Investment | Price benchmarks | 03
Rack server
The first tier where the facility starts to matter: an 8-GPU training node draws and dissipates as much as a whole rack did a few years ago. We design the node together with the power, cooling and network path that support it.
Typical configuration
CPU / services; storage; 8× GPU training node
Investment range
€10–60k | storage: €38–331k | 8× GPU: €440–520k
What this tier involves
- Typical role
- Virtualisation, databases, storage, AI training
- Power & thermal
- 3–12 kW per node; 8× GPU nodes frequently require liquid cooling
- Network
- Up to 8× 400G for training nodes
- Fee
- Basic €300–1,500 | Platinum €1–35k | Supreme €150–600/server/month
Indicative figures, August 2026, excluding VAT. High volatility on GPUs and memory — firm quotes are valid for 30–60 days. The final price is set at scoping.
What matters at this scale
At node level the server is sized as a system: processor, network, storage and cooling are chosen together, according to the node's role.
- Host per role
- Virtualisation calls for core count and memory; training calls for PCIe lanes and one network interface per accelerator
- Liquid cooling
- Densely accelerated nodes fall within the ASHRAE liquid-cooling classes; cold plates and loops are provided for in the design
- Local storage
- Local NVMe for scratch and checkpoint, separate from shared storage, with wear monitoring
- Firmware baseline
- Firmware and microcode held at a common reference version across all identical nodes
- Measurement
- Power draw and temperatures are metered per node, as the basis for later energy reporting