Integrated power for the next generation of data centers.
Coordinate conversion, storage, distribution and controls around critical-load requirements — from the grid interface to the critical load.
Discuss Your Power Architecture →From the grid interface to the critical load.
Define the power path as one system. Conversion, storage, distribution and controls are specified against the same operating envelope — not assembled from separate packages.
Illustrative architecture. The actual configuration depends on voltage class, isolation requirements and the target market’s certification path. Scope boundaries — LumiVolt supply, partner supply and third-party interfaces — are defined in the project documentation.
Planning for 800 VDC? Start with the system architecture.
800 VDC is an option, subject to interface, grounding and protection assessment. Evaluate conversion, energy buffering and controls against your load requirements before the bus voltage is fixed.
Each project states which stage applies to which function — conversion, buffering, protection and recovery. No function is presented as validated before the test evidence exists.
The bottleneck is conversion, not generation.
A conventional feed passes through a line-frequency transformer, an AC/DC stage, a DC/AC stage, and a distribution panel — each with losses, footprint, and its own failure modes. An SST replaces most of that chain with a single controllable power-electronic stage.
Smaller, lighter footprint
No line-frequency transformer core or oil tank. Meaningful where floor area and structural loading are the real constraint.
Precise control
Voltage, frequency, and power flow controlled electronically — not fixed by a transformer turns ratio.
Bidirectional by nature
Power flows both ways without a separate converter stage — relevant for storage integration and V2G-style applications.
Integrated LV distribution
The LV output stage is part of the same device — one enclosure instead of transformer plus switchgear cabinets.
DC-native where useful
A DC link at the SST output suits storage, EV fast charging, and increasingly DC-powered IT loads.
Grid support functions
Reactive power compensation, voltage regulation, and — in suitable configurations — grid-forming capability.
The rack went from 8 kW to 1 MW in six years.
Nothing upstream was redesigned for that. The power train is now the constraint on how much compute a building can hold.
| Year | GPU generation | Power per rack | Power per cabinet |
|---|---|---|---|
| 2024 | Blackwell | ~130–250 kW | ~40–100 kW |
| 2025 | Blackwell Ultra | ~250–500 kW | ~100–350 kW |
| 2028 | Rubin Ultra | ~900–1000+ kW | ~350–500 kW |
Ranges cover air- and liquid-cooled configurations. Rack and cabinet are counted separately in the source roadmap — confirm both definitions with the vendor before committing a design. These are the roadmaps our customers are designing to, not our forecast.
Four things break at megawatt density.
Each one has a workaround at 100 kW and no workaround at 1 MW.
Floor area collapses
At 8 kW a rack needs half a square metre of power equipment. At 500 kW the power room grows faster than the compute hall — the building becomes a substation with servers in it.
Copper stops making sense
Low-voltage distribution at 800 A per rack means busbar measured in tonnes. The copper alone can outweigh the electronics it feeds, and it all has to be supported, cooled and paid for.
Conversion stages stack up
MV transformer, LV switchgear, UPS, PSU — every stage costs one to two points of efficiency and a cabinet of floor space. At 10 MW that is hundreds of kW of heat nobody planned for.
The load stops being predictable
AI training loads swing on a millisecond scale between idle and full draw. A grid connection designed around a smooth commercial load profile will read this as a fault, not as demand.
Raising the distribution voltage is the only move that removes copper, floor area and conversion stages at the same time.
Applications we serve around the SST.
AI & Hyperscale Data Centres
High-density racks need high-density power delivery. The SST shortens the conversion chain and reduces the electrical footprint per megawatt.
EV Charging Hubs
Direct MV connection with an integrated DC output — fewer conversion stages between the grid and the vehicle.
Critical & Industrial Facilities
Sites where power quality, redundancy, and physical footprint all matter at once.
Commercial record and development status — stated plainly
Our SST platform is not a prototype and not yet a volume product. Two ratings are quotable: 13.8 kV / 3.5 MW (Rev. A7) and 10 kV / 2.5 MW. Three 10 kV / 2.42 MVA / 800 VDC units have been in service at a data-centre campus since September 2023, four 13.8 kV units are in delivery on a 2026 export order, and a 750 V series is grid-connected at a park site. Certification is still open, and the Quality page states exactly which standards are pending. If your project requires a type-tested unit today, tell us and we will say whether we fit before a proposal is written.
