Delta SOFC + SST Explained: On-Site DC Power for AI Data Centers

Bottom line: For AI data centers, the hard question has moved from "when do the GPUs arrive?" to "when do we get power?" This piece expands on this post and skips the SOFC basics. It covers three things instead: where the supply chain jams, how much a DC-coupled path actually saves, and where Delta's schedule really stands.

What happened

• At Energy Taiwan 2026, Delta showed an on-site SOFC power solution and its first physical SST, and announced a 550 kW SOFC order from Great Taipei Gas.
• The SOFC tech is licensed from Ceres Power (UK). Pilot production of about 5 MW starts at Tainan Plant 2 by end-2026, shipments begin in 2027, and mass production moves to Guanyin, Taoyuan in 2028.
• Delta's SST system converts medium-voltage AC directly to 800 V DC. Delta expects to ship about 100 units by year-end, with a bigger ramp in 2028.

Verification note: I couldn't confirm the post's "100 MW production line." Public reports describe a ~5 MW pilot line in Tainan and mass production in 2028. The figures "SST efficiency above 92%, 50% smaller footprint" and "2–3 overseas customers validating" are also not in Delta's public disclosures, so they're treated here as the post's illustrative figures.

Point 1: The bottleneck is time

The post cites 3–7 year grid queues, gas turbines sold out toward 2030, and roughly 3-year transformer waits. These figures are widely reported, though they vary a lot by region. The pattern is what matters: GPUs for a 100 MW-class site can land within months, while the interconnect takes years. Delta's hydrogen BU head Charles Tsai puts it as "you can buy chips, but not necessarily power." That's why US hyperscalers are moving to Bring Your Own Power.

SOFC scales in modules rather than waiting for one big turbine. And because there's no combustion flame, NOx is very low, which makes air permits easier to get.

Point 2: The DC-coupling efficiency math

A conventional chain runs MV AC → transformer → LV AC → UPS rectifier → inverter → rack PSU → chip DC, which is about 5–6 conversions. At 98% each, 0.98⁶ ≈ 88.6%. A native-DC SOFC through a SiC converter onto an 800 V bus, then a rack DC/DC, is about 2 stages: 0.98² ≈ 96%. On a 100 MW load, 7 points is about 7 MW, or roughly 60 GWh a year. This is a simplified estimate; real per-stage efficiencies will vary.

Point 3: About 40% less fuel

Delta quotes about 65% electrical efficiency for its SOFC, and up to 85% with heat recovery. A simple-cycle gas turbine runs about 35%. At 60% vs 36%, fuel per kWh is 0.36/0.60 = 0.6, which is about 40% less gas and matches the post. Combined-cycle plants reach 55–60%, but those are utility-scale plants, not something you put inside the fence.

Point 4: Why 120 kW racks force 800 V

Racks are going from tens of kW to 120 kW and beyond. At 54 V, the current at that power overwhelms busbars. At 800 V the current drops to about 1/15, and I²R losses fall with it. NVIDIA is pushing an 800 V HVDC architecture, and Ceres markets its new stack as "800V DC compatible," so SOFC + SST + 800 V lines up into a single chain.

Open questions

• SOFC volume: end-2026 is only a pilot. Real volume depends on the 2028 Guanyin plant.
• Stack life and degradation under 24/7 duty, and the cost of stack replacement, aren't public yet.
• SOFCs on natural gas still emit CO₂. The long-term path needs hydrogen or carbon capture.
• The 92% SST efficiency and 50% footprint figures need checking against Delta's datasheets once they're published.

Related videos

▶ Ditching the Grid! Delta's SOFC Secret & SST for Perfect AI Microgrids

▶ Efficiency vs. Power Factor: The Real Reason AI Data Centers Need SSTs

Will on-site generation become standard for AI campuses, or is it just a bridge until the grid catches up?

This one is getting lively discussion on the fan page. Come join in 👉 our Facebook page

Sources: Taipei Times, Oct 15 2026, DIGITIMES, Commercial Times, Ceres 2026 interim results | The Automation Core · Videos by @TheEngineeringCore-v