Ikioma Mini
The desk-sized entry point — two linked nodes, one private stack.
from €50,000
View →
Made to order · 2–8 weeks
Configure yours →
Ikioma
The recommended tier — 384 GB VRAM, four Blackwell GPUs, one integrated stack.
From €40,000
Hardware, delivered and deployed
Support & optimization
€13,000/year
Year 1 total
€53,000
[ why own it ]
01 · physics
Private by physics.
Your data never leaves the floor. Air-gappable, no egress, no telemetry — the trust boundary has a serial number you can point at in an audit.
02 · metering
Pay once, infer forever.
Own the machine and agents run around the clock. No per-token meter, no overage invoices — you pay for electricity, not tokens.
03 · ownership
Yours to keep.
Hardware is a capital asset, not a subscription line. Weights, tunes, and stack stay yours — perpetually, with trade-in credit on next-gen silicon.
01 · your work
Models that know your work.
Open-weight foundation models, post-trained on the categories of work your team actually does — document understanding, summarisation, structured extraction, code, agentic tool use. We benchmark against your task suite, not generic leaderboards — and the weights are yours, perpetually.
72B
class of open models
128ktok
context — long documents, long tasks
100%
of weights owned by you
02 · autonomy
Agents run around the clock.
Own the machine and your agents work all night — no rate limits, no per-token meter, no "try again later". Coding loops and long-horizon tasks run until they're done, and 128k tokens of context keeps them coherent start to finish.
24/7
unmetered inference
128ktok
context for long tasks
∞
no rate limits, no overages
03 · operations
You don't need to become an ML team.
The runtime drops in as an OpenAI- and Anthropic-compatible API, so your existing tools keep working. We handle model updates, security patches, and performance tuning; updates are signed and reversible, and a named engineer owns your account.
/v1
drop-in — OpenAI / Anthropic compatible
1binary
self-hosted runtime
1owner
named engineer per account
04 · economics
A capital asset, not a subscription.
Private inference is CapEx — it goes on the balance sheet, not the expense line. Most customers break even versus equivalent cloud spend within 4–7 months on typical agentic workloads; after that, the machine keeps compounding.
4–7mo
typical break-even vs cloud
10yr
security updates, in writing
€0
per token, after purchase
[ heritage ]
Built in the lineage of instruments, not platforms.
Assembled in Yokohama and Eindhoven. Boards burn in for 96 hours before they leave the floor. Every system carries a serial, a calibration sheet, and a name etched on the chassis — because the people who built it stand behind it for a decade, not until next quarter's earnings call.
96h
burn-in per system
before it leaves the factory floor
before it leaves the factory floor
2
assembly sites
Yokohama · Eindhoven
Yokohama · Eindhoven
10yr
security update commitment
in writing
in writing
[ the system ]
Hardware spec
GPU
4× RTX Pro 6000 Blackwell
VRAM
384 GB
CPU
32-core AMD Genoa
RAM
192 GB
Storage
4 TB high-speed RAID + 1 TB boot
PSU
2× 1600W
Dimensions
19″ W × 21″ H × 16.25″ D
Weight
90 lbs
lead_time
Made to order · 2–8 weeks
Model & inference stack Included
Preinstalled and pre-validated: the ikioma inference runtime, your choice of optimized open models, and the production middle layer (audit log, RBAC, evals) configured out of the box.
- ✓OpenAI / Anthropic API compatibility
- ✓Sandboxed tool execution
- ✓Signed audit log
- ✓Runs on-premises — your hardware, your premises
| Model | Params | Context | tok/s | Latency | Notes |
|---|---|---|---|---|---|
| Qwen 3 72B | 72B | 128k | — | — | TBD — burn-in benchmark |
| Llama 3.3 70B | 70B | 128k | — | — | TBD — burn-in benchmark |
Loading 3D view…
Configuration
Drag to rotate · Scroll to zoom
GB10 telemetry · simulated
GPU temp—
GPU clock—
Unified mem—
Blowers—
System power—
the hardware
The workhorse, up close.
Four Blackwell GPUs and the full cooling array. Drag to orbit, run the explode slider to lift the assembly apart layer by layer, X-ray the shell to see inside — or switch the configuration to the open frame or the GPU tower.
Illustrative 3D approximation — not manufacturing CAD. Schematic dimensions only.
[ which ikioma ]
Three versions, one stack.
Mini for the desk, Ikioma for the team, Pro for the fleet. Every version ships the same model and inference stack — the scale is what changes.