Transformers: Bottlenecks in Disguise

A 300 MVA power transformer — a hunk of grain-oriented steel and copper wire, doing essentially the same job it did fifty years ago — now takes up to four years to procure and costs two to three times what it did in 2019. Not because anyone invented a better transformer. Because the specialized steel inside it is made by five companies on Earth, and every one of them is booked out. The most unglamorous object in the entire electrical grid — a device so mature and so passive that nobody thought to build a startup around it — has become the tightest chokepoint in the AI buildout, and, not coincidentally, one of the more interesting venture capital land grabs of the past year.

The reason is straightforward physics colliding with an equally straightforward supply problem. A transformer converts voltage by induction through a magnetic core, and the size of that core scales with the inverse of frequency — run it at the grid's standard 50 or 60 hertz, and you need a genuinely large, heavy slab of steel to handle serious power. That's fine for a substation that doesn't move and has decades to amortize its footprint. It's a much worse fit for an AI data center that needs megawatt-scale power delivered fast, in a building that also has to leave room for GPUs and cooling. Layered on top of that physical constraint is a supply one: the same specialized electrical steel needed for a new transformer is also needed for the broader grid modernization and electrification buildout happening simultaneously, so demand for the raw material is spiking from every direction at once, while the number of companies capable of making it hasn't changed in decades.

The fix that's emerged is the solid-state transformer — not really a transformer in the classical sense at all, but a stack of power semiconductors doing digitally, at high switching frequency, what the magnetic core used to do passively. Swap iron and copper for silicon carbide and gallium nitride, and the same voltage conversion that required a room-sized object can happen in something closer to a large cabinet, on a timeline measured in months rather than years. It's the same transition that hit telephone switching and camera shutters a generation ago — a mechanical, physically-constrained technology getting quietly rebuilt out of semiconductors — except this time the payoff is measured in gigawatts rather than convenience.

What's notable is how fast capital has moved into this specific niche, and how blurry the line has already gotten between the venture-backed disruptors and the industrial incumbents they're supposedly disrupting. Three startups — Heron Power, DG Matrix, and Amperesand — raised roughly $280 million combined over about six months. Heron closed a $140 million Series B in February 2026, co-led by Andreessen Horowitz's American Dynamism fund and Breakthrough Energy Ventures, and is building a 40-gigawatt-per-year US manufacturing facility; its Heron Link device targets standard 34.5-kilovolt utility distribution voltage and outputs 600-volt DC aligned with the newest AI rack architectures, and NVIDIA has formally recognized it as a data center power partner. DG Matrix closed a $60 million Series A the same month, with its "Power Router" already in testing with Duke Energy and PowerSecure, a microgrid developer owned by utility Southern Company. Amperesand raised $80 million in November 2025 and is targeting 30 megawatts of commercial deployments this year.

But the startups aren't the only ones moving quickly, and the roster of companies now shipping SST products is a useful sign of how fast this niche has gone from research curiosity to commercial category. Enphase Energy, a name more familiar for rooftop solar microinverters than utility-scale power conversion, has launched its own SST offering aimed specifically at AI data centers — a notable pivot for a company built on distributed residential hardware, and a sign that the addressable market for solid-state power conversion has grown large enough to pull in adjacent players who didn't start in this business at all. SolarEdge has made a similar move. Neither company needed to invent the underlying semiconductor technology from scratch; both are applying power electronics expertise built for a different market onto the same fundamental problem the dedicated SST startups are solving, which says something about how portable this capability has become once the switching silicon itself is mature enough.

Here's the detail that complicates the tidy "startups vs. incumbents" story further: one of DG Matrix's earliest investors, and one of the utilities its rival technology is being piloted against, is ABB — a hundred-plus-year-old electrical equipment giant that took a strategic stake in DG Matrix's seed round and returned to participate in its Series A. ABB isn't watching this disruption from the sidelines. It's funding it directly, while presumably also building its own competing technology. And Eaton didn't wait to see whether the startups would eat its lunch — it acquired Resilient Power Systems in August 2025 for roughly $86 million and has since brought its medium-voltage solid-state transformer, the MVSST, to market as a named, shipping product line explicitly aimed at data centers and EV charging. The distinction between disruptor and incumbent, in this specific fight, has mostly already collapsed into a shared bet on the same underlying technology.

The honest complication is that this technology isn't fully proven yet, and the timelines matter. High-voltage silicon carbide devices — the class needed for the most ambitious grid-facing applications — have historically had processing yields as low as 50 to 60%, well below the maturity of lower-voltage parts. Long-term reliability questions, including degradation modes that only show up after sustained high-temperature operation, are genuinely still being tested in live conditions rather than settled in a lab. This isn't a technology where the engineering is finished and only the manufacturing scale-up remains; some of the hardest problems are still open, even as customers are already signing contracts for 2026 and 2027 delivery.

For investors, the useful way to read this isn't as a binary bet on which specific company's silicon wins. It's a bet on where the choke point actually sits, and for how long, and on which companies have the balance sheet or the manufacturing base to move fast on multiple fronts at once. Eaton is the clearest way to be exposed to this trade without picking a single unproven vendor — it's both directly commercializing SST technology through its own product line and, as an established manufacturer with existing customer relationships, positioned to acquire whichever startup's approach proves out first, the way it already did with Resilient Power. Enphase's entry from the distributed-solar side is worth watching for a different reason: it suggests the market is wide enough, and the underlying semiconductor building blocks standardized enough, that companies without a legacy transformer business can credibly compete for it, which should put pressure on pricing and timelines across the whole category. Infineon and STMicroelectronics sit a layer further down, supplying the silicon carbide and gallium nitride devices every one of these SST designs — startup, incumbent, or new entrant — ultimately depends on; their exposure to this trend doesn't require guessing which transformer architecture or which company wins, only that transformer demand keeps outstripping transformer supply, which every data point in this space currently confirms. Schneider Electric, already deep in the 800VDC reference-design ecosystem for AI data centers, is a reasonable proxy for the same underlying thesis from the systems-integration side.

The more durable prediction isn't about which company's transformer design becomes the standard. It's that the four-year transformer queue, if it doesn't get solved, becomes a permanent tax on the entire AI and electrification buildout — a hard ceiling on how fast gigawatts of new demand can actually be connected, regardless of how much capital or GPU supply is available. The race to commercialize solid-state transformers isn't really a race between a handful of well-funded companies anymore. It's widened into a race against a supply chain constraint large enough to cap the AI infrastructure buildout's actual pace, with an increasingly crowded field of specialists, incumbents, and unlikely new entrants all betting they can out-manufacture the bottleneck before it becomes permanent.