Skip to content
Kernwall
Hardware · for your IT and facilities team

One signed OS, on servers you can buy.

Kernwall OS runs on OEM servers from a planned compatibility list, from a desk box to your rack. Kernwall builds one box itself: the fanless Edge for your line. Size yours from your users and models.

Line-up

Six tiers, one OS image.

Every tier boots the same signed image and writes the same evidence for your auditor. Only the model, the headcount and the room change.

  • Own brand

    Kernwall Edge

    Visual inspection, speech, 8–30B models; a 120B MoE at FP4 for 1–3 users

    Hardware basis
    Jetson AGX/IGX Thor T5000, 128 GB, fanless IP65/67, TPM 2.0, self-encrypting disk
    Power · site
    40–130 W · Factory line, vehicle, forward site
  • OEM server

    Kernwall Desk

    1–5 users, pilots, a RAG demo, an evaluator lab

    Hardware basis
    DGX Spark 128 GB, or a Strix Halo 128 GB box
    Power · site
    ~150–250 W · Office desk
  • OEM server

    Kernwall Workgroup

    Up to 50 users on a 120B MoE, 70B at FP8, plus embeddings and speech

    Hardware basis
    2–4× RTX PRO 6000 in a tower or 2U, 512 GB–1 TB DDR5
    Power · site
    1.5–3 kW · Comms room on a 230 V / 16 A circuit
  • OEM server

    Kernwall Department

    100–300 users; a 600B MoE for a small team on the Station

    Hardware basis
    8× RTX PRO 6000 in 4U, or a DGX Station GB300
    Power · site
    5–7 kW (4U) · ~1.6 kW (Station) · Data closet (4U) · office (Station)
  • OEM server

    Kernwall Rack

    600B+ MoE for hundreds of users, fine-tuning, on-site teacher models

    Hardware basis
    HGX B300 8-GPU nodes
    Power · site
    ~14 kW per node; 20–40 kW racks · Data centre row
  • EU
    OEM server

    Kernwall Sovereign-EU

    Vision and small-LLM inference without US GPUs

    Hardware basis
    Dell or Supermicro server with Axelera Europa (629 TOPS, ~45 W)
    Power · site
    Low · Server room

Sizing

Your users and model decide the box.

Pick your users, model and context. The calculator names the tier, its GPUs and its power, and shows the working behind them.

Named users
Model
Context per request
EstimateMeasured
Build
2× RTX PRO 6000
Power · site
1.5–3 kW, 230 V / 16 A
sizing · working
50 named × 10% at once = 5 requests in flight
× 2 headroom = 10 concurrent requests to plan for
measured at 32K: 1 card holds 5, 2 cards hold 8
2 × 1-card copies hold 10 ≥ 10
What the calculator assumes

Millstone AI measured gpt-oss-120b on one and two RTX PRO 6000 cards under vLLM. At 32K context, one card held 5 concurrent requests and two held 8, with the first token under 8 seconds and every user above 15 tokens a second. The calculator plans for 10% of your named users at once, then 2× that for headroom. Past two cards it counts side-by-side copies of those set-ups: our assumption, not a measurement.

Compatibility

Three ways to get a box.

Buy one pre-built, bring your own, or take the Edge box. The compatibility list is planned, not published; its first entries are one Dell and one Supermicro Workgroup server, due in the build's second month.

Pre-built appliance
An OEM server from the compatibility list. An EU integrator ships it with the Kernwall OS image and your accreditation pack.
Bring your own server
A server you already own, if it matches a compatibility-list entry. Kernwall enrols your Secure Boot keys and seals the disk to its TPM, exactly as on a pre-built appliance.
Kernwall Edge
The one box Kernwall builds itself. Fanless, sealed to IP65/67, on a small bill of materials, and drawn from the edge systems the founder built for factories.

Brands we plan to testDell · HPE · Supermicro · Lenovo · NVIDIA-Certified Systems

Repairs, planned

Your drives stay with you. A unit goes back for repair without its disks, and Kernwall provisions the replacement from the signed image, so your plant data never leaves site in a return. The policy is planned, not yet published. It answers the VS-NfD repair clause and your fab's media rules alike.