The Solid State Transformer Has Been Hyped for Fifty Years; AI Data Centers Are Its First Real Buyer
The solid state transformer failed three times over — in the navy, in high-speed rail, in the smart grid — because all it offered was ‘smaller’. The copper problem and the 800-volt DC problem inside AI data centers are the first reasons it has ever had to exist.
The argument · tap a timestamp to hear it
Transformers are heavy because of grid frequency, not poor manufacturing
A transformer's size varies inversely with the frequency of the current running through it, and grid frequency is pinned at 50/60 Hz, which makes the equipment inherently bulky. A transformer is also a purely passive component: it cannot actively regulate or clean up defects in the waveform, so voltage sags and spikes pass straight through it. These two physical limits — size, and the absence of active control — happen to be exactly the two holes that power electronics can fill, and they are the starting point for every solid state transformer attempt since.
The solid state transformer's architecture was locked in back in 1968
A patent by GE engineer McMurray combined power semiconductors with a high-frequency transformer: first ‘chop up’ the low-frequency current to raise it to thousands or even tens of thousands of hertz, then change the voltage using a physically small transformer, then use a second set of semiconductors to bring the frequency back down to 50/60 Hz. That three-stage ‘step frequency up — transform — step frequency down’ pattern is the core architecture shared by nearly every solid state transformer design of the following six decades, and it can cut volume by 70%-80%.
Railways never wanted peak efficiency, only small enough and vibration-proof enough
The PETT that ABB built for trains reached 95%-96% efficiency at half the volume by 2012, and more than twenty prototype units ran on Alstom, Siemens and Bombardier rolling stock. It was never genuinely commercialized, because it cost more than 50% above a conventional traction transformer — and a conventional traction transformer was never chasing peak efficiency in the first place. It only had to be small enough and able to take the vibration. Hitting the technical spec is not the same as having a buyer.
Utilities simply did not buy the ‘energy router’ story
The FREEDM center proposed turning the solid state transformer into a multi-port ‘energy router’, so that connecting solar and electric vehicles would be as simple as plugging in a printer. That vision got ‘smart transformers’ onto MIT's annual list of breakthrough technologies and spawned startups including Varentec and Amantys, and the optimistic 2012 forecast had the market reaching $5 billion by 2020, growing 82% a year. But utilities, broadly, did not buy the ‘energy router’ story.
Copper and iron last fifty years; swapping in electronics does not pay
Haroon Inam said flatly on the SemiAnalysis podcast that spending several million dollars back then on an AC-to-AC solid state transformer was, in hindsight, ‘one of the dumbest decisions’ — an iron core with copper windings lasts forty or fifty years, so why replace it with equipment that is more fragile and more expensive. Some estimates put the cost at up to 5 times that of a conventional transformer, and a 2022 paper went further, arguing that for both solid state and hybrid transformers, capital cost has to come down 60% before the economics work.
— Haroon InamThe bottleneck in a 1-megawatt rack is copper, not efficiency
With AI rack power climbing to the 1-megawatt class, sticking with the 54-volt rack architecture could mean 200 kg of copper busbar per rack — expensive, and taking up space that generates no revenue. This is not an efficiency-optimization problem, it is a physical constraint: power equals voltage times current, and more current means thicker copper. The only way out is to lift the whole architecture to 800-volt DC, which is exactly the direction written into Nvidia's roadmap.
This round is not a vision deck; products are already headed into machine rooms
Once the 800VDC architecture opened up, entrants followed one after another: DG Matrix (over $60 million raised), Heron Power ($140 million Series B), and Singapore's Amperesand ($80 million, backed by Temasek and Walden Catalyst, which the host guesses is Lip-Bu Tan's fund). Taiwan's Delta Electronics is already discussing a product that converts medium-voltage grid power directly into 800-volt rack voltage, and in February 2026 mentioned installing it at one of Meituan's data center campuses in mainland China.
The gains are still only simulated; there is no measured data at all
The host himself points out that when a video opens with a question, the answer is usually ‘no’. So far the gains from solid state transformers in 800VDC data centers come mainly from simulation; there is no real operating data, and no way to know whether they beat a hybrid 800VDC design built out of conventional transformers. What sets this round apart from the smart grid and the trains is that the cost simplification from stripping out redundant rectifiers and UPS units is real, the copper problem is a hard forcing function, and the circle of builders and suppliers on the data center side is more concentrated — they can set their own standards and build fast in-house, without carrying the baggage of a fragmented utility industry.
In their own words · checked verbatim
How do you replace a technology that’s been battle tested for literally over a hundred years?
Power electronics coupled with high frequency transformers to shrink the whole device's footprint by a significant percentage - in some cases, 70-80% - while also providing some form of controllability.
This most advanced PETT boasted unprecedented efficiencies of 95-96%. It was half the size of a regular transformer - making it far easier to stick under the floor.
[Solid state transformers] would make connecting a solar panel or electric car to the grid as simple as connecting a digital camera or printer to a computer.
Alex Q Huang12:27
You're taking a hunk of iron and a hunk of copper... that's going to last 40, 50 years... why the hell would anybody in their right mind try to replace that with a bunch of electronics that are going to be more delicate?
Haroon Inam14:29
At 1 megawatt with the existing 54 volt rack architecture, you may need up to 200 kilograms of copper busbars - which is expensive and too heavy considering that this is space that is not generating revenue for the data center.
The rule goes that when someone leads off their video with a question, the answer is usually NO.
Figures
| Rated efficiency of a conventional transformer | 98%-99% | 2:08 |
| Year of McMurray's electronic transformer patent | 1968 | 3:11 |
| Volume reduction from a solid state transformer | 70%-80% | 3:11 |
| European railway AC frequency | 16.67 Hz | 4:12 |
| ABB PETT efficiency (2012) | 95%-96% | 9:23 |
| ABB PETT cost premium | more than 50% above a conventional transformer | 9:23 |
| 2012 market forecast (by 2020) | $5 billion, growing 82% a year | 13:29 |
| GridCo funding raised | $54 million (folded in 2018) | 14:29 |
| Estimated solid state transformer cost (vs. conventional) | about 5 times | 15:30 |
| Capital cost reduction required by the 2022 paper | 60% | 15:30 |
Glossary
- Solid State Transformer (SST)
- An approach that replaces the conventional low-frequency iron-core transformer with power semiconductors plus a high-frequency transformer.
- IGBT
- A type of power semiconductor switching device that became commercially available in the 1980s.
- Silicon Carbide (SiC)
- A wide-bandgap semiconductor that tolerates higher voltage and higher frequency than silicon; the key device for making solid state transformers practical.
- Galvanic isolation
- The property by which a transformer blocks any direct current path between circuits, preventing faults from spreading.
- 800VDC architecture
- The new power-supply voltage standard AI racks are adopting in order to save on copper busbar.
- PETT
- Power electronic traction transformer — ABB's solid state transformer variant, built for high-speed rail traction systems.
How to listen
Practitioners and investors following power electronics or data center power architecture, or anyone trying to understand why capital has suddenly placed a new bet on a very old technology.
The historical detours in the middle on the navy and on European railway frequencies can be skipped; they do not affect the key argument about AI data centers later on.