DEMIRARI AI HARDWARE

Tell us the model size, the latency target, the power budget, and the space you have. We engineer a machine to hit exactly that — no wasted silicon, no bottlenecks, no cloud.

Demirari AI Rack — bespoke 4U AI server

WHY BUILT-TO-ORDER

Off-the-shelf is a compromise you pay for twice.

Right-sized

Pay for the accelerators and memory your model actually needs — nothing more.

Constraint-driven

Power cap? Noise limit? Rack depth? We design to your real constraints.

Co-designed with the model

Spec'd alongside your Modelsmith build so weights, quantization, and memory align perfectly.

THE CONFIGURATOR

Design your machine.

Six decisions. We handle the rest. Watch the spec and estimate update live.

Selecting a workload pre-loads sane defaults for every step below.

COUNT

×4 / max 8

MODEL

Total VRAM: 192 GB

SYSTEM MEMORY · DDR5 ECC

NVME GEN5 STORAGE

COOLING

POWER BUDGET · CIRCUIT

30A @ 208V
15A30A60A100A
Demirari AI Rack 4U chassisDemirari AI Rack 4U

LIVE SPEC

Recalculating live

EST. THROUGHPUT

~564 tok/s

EST. LATENCY

<29ms

TOTAL VRAM

192 GB

MEM BANDWIDTH

3.4 TB/s

POWER

2.5 kW

NOISE

49 dB

FITS MODEL

326B INT4

ESTIMATED BUILD PRICE · ONE-TIME

$101,500 – $123,500

Estimate only. Final spec confirmed on your design call.

STARTING POINTS

Three chassis. Infinite configurations.

Demirari AI Edge

01 — CHASSIS

Demirari AI Edge

Desk-side / branch / factory floor. A silent workstation-class unit that lives next to your team, not in a datacenter.

For teams, sites, and edge inference.

1–2 GPU128–512GB RAMSilent — ≤35dBSingle-phase power
Configure this
Demirari AI Rack

02 — CHASSIS

Demirari AI Rack

The production workhorse. Standard rackmount depth, redundant everything, sized precisely to your serving load.

For always-on inference & RAG.

2U / 4UUp to 8 GPUUp to 4TB RAMRedundant PSU · DLC option
Configure this
Demirari AI Cluster

03 — CHASSIS

Demirari AI Cluster

Multi-node scale-out with a pooled memory fabric. Built when one chassis is no longer the right answer.

For 100B+ models and heavy fine-tuning.

NVLink / 400GbE fabricPooled memory2–16 nodesPB-scale storage
Configure this

ANATOMY

Every component, chosen on purpose.

Exploded blueprint diagram of a Demirari AI server

Accelerators

matched to model memory

CPU

dual EPYC / Xeon

Memory

DDR5 ECC, sized to KV cache

NVMe Gen5

fast model load & RAG

Networking

100 / 400GbE

Power

redundant, right-sized

Cooling

air or liquid

SPECIFICATIONS

The envelope we build within.

SubsystemNote
Accelerators2–4 × L40S / RTX 6000 AdaMatched to model memory footprint
CPU1 × EPYC 9004 / Xeon 6Sized for tokenization & orchestration headroom
Memory512GB–1TB DDR5 ECCSized to KV cache + RAG working set
Storage16–30TB NVMe Gen5Fast model load, hot vector indices
Networking2 × 100GbECluster-only NVLink; air-gap option
Power1.5–6 kW, single-phaseRight-sized to your circuit — 15A to 100A
CoolingAir, tuned fan curvesDLC unlocks dense 8-GPU builds
Acoustics≤35 dB quiet modeOffice-safe to datacenter-grade
Form factorEdge tower / 2UStandard 19" rack depth options
SecurityTPM 2.0 + secure bootWeights and data never leave the chassis
ComplianceBuilt for SOC 2 / ISO 27001 environmentsDocumentation pack ships with every unit

Hover a row for the engineering note. Every range is a real build we ship — not a datasheet aspiration.

COMPARE THE LINEUP

Pick a starting point. We'll tune the rest.

Demirari AI Edge

Demirari AI Rack

Demirari AI Cluster

GPUs
1–2
Up to 8
8 / node · 2–16 nodes
VRAM
≤96GB
≤1.1TB
Pooled · multi-TB
Memory
128–512GB
Up to 4TB
Pooled across nodes
Storage
4–16TB NVMe
Up to 245TB NVMe
PB-scale + object tier
Networking
10/25GbE
2×100GbE · 400GbE opt
400GbE / NVLink fabric
Noise
≤35dB silent
35–48dB
Datacenter
Power
Single-phase, ≤1.5kW
≤12kW, redundant
Per-rack, 3-phase
Cooling
Air (quiet)
Air or DLC
DLC standard
Recommended for
Teams, sites & edge inference
Always-on inference & RAG
100B+ models & heavy fine-tuning

Book a design call