Hardware rental

A dedicated machine,
and the engineer who runs it.

Three boxes, rented by the hour — an H200, a DGX Spark, and the two Sparks wired together as one 256 GB pool. One tenant at a time, billed for the hours you actually use. No minimum term and no yearly contract.

  • hourly
  • single tenant
  • UPS & cooling
  • Poland · EU
  • same-day slots

Ask about a slot See the rates

The three machines one tenant at a time

NVIDIA H200

141 GB HBM3e
$3.50 net / GPU-hour

The memory-bandwidth machine. 4.8 TB/s to 141 GB of HBM3e.

Accelerator
NVIDIA H200 Tensor Core GPU
Memory
141 GB HBM3e
Bandwidth
4.8 TB/s
Tensor cores
4th-gen, FP8 Transformer Engine
Host
Dedicated server, single tenant
Storage
Local NVMe scratch
Draw
~700 W under load
What it's for
  • Training and fine-tuning runs that need the whole GPU
  • Throughput inference at large batch
  • Memory-bound work — long context, big activations
  • Large-dataset analysis and rendering
Book NVIDIA H200 →

NVIDIA DGX Spark

GB10 · 128 GB unified
$0.50 net / hour

A 1-petaFLOP FP4 desk machine with 128 GB the CPU and GPU share.

Accelerator
GB10 Grace Blackwell Superchip
Memory
128 GB LPDDR5X, coherent CPU+GPU
Bandwidth
273 GB/s
AI throughput
1 PFLOP FP4 (sparse)
CPU
20-core Arm — 10× X925 + 10× A725
Storage / NIC
4 TB NVMe · ConnectX-7, 200 Gb/s
Draw
~170 W
What it's for
  • FP4 inference up to ~200B parameters
  • LoRA / QLoRA fine-tunes up to 70B
  • Agent loops that want the model on the same box
  • Air-gapped work — the machine can run with no uplink
Book NVIDIA DGX Spark →

2× DGX Spark cluster

256 GB unified · 200 Gb/s RDMA
$0.90 net / hour

Two GB10 nodes stitched over ConnectX-7. One 256 GB memory pool.

Accelerators
2× GB10 Grace Blackwell Superchip
Memory
256 GB unified across two nodes
Bandwidth
273 GB/s per node
AI throughput
2 PFLOPS FP4 (sparse)
Interconnect
ConnectX-7, 200 Gb/s QSFP, RDMA
Storage
2× 4 TB NVMe
Draw
~340 W
What it's for
  • FP4 inference up to ~405B parameters
  • Multi-node NCCL / RDMA rehearsal before you rent a real cluster
  • Models that will not fit one 128 GB node
  • Sharded serving with both nodes on one tenant
Book 2× DGX Spark cluster →
Side by side manufacturer figures
H200 DGX Spark 2× DGX Spark
Accelerator NVIDIA H200 GB10 Grace Blackwell 2× GB10 Grace Blackwell
Memory 141 GB HBM3e 128 GB LPDDR5X unified 256 GB unified
Memory bandwidth 4.8 TB/s 273 GB/s 273 GB/s per node
AI throughput FP8 Transformer Engine 1 PFLOP FP4 sparse 2 PFLOPS FP4 sparse
Node interconnect — ConnectX-7 200 Gb/s ConnectX-7 200 Gb/s RDMA
Local NVMe NVMe scratch 4 TB 2× 4 TB
Largest model, FP4 inference 141 GB resident ~200B params ~405B params
Fine-tuning headroom full-GPU runs up to 70B (LoRA) up to 70B, sharded
Power draw ~700 W ~170 W ~340 W
Rate, net $3.50 / GPU-hour $0.50 / hour $0.90 / hour
How it works reply inside one business day
  1. 01 Ask Mail me the dates and a rough number of hours. You get an answer inside one business day.
  2. 02 Terms A slot, a rate, and what the machine will have installed on it. Net USD, VAT invoice.
  3. 03 Run SSH in, the box is yours alone for the window. Questions reach the engineer, not a support desk.
  4. 04 Invoice Billed for the hours you actually used. No minimum term, no yearly contract.
Operating conditions who runs the machine — I do
Tenancy
One tenant at a time. Nothing else is scheduled on the box during your window.
Power & cooling
UPS-backed, air conditioned, monitored.
Connectivity
Fibre with an LTE backup link.
Location
Poland, European Union.
Availability
Around the clock. Only announced maintenance windows.
Billing
Net USD per hour actually used, VAT invoice.

Give me a date and a rough number of hours.

That is enough to confirm a slot. Setup, monitoring, security and upkeep are mine; the GPU time is yours. Technical questions go straight to the engineer.

hello@robertz.co