Exascaler storage
A parallel file system for data-intensive workloads, such as training or large document collections. Each customer gets its own space, with independent quotas and access control.
own space · with quotas
Services
AI infrastructure, inference and consulting, from Spain
Three service lines on the same NVIDIA B200 platform in a Tier III datacenter in Madrid: dedicated GPUs, open-model inference and consulting. Start with any of them and move to another without changing provider or contract.
In short
- Dedicated NVIDIA B200 GPUs from €1.50/h (1/8 fraction) and €8.00/h for a full GPU.
- Open-model inference from €0.06 per million input tokens on GPU Flow.
- A free 45-minute first consulting session with an engineer.
- Data in Spain, our own team and a standard dedicated environment 72 hours after signature.
A · Dedicated GPU
Dedicated NVIDIA B200 GPUs
Full GPUs, hardware-isolated fractions or entire nodes. On demand with no commitment to validate, or reserved with guaranteed capacity for production. Your environment, your image, your access. Every configuration is listed on pricing.
from€1.50/h
Best for technical teams who want their own isolated GPUs in Spain, from an hourly pilot to production.
See pricing →MIG fraction
1/8 · 1/4 · 1/2A slice of the GPU with its own memory and compute, isolated in hardware through MIG (Multi-Instance GPU), so no other customer uses your resources. 1/8 with about 23 GB for €1.50/h, 1/4 with 45 GB for €3.00/h and 1/2 with 90 GB for €6.00/h. For inference, OCR, voice and mid-sized models.
Dedicated pod
1 to 4 GPUs · passthroughOne to four full physical B200 GPUs inside our cluster, assigned in passthrough mode: your environment talks directly to the hardware, with no virtualisation layer in between. Each GPU has 180 GB of HBM3e memory at 8 TB/s. We deliver an image configured with your project's libraries. €8.00/h per GPU.
Full bare-metal node
8 × B200 · NVLinkAn entire eight-B200 server for you, outside the shared cluster, delivered assembled and with your own access. The GPUs talk to each other over NVLink 5 at 1.8 TB/s each and the node connects over InfiniBand NDR at 400 Gb/s. For training, fine-tuning and large models, at €64.00/h.
Development sandbox
from €0.15 per active hourA ready-to-use container with SSH, Python, PyTorch, Jupyter and coding agents. You pay only for the hours you use it, because it pauses when you stop working. It lets your team try models, prepare data or build before moving to a dedicated GPU.
B · Inference
Open models, served from Spain
Use the best open models without managing GPUs. Through an API compatible with the OpenAI and Anthropic SDKs, or on a private endpoint for your organisation only. Everything is processed in Madrid and nothing you send is used to train models.
from€0.06/M tokens
Best if you want AI in your applications and processes without running the infrastructure.
GPU Flow →Pay-per-token API on GPU Flow
pay as you goModels such as GLM 5.3, Qwen3.8, Gemma 4 26B, Nemotron 3.5 Lightning 30B or ALIA 40B, from the Barcelona Supercomputing Center, available on GPU Flow. You pay for the tokens processed, from €0.06 per million input tokens. If you already use the OpenAI or Anthropic SDKs, change the base URL and the key in your code and you are done.
Private dedicated endpoint
reserved GPUThe open model of your choice, or your own fine-tuned model, deployed on GPUs reserved for your organisation only. No queues shared with other customers, your own access rules and capacity sized for your volume. It is the right option when confidentiality or a sustained workload calls for it.
Voice and documents
Whisper · OCR · TTSAudio transcription with Whisper large-v3, reading of scanned documents with vision models such as Qwen3-VL 32B and Nemotron Nano 12B VL, and text-to-speech. For minutes, meeting recordings, paper case files or phone support, with the data processed in Spain.
C · Consulting
AI infrastructure, done with judgement
We built and operate our own infrastructure. We help you decide on yours with your workload measured on our cluster, before you commit budget, and we support you through design, deployment and operations.
first session45 min, free of charge
Best if you need to decide what AI infrastructure your organisation needs and want to do it with measured data.
