Think AI

A datacentre, condensed.

Five liquid-cooled nodes with ILM pre-installed, from a whisper-quiet, desk-side MicroNode to the eight-GPU RackNode. Datacentre-grade density, with no datacentre required.

Compare the family Talk to us
Think SuperNode photographed from a low angle against a dark background

The family at a glance.

Every tier runs the same ILM orchestration platform and the same self-contained liquid cooling philosophy. Choose the node that matches the scale, and grow through Constellation when the workload does. All compute figures are dense, not sparse.

VRAM figures are per-node aggregate pools.

Liquid cooled
by design.

Most machines spend their power moving air. think Node moves heat. One sealed enclosure, one liquid loop, one orchestration layer, engineered by the same team as a single system.

think Node cutaway
PSU CPU MEMORY, AIR COOLED CIRCULATION ACTIVE · SEALED LOOP TEMPERATURE PRESSURE FLOW ILM
01Accelerator bay

Eight accelerators. One enclosure.

Air-cooled machines are spaced apart so air can pass between them. Liquid does not need the gap. Cards sit shoulder to shoulder on a shared manifold, and density stops being an airflow problem.

Conventional 4 kW liquid-cooled datacentre plant room required
Outdoor dry cooler Chiller warm climates Facility loop valves, piping Pump skid N+1 5U server CDU 42U rack 42U rack + plant To scale To scale vs 42U rack think think think SuperNode 471 x 285 x 513 mm · 68.9 L · 3 kW 7.1x vs 5U + CDU
Component layout illustrative. Volumes measured across the true system, including every part of the cooling path.
True system volume 659 L 68.9 L
Cooling density 6.1 W/L 43.6 W/L

Measured across the true system volume: chassis, radiators, pumps and every part of the cooling path.

12.7x the thermal density
of a datacentre rack.

Measured across the true system volume: chassis, radiators, pumps, distribution, everything. This is the number that decides whether a machine needs a building around it.

Watts of cooling per litre of true system volume Higher is better
42U rack, air cooled10 kW · 1,199 L
8.3
42U rack, liquid cooled50 kW · with CDU and chiller
6.0
5U server plus CDU4 kW · true volume
6.1
NVL72 class rack100 kW · true volume
7.3
42U rack, high density20 kW · 1,199 L
16.7
5U server, air cooled2.5 kW · 83.2 L
30.0
think UltraNode6 kW · 149.6 L
40.1
think SuperNode3 kW · 68.9 L · 471 x 285 x 513 mm
43.6
think RackNode8.8 kW · 83.3 L
105.7
True system volume includes every part of the cooling system: coolant distribution units, chillers and piping where the design requires them. Rack figures from published vendor and ASHRAE reference data. think figures from customer proof of concept, CFD simulation and prototype testing.
105.7 W/L
Peak thermal density
RackNode, measured across the true system volume
15 to 20 C
Above ambient
Accelerator die temperature under sustained full load
Zero
Facility water
No chilled water loop, no CDU, no raised floor, no CRAC

From a desk
to a rack row.

The same thermal architecture, the same orchestration layer, scaled four ways. Nothing is bolted on later.

think MicroNode

The whole architecture, shrunk to a desk. Two accelerators on the loop, silent enough to sit beside you, and Constellation ready the moment it powers on.

MICRO
SUPER
ULTRA
RACK
Accelerators
2
Unified VRAM
192 GB
FP8 dense
2 PFLOPS
Cooling envelope
1 kW
Where it lives
Under a desk
All compute figures dense, not sparse. Every SKU ships with ILM pre-installed and joins a Constellation on day one.

No facility water

The entire thermal system lives inside the chassis. Nothing to plumb into the building.

No raised floor

Office, lab, or rack row. The machine does not care where you put it.

Serviceable under load

No-spill quick disconnects and active leak protection.

Four patents pending

Engineering worth protecting.

The family, in full.

Every configuration, one page. All compute figures are dense FP, never sparse. Every node ships with ILM pre-installed and a native Constellation interface.

