Offshore GPU serversWhole cards, by the second. Nobody sharing them.
RTX 4090, L40S, A100, H100 and H200 on EPYC hosts with local NVMe scratch, in any of the eight jurisdictions. Never a time-sliced fraction of a card, and nothing you train is ever looked at.
- G1 · 4090GPURTX 409024 GB GDDR6XvCPU8 vCPUEPYC 9004Memory48 GBDDR5 ECCScratch500 GBNVMe scratch0,27$/hr
$0,34elsewhere - G2 · 4090GPU2 × RTX 409048 GB GDDR6XvCPU16 vCPUEPYC 9004Memory96 GBDDR5 ECCScratch1 TBNVMe scratch0,53$/hr
$0,68elsewhere - G1 · L40SBest valueGPUL40S48 GB GDDR6vCPU12 vCPUEPYC 9004Memory96 GBDDR5 ECCScratch1 TBNVMe scratch0,57$/hr
$0,72elsewhere - G1 · A100GPUA10080 GB HBM2evCPU16 vCPUEPYC 9004Memory128 GBDDR5 ECCScratch1.5 TBNVMe scratch0,85$/hr
$1,07elsewhere - G1 · H100GPUH100 SXM80 GB HBM3vCPU20 vCPUEPYC 9004Memory192 GBDDR5 ECCScratch2 TBNVMe scratch1,59$/hr
$2,01elsewhere - G4 · L40SGPU4 × L40S192 GB GDDR6vCPU48 vCPUEPYC 9004Memory384 GBDDR5 ECCScratch4 TBNVMe scratch2,27$/hr
$2,88elsewhereSold out - G1 · H200GPUH200 SXM141 GB HBM3evCPU24 vCPUEPYC 9004Memory256 GBDDR5 ECCScratch3 TBNVMe scratch3,49$/hr
$4,39elsewhere - G8 · H100GPU8 × H100 SXM640 GB · NVLinkvCPU128 vCPUEPYC 9004Memory1.5 TBDDR5 ECCScratch8 TBNVMe scratch12,69$/hr
$16,08elsewhereSold out
Billed by the second on-chain, from ready to destroyed — stopped nodes cost nothing, and there is no minimum term. Reserve by the month for roughly 30 % less. Elsewhere is the cheapest published on-demand rate for the same card at a named GPU cloud; interruptible marketplaces of consumer hardware are not counted, because a card that can vanish mid-run is not the same product.
Included on every node
- Billed by the second, no minimum term
- Root on bare metal or a container
- CUDA 12, PyTorch and TensorFlow preloaded
- NVLink on multi-card nodes
- Dedicated IPv4 + IPv6
- No egress charge on model weights
- Reserve monthly for a further 30 % off
- Crypto settlement, no identity checks
Which card fits the job?
RTX 4090
Inference, image and video generation, fine-tuning models that fit in 24 GB. The cheapest way to keep something warm.
L40S
48 GB with proper encoders: rendering, transcoding, and inference that outgrew a 4090.
A100 · H100
Training runs and large-context inference where HBM bandwidth, not core count, is the wall.
H200 · 8 × H100
Models that will not fit anywhere smaller, and multi-day runs that must not be interrupted.
After a card that is not listed, or a host built around several of them? Specify it in the custom builder
The platform
The whole card, or we do not sell it.
Much of the cheap GPU capacity on the market is a slice of a card, a consumer machine in somebody's flat, or an instance that disappears when a higher bidder appears. None of that is on offer here.
- Accelerators
- NVIDIA Ada and Hopper RTX 4090, L40S, A100, H100 and H200. Whole cards passed through — never a time-sliced or virtualised fraction of one.
- Host
- AMD EPYC 9004 · DDR5 ECC Sized so the GPU is the bottleneck, not the loader. PCIe 5.0 throughout, NVLink between cards on multi-GPU nodes.
- Scratch
- Local NVMe, 0.5 – 8 TB On the node itself, not a network volume. Datasets stage at full disk speed instead of crawling over a fabric.
- Billing
- By the second Stop the node and billing stops with it. Reserve by the month for roughly 30 % less if the work is continuous.
- Runtime
- Bare metal or container Take the whole machine, or a container with CUDA 12, PyTorch and TensorFlow already in place. Custom images welcome.
- Privacy
- Nothing inspected We do not scan weights, datasets or output, and nothing is retained after you destroy the node.
Preloaded, or bring a container from any registry
- Ubuntu 24.04 LTS
- CUDA 12.4
- PyTorch 2.5
- TensorFlow 2.18
- JupyterLab
- Hugging Face TGI
- ONNX Runtime
- NVIDIA NGC images
- Any OCI registry
- Custom image
Ubuntu 24.04 LTS
CUDA 12.4
PyTorch 2.5
TensorFlow 2.18
JupyterLab
Hugging Face TGI
ONNX Runtime
Any OCI registry
Custom imageAvailable in any of the eight
The rate never changes with the region. Pick the jurisdiction, not the price.
- ReykjavíkIceland
- BucharestRomania
- SofiaBulgaria
- ChișinăuMoldova
- ZurichSwitzerland
- AmsterdamNetherlands
- Panama CityPanama
- SingaporeSingapore
Included at every size
Four things most of the trade charges for, and one of them twice.
-
Mitigation
1.2 Tbit/s DDoS protection
Always on at every edge, on the cheapest plan as much as the largest. No per-attack fee, no metered scrubbing and no protected-IP upsell.
-
Availability
99.99 % uptime SLA
Credits are worked out from the public incident record and applied to your balance automatically. There is no claim form, because there is nothing to claim.
-
Settlement
Crypto payment, no identity
Eight assets including Monero. No card, no bank detail, no document — not at signup, not at renewal, and not at any amount.
-
Moving in
Free migration
Whole servers, control panels, databases and mail, moved by an engineer at no charge. You test the copy before anything is cut over, and you keep the old host until you are happy.
Before you launch.
Is the card shared with anyone else?
No. Whole cards are passed through to your node — never a time-sliced MIG partition or a virtualised fraction. What the specification says is what you get for the duration, and nobody else is scheduled against it.
How is it billed exactly?
By the second, from the moment the node is ready to the moment you destroy it. There is no minimum term and no charge while it is stopped. Continuous work is cheaper reserved by the month — roughly 30 % below the hourly rate.
Can I bring my own image?
Yes. Boot a container from any registry, or take the node as bare metal and install what you like. Our images ship CUDA 12 with PyTorch and TensorFlow already built, which saves the usual first hour.
How do I get a large dataset onto the node?
Local NVMe scratch of half a terabyte to eight terabytes sits on the machine itself, so staging runs at disk speed rather than across a network volume. Ingress is never charged, and neither is egress on model weights.
Do you look at what I am training?
No. Weights, datasets and output are not scanned, indexed or retained, and destroying the node destroys the scratch with it. The same content policy applies here as everywhere else on the estate.
What if the card I want is sold out?
Accelerator supply moves in batches and the popular configurations do go. Tell an engineer what you need and roughly when; reserved capacity is allocated ahead of hourly demand, so a monthly commitment usually finds hardware faster.