Seal a rigid vessel, fill it with carbon dioxide above its critical density, and put the processors inside. The heat they make raises the pressure. The pressure drives the circulation. There is no pump, no compressor, and no refrigeration plant inside the vessel — the only moving thing in there is the fluid, and what moves it is the computation itself. Heat still has to leave the building: outside the vessel that is a circulation pump and dry-cooler fans. What the sealed design deletes is the chiller, not the loop.
Every figure on this page is computed from the Span–Wagner equation of state for CO₂, or from closed-form geometry. Nothing here is illustrative.
Every cooling architecture in the data centre today is a race to keep up with the chips — the cooling sets the ceiling and the silicon runs into it. In the 1370 mm vessel the vessel removes heat 43 % faster than the silicon inside it can produce heat. The binding limit is the die, not the fluid.
Both are sealed compute elements about the size of a loaf of bread. That makes it the honest unit of comparison.
Smaller than a 5090, runs hotter than a 5090, and carries ten and a half times the memory. Fill the vessel's 2.42 m³ with 1,453 graphics cards as pure volume — no boards, no airflow, an impossible best case — and you get 45 TB at 836 kW. The vessel does 138 TB at 386 kW.
159 kW/m³ — twenty-one times the power density of an air-cooled rack, eight times that of liquid. Floor area hides the fact that a rack is mostly air.
If modules must touch as many neighbours as possible and still tile space with no gaps, the mathematics leaves you very little choice.
Twelve faces, twelve neighbours — the kissing number in three dimensions, and therefore the most any equal body can touch. It is the Voronoi cell of the face-centred cubic lattice, so identical modules stack with no wasted volume.
A hollow shell of the same shape holds the modules on its interior facets. Because liner and module are the same solid at two scales, every facet lies parallel to a module face — and seating only works at particular sizes, never in between.
Those facets carry the wiring. An assembly that would normally be many boards joined by a network sits on a single surface instead, so no signal crosses a network hop to reach another module.
There are exactly five convex solids that can do this — the parallelohedra Fedorov enumerated in 1885. The rhombic dodecahedron is the only one offering twelve neighbours and a single facet shape.
In a sealed rigid vessel you cannot choose pressure and temperature separately — they are tied together by the equation of state. Fix how much CO₂ goes in, and everything else about the machine is already decided.
Not a rendering. Change the bore, the pressure rating, the charge density or the junction limit, and the pack re-solves: module count, capacity, flux, wall thickness, network diameter. Peel the vessel inner to outer, or open a single module and look at the dies from inside.
Separate from the machine above. Send us your cards and power supplies; we bench-test everything, measure your actual stack, and fabricate a sealed vat engineered around it. Nothing is due up front — the invoice comes thirty days after your build ships, and hardware that fails validation goes home with no order written.