MicroNode

The whole architecture, shrunk to a desk. Whisper-quiet, edge-ready, zero cloud dependency. SKU: MicroNode-S-A · v1.4

SoftwareILM orchestration platform, 1 year licence included
CPUAMD Ryzen 9950X, 16 cores / 32 threads
RAM256 GB (4 x 64 GB) U-DIMM DDR5
GPU2 x NVIDIA PRO 6000, 96 GB VRAM each; 192 GB aggregate GDDR7 ECC; 2 PFLOPS FP8 dense
Storage2 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write
Networking1 x Intel 5Gb Ethernet; Think AI Constellation single-port OSFP PCIe Gen5 100G NIC (transceiver not included)
Chassis and PSUCondensed compute node; Pro 2200W 80 PLUS Platinum
CoolingThink AI MicroNode custom liquid cooling system, whisper-quiet closed loop

SuperNode

Enterprise production workloads in one node. Replaces three to four air-cooled servers, sounds like none of them. SKU: SuperNode-S-A · v1.4

SoftwareILM orchestration platform, 1 year licence included
CPUAMD Threadripper 9985WX, 64 cores / 128 threads
RAM512 GB (4 x 128 GB) R-DIMM ECC DDR5
GPU4 x NVIDIA PRO 6000, 96 GB VRAM each; 384 GB aggregate GDDR7 ECC; 4 PFLOPS FP8 dense
Storage4 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write
Networking1 x Intel dual 10Gb Ethernet + IPMI; Think AI Constellation single-port OSFP NDR 400Gb/s PCIe Gen5 NIC
Chassis and PSUWorkstation and rackmount variants; Pro 3000W 80 PLUS Platinum
CoolingThink AI SuperNode custom liquid cooling system
Think SuperNode in three-quarter view with the side panel open, showing the liquid cooling loop
Three-quarter, side panel open 1 / 6

UltraNode

A rack of compute in one node. Seven accelerators held at peak clocks, continuously. SKU: UltraNode-S-A · v1.4

SoftwareILM orchestration platform, 1 year licence included
CPUAMD Threadripper PRO 9995WX
RAM1024 GB (8 x 128 GB) OC R-DIMM WRX90 6400 MHz
GPU7 x NVIDIA PRO 6000, 96 GB VRAM each; 672 GB aggregate GDDR7 ECC; 7 PFLOPS FP8 dense
Storage4 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write
Networking2 x Intel 10Gb Ethernet + BMC/IPMI; Think AI Constellation single-port OSFP NDR 400Gb/s PCIe Gen5 NIC
Chassis and PSUTower and rackmount variants; 2 x Pro 3000W 80 PLUS Platinum
CoolingThink AI UltraNode custom 6000W liquid cooling system

RackNode Shipping H2 2026

Rack-scale density behind dual EPYC processors.

Component RackNode
SoftwareILM orchestration platform, 1 year licence included
CPU2 x AMD EPYC 9575F (5th Gen Turin), dual socket
RAM1536 GB (12 x 128 GB) DDR5 RDIMM ECC
Accelerators8 x NVIDIA PRO 6000, 96 GB each
Total VRAM768 GB GDDR7
Compute8 PFLOPS FP8 dense
Storage4 x 8 TB PCIe Gen5 NVMe, up to 14.8 GB/s read, 13.4 GB/s write
Networking2 x Intel i350 1GbE + BMC/IPMI; 2 x Think AI Constellation single-port OSFP PCIe Gen5 400G NICs
Chassis and PSURackmount; 2+2 redundant 3200W CRPS (80+ Titanium)
CoolingThink AI RackNode custom liquid cooling system

All nodes: 3 year hardware warranty and 1 year software support as standard; Premium and Mission-Critical SLAs available. RackNode specifications derived from platform configuration and subject to final validation.

One vendor. One platform. No integration burden.

Every think AI node ships as a complete system: liquid-cooled compute paired with ILM, backed by a 3 year hardware warranty and tiered support.

ILM pre-installed

Every node ships with a 1 year ILM licence: turnkey, out-of-the-box orchestration for inference, fine-tuning, and training.

Constellation-ready

Native high-bandwidth fabric on every SKU. Start with one node; grow to a fleet with no re-architecture.

Warranty and support

3 year hardware warranty and 1 year software support as standard, with Premium and Mission-Critical tiers available.

Match a node to your workload.

Tell us your models and throughput targets. We will size the node, or the fleet, and show the arithmetic.

Contact Think AI