01 / QUEUE
The capacity exists, but not for you
Quota requests for a multi-node cluster at a hyperscaler run into weeks. Your team sits idle, the budget is already approved, and the project slips a quarter.
AI INFRASTRUCTURE PARTNER · EU DATA CENTRES
We build and run your AI infrastructure in Europe
In production from 45 days
You have the AI workload and the team. What you don't have is anyone to build and run the cluster underneath it. That's the part we own — design, migration, operation, and the security review — under local contract, in your language. You get the schedule and the cost in writing before you commit.
Book an architecture session
Where AI projects stall
The model is not the bottleneck.
The infrastructure underneath it is.
02 / COST
You pay hyperscaler rates for research work
Published on-demand H100 pricing at the large clouds runs roughly 1.8–2.9× above specialist European providers, before egress. On exploratory training, that difference decides how many experiments you get to run.
03 / PEOPLE
Nobody owns the cluster
An in-house AI infrastructure engineer costs from €120,000 a year in Germany and Switzerland, and you need more than one for round-the-clock operation. Mid-sized companies rarely have that headcount to spare.
What we do
Three stages, priced separately
Most clients start with Launch and add Operate once the first workload is in production.
Each is priced separately, so you can stop after the first.
Stage 01
Launch
From signature to a running workload. Fixed scope, fixed price, from six weeks.
- Workload assessment and sizing
- Cluster and network design
- Migration of the first workload
- Scheduler and MLOps setup
- Security review support
- Handover training for your team
Stage 02
Operate
We run the contour so your engineers don’t have to. Monthly retainer. Monthly rate via:
- Monitoring and incident response
- Capacity and queue management
- Patching and upgrades
- Named engineer on your account
- Quarterly architecture review
Stage 03
Optimise
Once the workload is steady, we bring the cost per useful result down.
- Utilization and idle analysis
- Checkpoint and restart tuning
- Precision and batch size tuning
- Inference cost per token
- Commitment sizing
- Reported against a baseline
How a launch engagement runs
Day 0–5
Architecture session.
Sizing, lead time and cost in writing.
Day 5–14
Contour stands up. First training run executes on your data.
Day 14–30
Compliance pack delivered while remaining pipelines migrate.
Day 30–45
Production under monitoring. Your team trained, runbook handed over.
Indicative. The schedule for your project is fixed at the architecture session.
Economics
Compare the finished run,
not the hourly rate
Published hourly rates are the easiest number to check and the least useful one to decide on. They ignore restarts, idle time, egress and the weeks your team waits for a quota.
At the architecture session we model your actual workload — how many runs, how long, how often they fail and restart — and cost it across the providers that meet your residency requirement. You get the comparison in writing, including the option of staying where you are.
Get the comparison on your workload
INDUSTRIES
Who we build for
Manufacturing & engineering
You have R&D, not a platform team
Vision on the line. Simulation and digital twins, predictive maintenance, generative design. Two to five ML people and nobody who runs clusters. Procurement in German, contract under local law.
How this works →Pharma & life sciences
Reproducibility matters more than price
Molecular modeling, candidate screening, bioinformatics, imaging. Predictable baselines, strict expectations on data handling, and no local capacity at this scale in Switzerland.
How this works →AI-native companies
You need the cluster this month
Post Series A or B, product already in production. Fine-tuning and inference at scale. No quota queue, no lock-in, no egress bill, and a contact who answers the same day.
How this works →
Before your security team asks
Common questions
Every regulated deal reaches the same three questions. We’d rather you hear our answer at the first meeting than at the review board.
Full residency & jurisdiction packFrom six weeks, depending on cluster size and how much needs migrating. The first training run typically lands in the third week. The schedule starts when compute is allocated, and we fix the date in writing at the architecture session — before you commit.
Less than the published hourly rate suggests, and more than the cheapest quote you will find. The number that matters is the cost of a completed training run. We model that on your workload at the architecture session.
Data residency is fixed at the architecture session, against your requirements. You hold the encryption keys and the platform contract in every case.
It depends on your sector. Public procurement and parts of finance require BSI C5 in Germany or SecNumCloud in France. We check what applies to you before the architecture session and publish what a provider actually holds.
Next step
Book an architecture
session
Ninety minutes with an engineer, not a sales call.
You bring one real workload. We leave you with a sizing, a lead time
and a cost — whether or not you work with us.
- We size your actual workload, not a reference profile
- You get the configuration and cost in writing
- We say on the call if you’re better served elsewhere
- No charge, no obligation, no deck