Scannit
Bare-metal · NVIDIA

Shattering AI’s biggest bottleneck.

Bare-metal NVIDIA GPU clusters for AI teams, leased for 1 to 60 months. Short terms run on servers we already operate. Long terms are built to order.

123456789101112FABRICNODE-01 · 8× B200NODE-02 · 8× B300NVMe

GPU
8× NVIDIA B200
VRAM
1.4 TB HBM3e
CPU
2× Xeon Platinum 8570
Network
8× 400 Gb/s
Standard buildsRTX PRO 6000B200B300

Hardware

Standard builds.

Three builds to start from. Each one changes to your spec, and we source other NVIDIA GPUs on request.

Availability changes quickly, so we confirm it with the price in your quote.

Standard build 01

8 GPUs per node

NVIDIA RTX PRO 6000

For inference, fine-tuning and rendering.

GPUs
8x RTX PRO 6000 Blackwell
Total VRAM
768 GB GDDR7 ECC
GPU interconnect
PCIe Gen5
CPU and memory
2x Intel Xeon 6767P, 1 TB DDR5 ECC
Networking
RoCE or InfiniBand, per build

Standard build 02

8 GPUs per node

NVIDIA B200

Blackwell system for large-model training and inference.

GPUs
8x NVIDIA B200 SXM
Total VRAM
1.4 TB HBM3e
GPU interconnect
NVLink 5, 1.8 TB/s per GPU
CPU and memory
2x Intel Xeon Platinum 8570, 2 TB DDR5-6400
Networking
8x 400 Gb/s ConnectX-7

Standard build 03

8 GPUs per node

NVIDIA B300

Blackwell Ultra system with more memory per GPU, for larger models and longer context.

GPUs
8x NVIDIA B300 SXM
Total VRAM
2.1 TB HBM3e
GPU interconnect
NVLink 5, 1.8 TB/s per GPU
CPU and memory
2x Intel Xeon 6776P, 3 TB DDR5-6400
Networking
Configured per build

More GPUs on request

Need a different setup? Customize it

Custom network topologies, InfiniBand configurations and storage expansions on request.

Lease terms

Bare-metal leases from 1 to 60 months.

Every lease runs from one node to a full cluster.

One quote

Price, availability and terms for your build, in writing.

Short or long terms

We scope each cluster to your workload. New hardware is built to order on terms of 24 to 60 months. Shorter terms, from 1 month, run on hardware we already operate.

Managed on request

Bare metal by default. Managed services when you ask for them.

At a glance

Lease term1 to 60 months. New builds 24 to 60.
RatesQuoted per request
AccessBare metal, no hypervisor. Managed services on request.
Cluster sizeFrom 1 node; 8 GPUs per node in the standard builds
LocationUnited States

Process

From request to handover

  1. 01

    Request

    Tell us the hardware, the cluster size, the lease term and when you need it.

  2. 02

    Scope and quote

    We scope the build with you and send price, availability and terms.

  3. 03

    Source and install

    We source what the build needs and install it in a SOC 2 audited data center in the US.

  4. 04

    Test and hand over

    We burn in the cluster, run the acceptance tests we agreed with you and share the results. Then we hand it over as bare metal.

Reserve compute

Facilities

Bare metal in SOC 2 audited US data centers.

We source the hardware, install it and run it. You get one contact for the hardware and the site.

Bare-metal compute

Bare-metal access to NVIDIA GPUs, with no virtualization layer. Built for data-intensive AI training and high-throughput inference.

  • Local NVMe storage

    Up to 30.7 TB of NVMe per node, for staging datasets next to the GPUs.

  • No virtualization overhead

    Direct hardware control, so the GPU does the work you scheduled.

Enterprise facilities

Enterprise colocation data centers in the United States.

  • Sites

    US data centers, chosen per build and confirmed in your quote.

  • Clean handover

    Every node is wiped and reimaged before a lease starts.

  • Operator audits

    The data center operator is audited to SOC 2 Type II every year and certified to ISO 27001.

  • Uptime target set in your contract

    Agreed per lease, to fit your workload.

Reserve

Reserve compute

Tell us what you need. We come back with price, availability and terms.

  • Bare metal, leased for 1 to 60 months.
  • Custom network topologies, InfiniBand configurations and storage expansions on request.
  • We use these details to answer your request.

8 GPUs per node

1 node = 8 × B200

Storage needs, networking topologies, location or deployment timelines.

This is a request, not an order. We come back with price, availability and terms. Privacy policy

By sending this request, you agree to our Terms of Service.

Own GPU servers? We find them work, and you keep 75%.

See how it works