Unitsolver Compute Shell + Cell
From steel coil to running compute. Time is the most expensive asset you own.
One owner of the whole AI Factory: the building and the Vera Rubin compute cells made on our own robot line in the Unitsolver factory, with the balance-of-system from leading manufacturers integrated into the same model. Designed end to end by our AI-coordinated engines, from a single cell to 66 MW in one reference hall. We are a start-up building an ecosystem that fits the NVIDIA DSX Reference Design for Vera Rubin, with our formats being finalised to NVIDIA’s requirements. Coming to market in early 2028.
A Coil to compute
The building phase of an AI factory is usually a site project, assembled piece by piece. We make that phase ours: the Shell and Cell structures come from our own production, and the BoS equipment from leading manufacturers is integrated into them. One steel coil feeds one robot line; the same welded faces become the hall and the compute cells. Compute Shell and Compute Cell are produced in the Unitsolver factory; each cell is fitted out and tested there, transported to the Unitsolver AI factory site and lifted into place. The film illustrates what happens along this production chain.
Unitsolver · building the ecosystem since 2023
An AI factory moves at the pace of its site: permit, footings, power and liquid connections, the building itself. We work the way that bottleneck demands, industrialising the physical build so the facility keeps pace with the compute it hosts.
Since 2023 we have built an ecosystem that aims for a technological lead. Its foundation, our research into physical construction, was completed in autumn 2025.
Now we are validating the integration together with leading companies in the field, to the standard of excellence set by the NVIDIA DSX Reference Design.
GPUs depreciate from the day they ship. Every month the building is late is a month of compute nobody sells. We take the two slowest things off the critical path — engineering and the site-built shell — and build the compute cells in the factory while the hall goes up.
Empty field to running compute in about 10 months (target): shell, power, cooling and cells included. Factory-built AI infrastructure still needs a prepared, connected site. Unitsolver manufactures that site’s building and cells, so the compute arrives into a facility that is already finished, cooled and connected.
Each month earlier is worth up to $0.8 M per MW of IT in GPU-hours to the operator, at today’s rental rates. calc
Design drops from ~10 % of capex to about a fifth of that — the building is generated, not redrawn. target
No cast floor slab. Footing lines on site; everything above them goes in as factory steel.
B The product family
Shell, Cell and every BoS pod are welded from the same patented steel cassettes, on the same robotic line, under the same structural engineering.
The models below are semantic references from the core of the concept: our computational BIM model illustrates the design, not every detail as it will be built. Our construction technology is patent pending worldwide, and we are protecting many new, unique solutions across the whole production chain as we bring Europe’s most ecological CE-marked AI factory to market.
The AI factory building itself. Roof, envelope and floor system from factory cassettes, bolted dry onto footing lines — sized from the start for the cells, busways, liquid mains and pods it will carry. Grows by the cell, not by the construction project.
A closed, laser-welded compute enclosure, fitted out and acceptance-tested in the factory: one row of twelve Vera Rubin NVL72 racks and a network rack, a 600 mm rear service aisle for manifolds and quick-disconnects, and one connection unit for liquid, power and data. Every cell is cabled, piped and functionally tested in the factory (FAT) before shipment: the factory-first principle, extended from the rack to the building. The compute trays go in on site, once the tested cell is lifted into place. Every cell is designed together with its OpenUSD digital twin, following NVIDIA’s SimReady conventions.
The 76–81 % of a liquid-cooled facility's cost that sits in MEP gets the same treatment: equipment from leading manufacturers, skid-built on our cassettes, tested in the factory, set outside the hall and plugged in.
C Our thesis
One project.
An AI factory brings together compute, cooling, power, the site and the building. We take ownership of the whole as one product: pick the size you need today, and grow on the same frame, the same interfaces and the same warranty boundary.
One Vera Rubin cell: twelve NVL72 racks. Sovereign pilot, research cluster, edge node.
3–4 cells. The sweet spot: CDU redundancy fully used, fixed site cost under ~5 % of capex.
A generated hall: 18 positions in three bands. Phase 1 installs 6 cells, 13.2 MW; the other 12 positions are serviced and waiting.
The 100 × 60 m reference hall with all 30 positions filled: up to 66 MW of Unitsolver enclosure capacity calc. Shells repeat on one site as one project.
