Start and stop when you want
Dedicated GPU, per-second billing, stop whenever you want.
Deploy GPUs across Europe in minutes
Compare offers from European data centres, pick your site and pay per use.
Per-second billingPrices in €Data stays in the EUEU company
38sites
23countries
0offers
0,0 ssearch
00:00:00€0.0000
01What halgrid is
See every GPU in Europe
Europe has more GPUs than it uses. They sit idle in regional data centres, universities and telco sites while teams queue for hyperscaler quotas.
We connect those sites into one network and put it under one lens: where each GPU is, what it costs and who operates it. You pick the site, deploy in minutes and pay per use. Like a power grid: produced in many places, delivered through one plug.
Built for AI startups, research labs and companies with sensitive data. Independent developers too.
02How it works
Go from GPU model to running instance
Self-serve from start to finish. Your first account is set up with you, in person.
- 01
Choose a GPU
CatalogueModel and memory, from 24 GB to an H200. One GPU or a cluster.
gpu h100-80gb count 4 - 02
Compare offers
SitesSite, country, price per hour, availability. Filter by location or price.
site Frankfurt · DE price €1.94/h - 03
Deploy
RuntimeSSH, your container or a template: vLLM, PyTorch, ComfyUI, Whisper.
image vllm/vllm-openai status running - 04
Pay per second
BillingIn euros, one invoice. Stop the instance, stop the meter.
runtime 03:12:40 cost €6.21
03GPUs and pricing
Know the price before you deploy
Each site publishes its price per GPU-hour. No quotes, no negotiation. Reference prices below, for planning.
| GPU | Memory | Best for | Reference price | |
|---|---|---|---|---|
| RTX 4090NVIDIA · Ada | 24 GB | Image generation, small models, prototyping | from €0.35/h | Request → |
| L40SNVIDIA · Ada | 48 GB | Inference, media pipelines | from €0.85/h | Request → |
| RTX PRO 6000NVIDIA · Blackwell | 96 GB | Media generation, cost-effective inference | from €1.40/h | Request → |
| A100NVIDIA · Ampere | 80 GB | Batch inference, embeddings, fine-tuning | from €1.30/h | Request → |
| H100NVIDIA · Hopper | 80 GB | LLM inference, 70B-class models | from €1.90/h | Request → |
| H200NVIDIA · Hopper | 141 GB | Large models, long context | from €2.60/h | Request → |
| B200NVIDIA · Blackwell | 180 GB | Large-scale inference, frontier models | from €4.50/h | Request → |
Reference on-demand prices per GPU-hour, European market, September 2026, excl. VAT. Interruptible and reserved capacity: quoted per site.
Pay less when your job can wait
Idle capacity at a discount. May be paused when demand rises.
Lock in capacity for months
GPUs locked at one site for one to six months, at a lower rate.
Call an endpoint, skip the ops
OpenAI-compatible endpoint for open-weight models, on your site.
04The network
Run on data centres we have already vetted
We own no GPUs. Every operator passes one standard before its capacity reaches you.
What every site must pass
- Verified legal entity and physical location
- Access control, redundant power and cooling
- Hardware benchmarked by us, not declared
- Minimum network and storage performance
- Continuous monitoring, incident reporting
- Data deletion procedure, signed contract
Operate GPUs in Europe? Join the network.
Add your capacity
05Trust
Sovereignty you can check
Where the hardware is, who runs it, which law applies. Three answers you can verify.
You choose the site
Every offer names its site and country. Your job runs there and nowhere else.
- Site and country on every offer
- Per-job log: site, hardware, time
- Data leaves the EU only if you choose
Named European operators
Each site is run by an identified European operator under contract with us.
- Public subprocessor list
- ISO 27001 or Tier III facilities only
- Operators vetted before they join
EU company, EU contract
An EU company under EU jurisdiction. No US parent, so no CLOUD Act exposure.
- Article 28 DPA with every account
- Encryption in transit and at rest
- Dedicated GPUs, isolated networks
06Straight answers
Questions before your first job
What happens if a site goes down?
Your instance runs on one site, like any dedicated server. If it fails, we notify you, help you restore from persistent storage and provision equivalent capacity on another site. Automatic failover is on the roadmap; we will not claim it before it exists.
How is the price set?
Each operator prices its own capacity. We show every offer with its site and price, add a transparent platform fee, and you pick. The table above shows market references; the real offer is the one you see before deploying.
Which workloads are not a fit?
Large training jobs split across sites. Sites connect over the internet, not InfiniBand, so a training job stays inside one site. Inference, fine-tuning, batch, embeddings, rendering and single-site clusters fit well.
Where is my data?
At the site you selected, nowhere else unless you move it. Each job logs site, hardware and time. Our control plane runs in the EU.
Are you subject to the US CLOUD Act?
No. We have no parent or subsidiary in the United States. Like any EU company we answer lawful requests from European courts, and we would tell you unless the law prevents it.
How do I start?
Click Try halgrid and tell us what you will run. We set up your account and first GPUs with you, personally.
Do you offer managed inference endpoints?
Managed inference: an OpenAI-compatible endpoint for open-weight models on the site you choose, operated by us. Ask us to set one up.
Ready when you are
No quotas, no procurement cycle, no data leaving the EU. Tell us what you will run and we set up your first GPUs with you.