Hook: The industry keeps talking about “the 800 VDC SST” as if one magic box will feed tomorrow’s AI halls. On September 22, 2026, ABB answered with a different shape: Infinitus—an integrated source-to-rack DC portfolio that splits the problem into five building blocks, not a single converter SKU (Converge Digest summary of ABB launch, Sept 22, 2026; cross-check ABB primary materials before final cite).
The timing is not marketing fluff. ABB positions today’s AI racks near ~200 kW, with next silicon aiming at 1 MW per rack and beyond. At that density, every extra AC/DC/AC hop steals watts, floor space, and cooling budget from compute (StorageReview / ABB positioning).
Key takeaway: Infinitus is less “buy our SST” and more “design the whole DC chain”—MV in → conversion → distribution → protection → cooling—so a megawatt rack does not inherit a 480 V AC facility’s conversion tax.
What happened
ABB launched Infinitus as a direct-current power portfolio for AI infrastructure. Coverage describes five principal blocks:
- Medium-voltage powertrains — bring power from the grid or on-site generation into the site.
- DC sources and power quality — convert AC to DC outside the white space and stabilize the bus.
- 800 VDC distribution — move high power density through denser conductors in less footprint.
- DC power protection — overcurrent and fault isolation for people, servers, and hardware on a DC bus.
- Cooling optimization — DC-connected variable-speed drives and motors across grey and white space.
At the center of the conversion story sits a solid-state transformer intended to step medium-voltage AC toward 800 VDC, collapsing stages that a classical AC UPS + PDU path still keeps. ABB frames the portfolio for both DC-native halls and hybrid AC/DC retrofits—important, because almost nobody rebuilds a campus overnight (same Converge Digest / StorageReview coverage).
Roadmap claims from launch coverage: major components into production within roughly 12 months; first complete DC-native facilities using the full portfolio in about two to three years. The launch also extends ABB’s NVIDIA 800 VDC collaboration announced in 2025 for gigawatt-scale AI sites (vendor timeline—treat as roadmap, not a purchase order).
Engineering mechanism: why “one SST” is the wrong mental model
A solid-state transformer can erase several conversion stages. It does not erase the rest of facility physics:
- 1) Distribution geometry
At hundreds of kilowatts to a megawatt per rack, ampacity, busbar layout, and voltage drop dominate cable trays and white-space volume. 800 VDC is a density play: same power, lower current—if the bus and connectors are designed for DC. - 2) Fault energy and selectivity
DC arcs do not have a natural current zero every half-cycle. Selectivity between feeder and rack breakers becomes a microsecond-class problem. An SST upstream without coordinated DC protection is an incomplete story—exactly the gap our earlier SSCB pieces hammered. - 3) Cooling is part of the electrical budget
Fans, chillers, and CRAH motors are not “HVAC elsewhere.” When cooling rides on DC-connected drives, every percent of cooling efficiency frees IT watts. Infinitus explicitly puts cooling optimization inside the same portfolio—signaling that power and thermal are one system at 1 MW/rack.
Efficiency claim (directional):
- ABB / BCG analysis cited in launch coverage estimates >5% energy-efficiency gains from DC distribution versus conventional AC paths. For a 500 MW facility, 5% is on the order of 25 MW that could otherwise feed IT load—or be avoided as generation/cooling plant. These are estimates, not fleet-measured PUE deltas (Converge Digest).
That framing matters for engineers: Infinitus is arguing that the bottleneck is system architecture—how many conversions, how faults clear, how cooling is powered—not only semiconductor switch loss inside one chassis.
The misconception worth killing
Misconception: “Specify an 800 VDC SST and the AI campus is ready.”
Reality check: An SST is one conversion layer. Without DC-rated distribution, protection coordination, and a cooling path that matches the electrical topology, you still own AC-era losses, fault risk, and grey-space sprawl—just with a fancier front end.
Hybrid AC/DC support in Infinitus is the pragmatic tell. Operators will stage DC into new halls or new rows while legacy UPS rooms keep running. The engineering question becomes where the AC/DC boundary sits, and who owns protection at that boundary—not whether DC wins in a slide deck.
Limits / 待核:
- “>5% efficiency” and “25 MW on a 500 MW site” are ABB/BCG estimates—verify against ABB primary launch docs and independent measurements before treating as measured results.
- “Main components in 12 months / full DC-native facilities in 2–3 years” is a roadmap, not currently shipping end-to-end plant.
- Rack power figures (~200 kW today → 1 MW+) are industry positioning; actual SKUs vary by GPU generation and liquid-cooling maturity.
- DC protection performance (clearing time, selectivity, arc behavior) needs product-level data—portfolio branding alone does not specify breaker behavior.
Takeaway judgment
If you design or buy AI power infrastructure in late 2026, stop evaluating “SST vs transformer” in isolation. Ask:
- Where does medium voltage become 800 VDC—and how many stages die in that hop?
- Who clears a rack-level DC fault in microseconds without collapsing the bus?
- Is cooling on the same electrical story, or still an AC afterthought stealing PUE?
ABB’s Infinitus is a credible vendor signal that the industry is productizing source-to-rack DC as a layered stack, aligned with NVIDIA’s 800 VDC direction. Watch three checkpoints before you redesign a campus around it:
- Named SKUs and lead times for each of the five blocks—not only the SST slide.
- Independent efficiency and fault-clearing tests on shipping hardware.
- A clear hybrid migration path that does not strand existing UPS investment.
800 VDC will not arrive as one box. It arrives as a coordinated chain. Infinitus is ABB naming that chain—and that naming is the useful engineering news.
Related: @TheEngineeringCore-v · @Vkinng · vkinngworld.blogspot.com · prior TEC pieces on SSCB protection and onsemi EPP packaging for adjacent layers of the same stack.
Sources: Converge Digest (2026-09-22 ABB Infinitus); StorageReview Infinitus coverage; OEM Update / ElectronicsForU secondary summaries. Efficiency and timeline figures marked as vendor/BCG estimates—spot-check ABB primary materials before absolute citation.
—— The Engineering Core
