Models. Private inference. GPUs.

Building the compliant full-stack AI cloud in Europe.

European GPU infrastructure, bare metal and hosted open-source models — optimized for performance, sovereignty and cost.

Get an API key See the hardware

Built by founders and engineers with 15+ years in highly regulated industries. Billed by the second.

In shortAvailable

DGX Spark: 4 ready nowrented monthly · more arriving
Models: €0.015per million tokens, from
GPUs: €0.17per GPU-hour, from
Capacity: SE + ESSweden and Spain — more EU regions in deployment
Deploys: Daysselected GPU systems, subject to availability
Your data: Neverused to train a model

Three products

Run AI your way.

Same account, same EU regions, three amounts of control.

Serverless · Managed · VM

Serverless Models

Choose a hosted model and call the API. We operate everything underneath.

Available

€0.29/M input tokens, from

Explore models
Managed GPU

Tell us what to run. We deploy and operate the environment for you.

Available — DGX Spark today

€495/month, dedicated machine

Explore managed compute
GPU VM

Rent the GPU directly with VM and SSH access. You control the stack.

Available — more tiers on request

€0.17/GPU-hr, from

Explore GPUs

The fleet

Infrastructure behind AxForge.

Named hardware, honest statuses. Consumed through the three products above.

EUR · ex VAT

The AxForge hardware fleet with memory, status and list price
System	Memory	Status	Price
NVIDIA DGX SparkGB10 Grace Blackwell · 8 live, 3 arriving	128 GB unified	Available	€495/month
RTX 6000 ProBlackwell workstation	96 GB GDDR7	On request	€1.73/GPU-hr
4× RTX 5090One system	32 GB GDDR7 each	On request	€0.54/GPU-hr
4× RTX 3090Budget tier — and honest about it	24 GB GDDR6X each	On request	€0.17/GPU-hr
4× RTX 3060 ×2Starter tier — two 12 GB GPUs per machine	12 GB GDDR6 each	Available	€0.10 / hour
NVIDIA H100 / H200Bought against registered demand	80 / 141 GB HBM	On request	—
NVIDIA B200Planned	192 GB HBM3e	Coming	—

Models are €0.015 – €1.77 per million tokens on the same account. Committed-use pricing on request. Hardware without a price carries none until it lands — nothing here is invented.

eu-se-1 Stockholm
eu-es-1 Málaga
More EU regions — in deployment

What we took out

Six things you expect. None of them are here.

Every one of them is normal somewhere else. You have probably paid for a few.

One line each

No sales call to learn the price. Every price is printed on this page.
No egress fees. What you pull out is not metered against you.
No minimum commitment. Billed per second, so you stop when you stop.
No invented SKU names. DGX Spark is the part, and the invoice line.
No training on your data. In the contract, and asserted on every response.
No lock-in to one level. Move up or down on the same account.

The proof

Headers are the promise.

One request. The response says where it ran and what it cost.

Returned on every call

api.axforge.ai/v1TLS 1.3 · terminated in the EU

# one variable changes. your client does not.
$ export OPENAI_BASE_URL=https://api.axforge.ai/v1
$ curl -sSD- "$OPENAI_BASE_URL/chat/completions" \
    -H "authorization: Bearer $AXFORGE_KEY" \
    -H "x-axforge-region: eu-se-1" \
    -d '{"model":"qwen3.5-397b-a32b", ... }'

< HTTP/2 200
< x-axforge-region:      eu-se-1
< x-axforge-retention:   none
< x-axforge-training:    disabled
< x-axforge-accelerator: nvidia-dgx-spark
< x-axforge-cost-eur:    0.000163

Where it executed: eu-se-1 · StockholmPinned on the key, echoed on the response.; In region
What happened to the data: Retention none · training disabledNothing written after the response.; Verified
The machine class: NVIDIA DGX Spark · 128 GBThe same part you can rent at GPU VM.; Named
The price: €0.000163 for this callMetered per token, billed by the second.; Metered
Read it yourself: Your logs, not our slidesAn auditor can check the headers instead of asking us.; Every call

Sovereignty

Where your data lives, as facts.

The six questions a procurement team asks, answered before they ask.

Enforced per request

Sovereignty facts: regions, data at rest, training, operator, subprocessors and egress
Fact	Answer	Detail
Regions	eu-se-1 · eu-es-1	Stockholm and Málaga — additional EU regions in deployment
Data at rest	EU only	Prompts, outputs, weights and logs, keys held in region
Training	Never	Not our models, not a third party's; there is no toggle
Operator	AxForge	EU entity, EU-resident staff, EU control plane
Subprocessors	Named in the DPA	All EU-established, 30 days' notice before a change
Egress	None outside the EU	No egress fee inside it; metadata kept 30 days, in region

Residency

Residency is enforced at the edge — a request pinned to a region fails rather than leaves it.

Audit position

We make no certification claims on this page; ask for the current audit position and you will get a straight answer.

Figure — the path of one request Application → EU ingress → named accelerators

Your application Any OpenAI-compatible client, on your own infrastructure.

EU region boundaryeu-se-1

EU ingress · api.axforge.ai Auth and metering. The region is pinned on the key.

Inference on named accelerators Your dedicated systems, in that region only.

Training corpus

No path leads into it. Your prompts, outputs and weights never train a model.

Request path

EU region boundary

No path — training corpus unreachable

Start here

Tell us the shape and the timeline.

Or skip the conversation entirely and take a key now.

Get an API key Talk to an engineer

A key works with any OpenAI-compatible client the moment it is issued.

What we need from you

For models: Model, tokens per day, latency target, region.
For GPUs: Accelerator, count, fabric, region, term.
You get: A written configuration and a price.

Building the compliant full-stack AI cloud in Europe.

Run AI your way.

Serverless Models

Managed GPU

GPU VM

Infrastructure behind AxForge.

Six things you expect. None of them are here.

Headers are the promise.

Where your data lives, as facts.

Tell us the shape and the timeline.

Keys are issued in order.

Tell us the shape. An engineer replies.