Ingot
Order now

IG-4Rack unit

Inference that stays inside the rack.

A rack unit that runs your models on your own metal. No egress, no per-token meter, and a front panel you can read from across the room. This one is live.

Explore specs Look inside

Turbine · primary

Thermal load

LOAD %

Warming rays

Standby. Element cold.

Tokens / sec
Core temp
°C
Draw
W
Queue
jobs
Built forOn-prem AI

Why a box

Some data never leaves.

Follow one request through the building. It goes to the unit, the answer comes back, and that is the whole trip. Anything that tries to go further stops at your network edge.

  1. Nothing egresses

    The prompt, the documents and the output stay inside your network. There is no request to review, because there is no request.

  2. Flat cost

    A unit costs what a unit costs. The bill does not move because somebody shipped a chatty retry loop on a Friday afternoon.

  3. It keeps working

    No deprecation notice, no silent model swap, no provider incident. The weights on the drive are the weights you evaluated.

Inside the box

Four parts, no mystery.

Point at a part to find it in the chassis. Everything that wears out is replaceable from the front panel or the top lid.

  1. 1Accelerators

    Four cards with 384 GB of memory between them. Enough to hold a large open-weight model whole, with room left for long contexts.

  2. 3Turbine row

    Five fans pull air from front to rear. At normal load you hear the room, not the box.

  3. 4Power

    Two redundant 1,600 W supplies. Lose one and the unit keeps serving while you swap it.

  4. 2Storage sleds

    Four hot-swap NVMe sleds, 30 TB in all, encrypted at rest. Pull the sleds and the weights leave with you.

IG-4 · lid removed · air enters at the front and leaves at the rear

Spec sheet

IG-4, by the numbers.

Copied off the rating plate on the back of the unit. Every reading on the panel stays inside these limits.

Ingot
IG-4Rack inference unit · Rev. C
Form factor
2U, 19-inch rackSliding rails included
Accelerators
4 × K-class, 96 GB each
Dimensions
438 × 88 × 610 mm
Throughput
Up to 1,840 tokens/sTurbo, mid-size model, batch of 8
Weight
24 kg
System memory
512 GB ECC DDR5
Power
2 × 1,600 W, redundant80 PLUS Titanium
Storage
4 × 7.68 TB NVMe, hot-swap
Draw
284 W nominal · 694 W turbo
Network
2 × 100 GbEPlus 1 × 1 GbE management port
Acoustics
38 dB(A) at 1 m, nominal
Environment
10 to 35 °C, front-to-rear airflow
Input100 to 240 V~ · 50/60 Hz · 2 × 10 A
FXNRK·7
SN IG4-C-000731

Fictional product. All figures, marks and serials are invented.

Box versus endpoint

What changes when you own it.

A hosted endpoint is the right call for plenty of work. Flip the switch to see what changes when it isn't.

Hosted endpoint Ingot
Where data goesOut to a provider's region, kept under their retention policy.Nowhere. The network cable is the boundary.
CostMetered per token, so the bill follows traffic.Fixed. One unit, depreciated like any other server.
Model changesWhen the provider decides.When you decide, and never silently.
OutagesTheir incident is your incident.Your rack, your redundancy, your schedule.
Latency floorA round trip over the internet, plus their queue.Your LAN.

Install

Racked in an afternoon.

No consultants and no onboarding call. If your team can rack a server, they can rack this. Here is the whole job, from the box to a live endpoint.

  1. 01Rack itTwo rails, four screws, two power cords20 min
  2. 02Connect the networkOne 100 GbE link, no outbound route10 min
  3. 03Load your weightsFrom a drive or an internal share1 h
Install log · IG-4 T+00:00

    Configurations

    Three sizes, one panel.

    Every unit ships with the same front panel and firmware. Pick one out of the rack. Prices are illustrative, like everything else here.

    Shown above

    IG-4

    The bench unit on this page. Most teams start here.

      $64,000one-time
      Configure IG-4

      Questions

      Before you ask procurement.

      The five things every security review asks first.

      Which models will it run?

      Any open-weight model you hold a licence for, up to roughly 380 GB at the precision you choose on a IG-4. Larger models run quantised, or across the eight cards of a IG-8.

      Does it ever need the internet?

      No. Updates arrive as signed offline bundles that you load from a drive, so the unit can spend its whole life on an air-gapped network.

      What happens when a part fails?

      Drives, fans and power supplies swap while it runs. If an accelerator fails, the unit drops to three cards and keeps serving at reduced throughput until the replacement arrives.

      Can several units work together?

      Yes. Units on the same network pool into a single endpoint and share requests between them, and a model too large for one unit can be split across two.

      How loud is it, really?

      About 38 dB(A) at a metre under normal load, which is roughly a quiet office. Turbo is louder, and the red lamp on the panel tells you when it's on.

      Nothing here is real

      It is a concept machine.

      Ingot is invented. There is no unit, no firmware and no telemetry, and every reading on this page is simulated in your browser.

      Turn the keyBack to the panel