Why data centre power architecture is changing
Data centre power design has been stable for decades. Power arrives at medium voltage, a line-frequency transformer steps it down, an uninterruptible power supply conditions it, and it is distributed at low voltage to the racks. That architecture worked because the loads it served were predictable and the constraint was cost per megawatt.
That is no longer the situation. Rack density has increased by an order of magnitude. Loads are increasingly DC-native. Sites are constrained by available floor area and by grid connection capacity rather than by energy supply. And the electrical infrastructure now represents a significant share of the cost of a megawatt of IT capacity.
The change is being driven from the load side rather than the supply side. For as long as loads were AC and predictable, the conventional chain was an adequate answer to a solved problem. Once loads became dense and DC-native, the chain stopped being the cheapest way to answer the question, and alternative architectures started to be worth the engineering effort.
This page sets out why the conventional chain is becoming a constraint, and where solid-state transformer architecture changes the calculation.
Read the detailed engineering discussion
Where the conventional chain binds
A line-frequency transformer is a large, heavy, oil-filled device. For a given megawatt capacity it occupies space that could otherwise hold IT equipment — and in a dense facility, floor area is the scarce resource.
Power passes through AC/DC and DC/AC stages that each introduce losses. At data centre scale, a small efficiency improvement compounds into a large operating cost over the life of the facility.
A transformer's output voltage is fixed by its turns ratio. It cannot regulate voltage dynamically, cannot control power flow direction, and cannot provide reactive power support without additional equipment.
Racks, storage, and fast charging increasingly operate natively on DC. The conventional chain converts AC to DC and back, when a DC bus would avoid those conversions entirely.
What SST architecture changes
A solid-state transformer replaces the line-frequency transformer, the separate power conversion stage, and the LV distribution stage with fewer, smaller, electronically controlled stages.
Voltage, frequency, and power flow are controlled electronically rather than fixed by physical construction. This enables functions that would otherwise require separate equipment.
Power can flow in either direction without an additional converter, which matters both for storage integration and for any application where the site both consumes and exports.
Storage, fast charging, and DC-powered IT loads can connect to a DC link at the output stage, avoiding a conversion on each side.
Two ratings are available and can be quoted, and units are in service — three 10 kV class AIDC units in service since 2023, plus a four-unit export order in delivery. Certification is not finished, and we publish that item by item rather than leaving you to find out. Alongside the product we offer architecture review, reference design, and a joint validation plan for the functions that still need evidence.
Where this applies
Where rack density is the binding constraint and electrical footprint per megawatt directly determines how much compute fits in the building.
Sites connecting directly at medium voltage with high aggregate charging demand, where avoiding conversion stages improves both efficiency and equipment count.
Sites already operating DC bus infrastructure where an AC-based conversion chain is an avoidable inefficiency.
Technical read: four parameters that define an SST design
A typical solid-state transformer rectifies medium-voltage AC onto a DC link, transfers across an isolating medium-frequency transformer, and then either inverts back to low-frequency AC or presents a DC output directly. Each stage can be built on silicon-carbide devices, which permit higher switching frequency, which in turn shrinks the isolating transformer dramatically.
The size of a transformer core falls roughly in proportion to the frequency it operates at. Raising the link from fifty hertz to tens of kilohertz is what allows a transformer that once filled a room to fit inside a cabinet — and that change in volume is the whole argument for the architecture, before any control benefit is counted.
Because the DC link carries stored energy, a solid-state transformer can hold up its output through a supply disturbance of a defined duration. How long it holds is set by the link capacitance and the load, and it has to be stated as a number in the specification rather than implied as a general guarantee.
Replacing several passive devices with one active device concentrates failure modes. If the facility requires N plus one, that redundancy has to come from parallel units or from a bypass path designed in from the start; it cannot be added afterwards as an option. Reliability is argued at system level, not at device level.
Integration considerations
Power-electronic conversion behaves differently from electromechanical transformation under fault conditions. Protection strategy and ride-through behaviour must be designed deliberately, not inherited from conventional practice.
A single device replacing several must be assessed for failure modes at system level. Where redundancy is required, the architecture must provide it explicitly.
Power-electronic equipment at medium voltage follows a different certification path from conventional switchgear. This affects the timeline and should be planned early.
Removing the transformer reduces footprint but concentrates losses in a smaller volume. Cooling design becomes more, not less, critical.