy transformer, low-voltage breakers and fuses, power-factor-correcting capacitors, and some protection complexity, since a power electronics front end that faults pushes only slightly above rated current instead of hard-faulting and catching fire.

The data center case is the same argument with more to delete. Racks are still fed with alternating current the way they were in the 1990s, which is fine at 10 kilowatts a rack and increasingly absurd at 100 kilowatts or a megawatt. At that density the right answer starts to look like EV charging or a battery plant: distribute direct current at higher voltage and raise the rack backplane from 48 volts, itself a telecom leftover, to 800 volts or more. One device can then go from 34 kilovolts straight to 800 volts with no intermediate transformers and no gray space full of uninterruptible power supplies and distribution panels. The GPUs are still where the money is, but the scarce inputs are certified electricians and copper, and both fall sharply when you stop distributing power at low voltage. The long-run grid version is that one of these can cost about what an oil-filled transformer costs while also doing the jobs of the components that surround it: overcurrent protection and fault isolation, tap-changer voltage correction, three-phase balancing, synchronous-condenser-like inertia, capacitor banks for power factor. Those functions raise utilization, and since a rate case divides total utility cost by kilowatt-hours served, higher utilization of existing poles and wires is the lever on affordability.