Book the session →Sizing
memory · compute · networkWe translate your use case (models, context length, concurrent users and expected response time) into the right configuration. We do not estimate it on paper: we measure it by running your workload on our B200 cluster, so the decision rests on real data rather than spec sheets.
NVIDIA reference architecture
DGX Spark → HGX B200The right product for each stage, following NVIDIA's reference architecture: DGX Spark on the desk for pilots and development, RTX PRO workstations for departments and HGX B200 for production. We help you fit each piece with the network, storage and software you need.
GPU datacenter design
density · coolingPower per rack, direct-to-chip liquid cooling, high-speed InfiniBand networking and parallel storage. We apply what we learned deploying and operating NVIDIA Blackwell on our own cluster, so your datacenter is ready for the density of today's GPUs.
Managed operations
your hardware, our teamIf your organisation has its own infrastructure, we run it for you: monitoring, updates, maintenance and support, with NVIDIA Base Command Manager and our team in Spain. Ownership of the hardware and the data stays with you at all times.
Add-ons
What you can add to any environment.
We include them in the proposal depending on your case. Ask us about any of them.
High availability
Replacement GPUs in the same cluster. Because your data lives on Exascaler storage rather than on the GPU, the environment restarts on another GPU without moving anything.
standard < 24 h · 24/7 < 4 h
Dedicated connectivity
An encrypted point-to-point link between your office or datacenter and our cluster in Madrid, through a connectivity partner, so your traffic never crosses the public internet.
post-quantum cryptography available
What to choose
Your need, the right option.
A quick guide. If your case does not fit any row, tell us about it and we will work it out together.
| Need | Line | Mode | From |
|---|---|---|---|
| Try an open model this week | B · Inference | Pay-per-token API on GPU Flow | €0.06/M tokens |
| Build and experiment with coding agents | A · Dedicated GPU | Development sandbox | €0.15/active hour |
| Serve a mid-sized model, OCR or voice in production | A · Dedicated GPU | MIG fraction 1/8 to 1/2 | €1.50/h |
| Give AI to the whole workforce with sensitive data | B · Inference | Private dedicated endpoint | Based on volume |
| Fine-tune or train a model on your data | A · Dedicated GPU | Dedicated pod or 8 × B200 node | €8.00/h |
| Decide your organisation's AI infrastructure | C · Consulting | Sizing | 45 min, free |
Frequently asked questions
What is the difference between on-demand and reserved GPUs?
On demand you pay by the hour, with no commitment and subject to availability; reserved guarantees the capacity, with a six-month minimum. A full B200 costs €8.00/h on demand or €5,840 per month reserved (730 hours). For longer commitments, let's talk.
What is a MIG fraction and is it secure?
It is a slice of a B200 GPU isolated in hardware, with its own memory and compute. MIG (Multi-Instance GPU) is an NVIDIA technology that separates fractions at chip level, so no other customer can see or use your resources. It ranges from 1/8 (about 23 GB) to 1/2 (90 GB).
Can I use my own model?
Yes. You can deploy it on a dedicated pod with your own image or ask us for a private endpoint that serves it to your organisation only. Your models and their weights remain yours and you can take them with you whenever you want.
How long does it take to get my environment ready?
A standard dedicated environment is ready 72 hours after signature. The GPU Flow API is self-serve, so you can start using it as soon as you sign up.
Do prices include VAT?
No. All published prices exclude VAT and are listed on the pricing page. Add-ons (storage, high availability and dedicated connectivity) are included in the proposal depending on your case.
Can I switch mode or service line later?
Yes. All three lines run on the same platform, so you can move from the API to a dedicated GPU, or from a fraction to a pod, without changing provider. Your data, models and weights are yours at all times.
Next step
Tell us about your case and we will propose the configuration.