A short circuit on an 800 V DC bus can climb by roughly 800 amps every microsecond. A mechanical breaker needs 10–20 milliseconds to clear it. Do the math and you see why every 800 VDC announcement this month quietly comes with a second product: the solid-state circuit breaker.
October has been the month the AI data center power chain went DC in public. At the OCP Global Summit (Oct. 12–15), DG Matrix showed its Interport 360 Flex skid that takes 34.5 kVAC in and delivers native 800 VDC to the white space. On Oct. 21, Vantage secured more than 1 GW of Heron Power's solid-state transformer capacity. ABB's new Infinitus DC portfolio lists "DC Power Protection" as one of its five building blocks and leans on what ABB calls the first fully IEC-certified solid-state circuit breaker.
The transformer gets the headlines. The breaker is the part that decides whether 800 VDC is safe to put next to a megawatt rack. Here is the engineering, with numbers.
1. DC has no current zero, and the fault rises at V/L
An AC breaker cheats: the current crosses zero 100 or 120 times a second, and the arc simply goes out at the next zero. DC never crosses zero. To stop it mechanically, the arc itself has to build a voltage higher than the bus, which at 800 V means a big, hot, slow arc chute. Infineon and SolarEdge put it bluntly: without a natural current zero, mechanical interruption "faces challenges with arcing" and is comparatively slow.
The second problem is speed. Right after a bolted fault, only the loop inductance limits the current:
di/dt = Vbus / Lloop → 800 V / 1 µH = 800 A/µs
The 1 µH loop is an illustrative assumption for a short busway run; real values depend on layout. If the loop resistance is 10 mΩ, the current heads toward I = V/R = 80 kA with a time constant τ = L/R = 100 µs. In other words, a fault reaches tens of kiloamps long before a mechanical contact has even started moving.
2. Let-through energy: why microseconds beat milliseconds by ~10⁶
What damages cables, busbars and server power shelves is the let-through energy, I²t. Use the same illustrative circuit and compare two breakers.
Solid-state breaker. Say it trips at 500 A (4× a 125 A rating) and needs 2 µs to detect and turn off its switches. The current hits 500 A after 0.625 µs, then keeps rising for 2 µs:
Ipeak = 500 A + 800 A/µs × 2 µs = 2,100 A
I²t (linear ramp) = Ipeak² × t / 3 = 2,100² × 2.625 µs / 3 ≈ 3.9 A²s
Add the decay phase from point 3 (about 5 µs, another ≈7.7 A²s) and the total is roughly 12 A²s.
Mechanical breaker. With 10 ms clearing (the low end of the 10–20 ms that Infineon cites for electromechanical breakers), the current settles near 80 kA for most of that time:
I²t ≈ 80,000² × 0.010 s ≈ 6.4 × 10⁷ A²s
That upper bound ignores arc-voltage current limiting and source impedance, so the real gap is smaller. But even a 100× correction leaves the solid-state breaker tens of thousands of times gentler. This is the point of NVIDIA's 800 VDC white paper: interrupt "before the fault reaches its maximum value," and you cut both peak current and arc-flash incident energy.
3. The hidden job: dumping the inductor's energy
Turning off a transistor does not make the current vanish. The loop inductance still holds E = ½LI², and it will push the voltage as high as it needs to keep current flowing. Every solid-state breaker therefore has a clamp, usually a metal-oxide varistor (MOV) or TVS diode, in parallel with the switches.
tfall = L × Ipeak / (Vclamp − Vbus) = 1 µH × 2,100 A / (1,200 − 800) V ≈ 5.3 µs
Eclamp = ½ L Ipeak² × Vclamp / (Vclamp − Vbus) = 2.2 J × 3 ≈ 6.6 J
Notice the factor of 3. While the current decays, the source is still pushing 800 V into the clamp, so the MOV absorbs three times the stored energy, not just ½LI². Two design levers follow directly. A higher clamp voltage clears faster and absorbs less energy, but the switches must then block that clamp voltage with margin, so a higher clamp means higher-voltage, costlier devices. And because Ipeak enters squared, shaving 1 µs off detection time cuts clamp energy far more than you'd guess.
Key point
An SSCB is a race against di/dt. Every microsecond of detection delay adds about 800 A of peak current on an 800 V, 1 µH loop, and clamp energy grows with the square of that peak. Fast sensing matters as much as the power switch.
4. The price: conduction loss, and why big SSCBs need liquid cooling
A mechanical contact is a piece of metal pressed on metal. A solid-state breaker carries the full load current through semiconductors all day, and it has to block in both directions when batteries and regenerative loads sit on the bus. That costs I²R:
Pcond = I² × Ron,total
125 A × 125 A × 4 mΩ ≈ 63 W (on 800 V × 125 A = 100 kW → 0.06 %)
1,250 A × 1,250 A × 0.4 mΩ ≈ 625 W (on 1 MW → 0.06 %)
The on-resistance values are illustrative, assuming ten times more silicon in parallel for the bigger breaker. The percentage stays small; the absolute heat does not. 63 W can be air-cooled inside a power shelf. 625 W in a breaker-sized box usually cannot. That maps neatly onto NVIDIA's two initial SSCB targets: 125 A air-cooled units for 800 V-to-54 V power shelves, and 1,250 A units, "likely utilizing liquid-cooled designs," for native 800 VDC rack interfaces. It is also why the device choice matters. Infineon and SolarEdge picked SiC JFETs for their low on-resistance per area and short-circuit robustness.
To verify
- All circuit values in points 1–4 (1 µH, 10 mΩ, 500 A trip, 2 µs delay, 1,200 V clamp, 4 mΩ / 0.4 mΩ on-resistance) are illustrative assumptions, not vendor specs.
- Ideal Power's B-TRAN 800 V / 63 A reference kit reportedly shows about 1.4 µs semiconductor interruption and about 15 µs total clearing; check the latest datasheet before quoting.
- ABB does not give a launch date or ratings for the Infinitus SSCB on its press page; first full DC-native Infinitus sites are expected "within two to three years" per ABB.
What to take away
- 800 VDC is a protection problem before it is an efficiency win. ABB and BCG estimate DC distribution gains of more than 5% (ABB), but none of it ships without breakers that beat di/dt.
- Judge an SSCB on three numbers: detection-plus-turn-off time (sets Ipeak), clamp energy rating (sets survivability on repeated faults), and on-resistance (sets the heat bill every hour it is closed).
- Watch the cooling spec. Above a few hundred amps, the breaker joins the liquid loop, which ties protection design to the facility's cooling design.
Want the upstream half of this story? Our video Efficiency vs. Power Factor: The Real Reason AI Data Centers Need SSTs walks through the solid-state transformer that feeds these 800 V buses.
Sources: DG Matrix (Oct. 13, 2026), Vantage / Heron (Oct. 21, 2026), ABB Infinitus, Infineon / SolarEdge (Sep. 9, 2026), NVIDIA 800 VDC white paper, Ideal Power B-TRAN RDK.
