Skip to content
AccountMenu

Buyer guides.

Each guide answers one question buyers of on-premises AI hardware ask, with figures from Wick's published specifications and cited sources.

  1. Wick Ultra vs NVIDIA DGX Station: GPU memory and specs

    Wick Ultra holds 4.6 times the GPU memory of NVIDIA DGX Station GB300: 1,152 GB on four NVIDIA Rubin GPUs against 252 GB on one Blackwell Ultra GPU, with 200 petaflops of FP4 compute against its 20.

    Wick Ultra: 4.6 times DGX Station's GPU memory, from $656,000.

  2. How much GPU memory do open models need? A 2026 table

    An open model needs 0.56 GB of GPU memory per billion parameters at 4-bit (NVFP4, 4.5 bits per weight with its scales), one at 8-bit and two at 16-bit, for its weights alone: DeepSeek V4 Pro's 1.6 trillion parameters take 900 GB at 4-bit.

    Wick Ultra holds the 4-bit weights of 2.04 trillion parameters, from $656,000.

  3. Hardware to run a 1-trillion-parameter model on premises

    A 1-trillion-parameter open model such as Kimi K2.5 needs 577 GB of GPU memory for its weights at 4-bit and 1,026 GB at 8-bit, before context.

    Kimi K2.5 for 399 people at once on Wick Ultra, from $656,000.

  4. Hardware to run a 70B or 400B model on premises

    A 70-billion-parameter model needs 140 GB of GPU memory for its weights at 16-bit and a 405-billion-parameter model 810 GB, before context.

    Llama 3.3 70B at full 16-bit precision on one Wick, from $199,000.

  5. Running open models at full precision: BF16 memory needs

    Running an open model at full 16-bit precision takes 2 GB of GPU memory per billion parameters for the weights alone: Qwen3.5 397B needs 794 GB in BF16, which one Wick Ultra's 1,152 GB holds, and GLM-5 needs 1,488 GB, which one Wick Cluster node's 2,304 GB holds.

    Qwen3.5 397B in BF16 on one Wick Ultra, from $656,000.

  6. AI workstation vs AI server for a team: what fits where

    The largest AI workstations hold 384 GB of GPU memory across four NVIDIA RTX PRO 6000 cards, or 252 GB on the one GPU of DGX Station GB300, while an eight-GPU server such as NVIDIA DGX B200 holds 1,440 GB in a data center rack.

    Server-class memory without the rack: 1,152 GB in Wick Ultra, from $656,000.

  7. Private AI compute for a research lab

    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.

    Wick for one researcher from $199,000, Wick Ultra for the group from $656,000.

Questions about price, payment, the room or setup: Wick Help