AI compute your lab owns.
A research lab can own its AI compute outright: Wick Ultra puts 1,152 GB of GPU memory on four NVIDIA Rubin GPUs in the lab itself, from $656,000, and Wick for one researcher starts at $199,000.
Models and datasets stay on the lab's own hardware, and the lab's jobs run with no shared queue.
From $656,000Includes delivery, the outdoor unit and the first year of service.
Which Wick for which group
Every figure is per unit, or per node for Wick Cluster.
| System | Built for | GPU memory | 4-bit weights up to | Full fine-tuning up to | Price |
|---|---|---|---|---|---|
| Wick | One researcher | 288 GB | About 576 billion parameters | About 18 billion parameters | $199,000 |
| Wick Pro | A small model team | 576 GB | About 1.15 trillion parameters | About 36 billion parameters | $339,000 |
| Wick Ultra | A frontier model team | 1,152 GB | About 2.3 trillion parameters | About 72 billion parameters | $656,000 |
| Wick Cluster | A training floor | 2,304 GB | About 4.6 trillion parameters | About 144 billion parameters | $1,234,000 per node |
Why labs own compute
Shared university clusters suit many projects. A lab buys its own when a data agreement keeps a dataset in the building, or when a project needs large GPU memory on demand rather than in a queue.
In a 2026 survey of 1,500 enterprise architects and infrastructure leads, 66% had moved AI workloads from public cloud to private cloud or on premises in the past year, and 42% named data security, governance and compliance as the main reason.
In the lab
Wick, Wick Pro and Wick Ultra stand against an outside wall and send their heat to their own outdoor unit, fitted on delivery day. Wick Cluster joins a facility with three-phase power and outdoor units or building water. Wick reviews your room with you before you accept a price.
Every system has 10 GbE and fiber (4 × 800 Gb/s fiber on Wick Ultra) to reach the lab's network and storage, and holds up to 491.52 TB of protected data SSD you choose.
Buying for a lab
Sending a build is free. A 20% deposit holds it once you accept your written price, refundable until it ships, and the balance is due once it runs, by wire or ACH. First year of service included in every list price.
Questions
Can a lab train its own models on Wick?
Yes, within GPU memory. Full training holds about 16 bytes per parameter before activations: one Wick Ultra fully trains or fine-tunes models up to about 72 billion parameters, and one Cluster node up to about 144 billion.
Who looks after it?
Wick delivers it, connects power and network and switches it on. First year of service included in every list price, and no IT team is needed.
Can we run our own software stack?
Yes. Wick OS is included and optional. Install it on delivery day, or run your own operating system and tools on the bare hardware.
Sources
- Virtualization Review, reporting Cloudera and Wakefield Research, Cloudera survey: 66 percent repatriated AI workloads from public cloud, August 12, 2026
- San Diego Supercomputer Center, Hao AI Lab receives NVIDIA DGX B200, May 23, 2025
Wick figures are read from its published specifications and list prices. Checked October 27, 2026.