01Product 02Specs 03Who it's for 04How to buy Request specs

01 PRODUCT

CS1

150 KW CLASS · PUMPED-LOOP MODULE

One standard thermal module with a part number. Sized to a full 150 kW heat load, anchored on a single NVIDIA GB300 NVL72-class rack — 72 GPUs, 36 Grace CPUs, 9 switch trays, 117 cold-plate interfaces, ~140 kW of rack heat within the 150 kW rating.

CS1

Standard size — 150 kW class.

CS1-2

Size variant — different heat load / radiator area.

CS1 Rev B

Design revision — same size, updated design.

/ 01 THE THERMAL CHAIN

Six links. One clean interface.

Heat leaves your silicon and never comes back. Our scope starts at the lid of your package and ends at deep space — everything in between is one module, one spec, one supplier.

TIM2 interface layer

CUSTOMER LID → OUR PLATE

Hands heat from the customer's GB300 lid into our cold plate. Our spec begins at the lid — the die-to-lid TIM inside the package stays the customer's.

Cold plate OUR EDGE

METAL MICROCHANNEL · 117 INTERFACES

A metal microchannel plate on every GPU, CPU, and switch tray — our main differentiator, where thermal resistance is won. Baseline today; an aggressive direct-die / liquid-metal option follows.

Pumped loop

PGW · ~2 KG/S · 20 K RISE

A closed, warm propylene-glycol-water loop carries the full load at ~2 kg/s across a 20 K rise. Warm on purpose: rejection scales with T⁴, so every kelvin buys radiator area.

Redundant pump

N+1 · HOT-SWAPPABLE · BLDC · ~1–1.5 KW

The only moving part in the system and the #1 failure mode — so it gets the reliability budget: N+1 redundant, hot-swappable brushless pumps.

Radiator

TUBE-AND-FIN · ~110 M² · ~345 K

Metal tube-and-fin panels of flight-heritage class, radiating from both faces at ~345 K, flown knife-edge to the sun so the panels see cold sky, not sunlight.

Structure

BOOMS · WHIPPLE SHIELD · BRACKETS

Deploy booms, integrated Whipple micrometeoroid shielding, and mounting brackets — scope shared with the bus and settled per customer.

/ 02 CONNECTION POINTS

One core. Two mounts.

The same pumped-loop core connects to your vehicle two ways. Pick the path that fits your architecture; the fluid interface is identical.

Direct GPU mount

DIRECT-ATTACH · HIGHEST PERFORMANCE

Cold plates bolt straight onto the GPU packs; the loop runs directly to the pump module and out to the wing. Shortest thermal path, lowest resistance, tightest integration.

Solar-wing back mount

RADIATOR ON WING STRUCTURE · EDGE-ON

Radiator panels ride the back of the solar-wing structure — cells face the sun, radiator faces cold sky, panels stay edge-on. Saves dedicated boom structure where the wing already exists.

/ 03 ANCHOR LOAD

Sized to the rack that matters.

CS1 is engineered around the NVIDIA GB300 NVL72-class rack — the densest compute block headed to orbit. If the module holds the line there, everything below it is margin.

See full specifications

GB300 NVL72-CLASS RACK — HEAT BUDGET

72 × GPU @ ~1.4 kW~101 kW

36 × Grace CPUincl.

9 × switch trayincl.

Cold-plate interfaces117

Rack total~140 kW

CS1 rating150 kW

/ 04 WHERE WE WIN

Heritage panels weren't built for 100 kW compute. CS1 was.

Cold-plate performance

Optimized chip-side first, where incumbent heat-pipe and OHP panels never had to compete.

Reliability

One moving part, made redundant and hot-swappable. Everything else is metal and fluid.

System mass

Judged at the system level — plates, loop, structure, shield — not per-panel brochure numbers.

Cost & speed

Proven metal construction, designed for volume manufacture — cheap and fast to build.

Get the detailed CS1 data sheet.

Request specs Submit LOI / PO intent

CHILLER SPACE