The critical points of an AI factory are its interfaces: where a liquid joint, a busway or a cell meets the hall. We design both sides of every interface in one model, from the steel to the integrated BoS equipment, and test them in the factory before they meet on site.
What an AI factory has to meet — and where it sits in our design. Every figure traces to a calculation, a test or a vendor document.
| Requirement | How Shell + Cell meets it | Basis |
|---|---|---|
| Compute density | Twelve Vera Rubin NVL72 racks in one 41.6 m² cell, up to 2.2 MW per cell. | calc |
| Liquid cooling | Liquid at the rack rear on a drip-trayed DN150 loop, serviced in place from a 600 mm rear aisle. Zero crossings with power or data. | calc |
| Water | Closed liquid loop to dry coolers on structural AP200 decks: the heat leaves without evaporating water, and the warm return is ready for heat recovery. Vera Rubin runs on warm-water liquid cooling (NVIDIA: coolant up to 45 °C), which suits closed-loop dry coolers in a Nordic climate. | design rule |
| Power | Redundant A+B busway above the rack row; path to 800 VDC distribution. | calc |
| Structure | AP200 floor cassettes spanning 3.8 m between the full-height long walls, 1.5 mm where the cell bands carry the racks; every span checked by the FEM engine. | EN 1993-1-3 |
| Fire | Aspirating detection + IG-541 inert gas. No halon, HFC or PFAS agents. Leak-tight welded envelope, tested at the factory. | EN 15004 |
| Schedule | ≈10 months field to compute; cells FAT-tested in parallel with the Shell. | target |
| Grid | Phased in blocks that fit distribution-voltage connections, grown as capacity arrives. | design rule |
| Lifecycle | 50-year frame, 4-year payload. Cells lift out, get re-fitted for the next rack generation, go back in. | design rule |
| Carbon | SSAB Zero™ steel frame, −70 % embodied carbon vs. concrete. | EPD |
| Digital twin | Cell and Shell are modelled as OpenUSD concept assets following SimReady conventions, intended for NVIDIA DSX Sim workflows; checked against the public AI Factory validator rules (own check, not NVIDIA qualification); cell-to-shell connections matched in the model. | own check |
D Predictability
Unitsolver is built on one idea: a unit can be modulated to meet any technical, geometric or human need — if the reasoning behind it is deterministic. The same input always produces the same building, down to the last profile. That is what makes cost, schedule and quality predictable before a single coil is unrolled.
Same input.
Same building.
Every time.
The engine turns a brief and a drawing into a verified structure, the drawings, the cutting files and the carbon figure in one run. Change the rack generation or the layout, and the whole chain regenerates — nothing is redrawn by hand, so nothing is re-interpreted between model, analysis and machine.
A fully robotised Prima Power line: coil, punch and shear, robotic bending, assembly and laser welding. Every station takes its parameters from USBSos — no paper work orders.
USBS factory · production lineWhat a whole facility costs the line. One reference facility is a 6 000 m² Shell — 6 000 m² of AP200 floor, 6 150 m² of YP200 roof and 2 560 m² of US100 envelope — plus 30 Vera Rubin Compute Cells at 172.6 m² of cassette each. That is about 19 900 m² of steel cassette, roughly 465 t of steel. Against a line output of 250 000 m² of cassette a year it is about 8 % — one line builds roughly twelve complete AI factories a year, and capacity for the cells themselves is set by fit-out bays, not by the line.
E Embodied carbon
Compute Shell and Cell are welded from SSAB Zero™ — steel made from recycled scrap with fossil-free electricity, carrying a verified EPD. The frame clears the carbon line on its own: no offsets, no certificates bought after the fact. The cooling adds no water use: dry coolers reject the heat without evaporation.
From 2029, embodied carbon enters EU data-center reporting. A frame that is light, bolted and relocatable is also the frame that scores.
Frame embodied carbon, A1–A3, indexed to a concrete frame = 100. Frame only; MEP and payload excluded. EPD
F Ecosystem
Tell us your site, your power and your first megawatt. We come back with a Shell, a cell count, a schedule, a carbon figure and a price for the whole facility: generated from your site, not estimated.