Solution 01 — Private model hosting

Your models. Your data. EU soil.

Dedicated, single-tenant model instances operated by CIPHER in European data centres — the operations burden stays with us, the isolation stays with you.

Book a consultationSee the architecture

1

Tenant per GPU

no shared inference

99.9%

Uptime target

contractual SLA

2 wks

Typical go-live

single instance

EU

Regions

Amsterdam, Frankfurt

The problem

What keeps your compliance officer awake

01

Prompts leave the EU

Public AI APIs process your input on infrastructure you cannot name or audit.

02

Fine-tuned weights leak know-how

Uploading adapters to a shared endpoint moves your IP to someone else's disk.

03

Shared inference is noisy

Other tenants' traffic sets your latency and your throughput ceiling.

04

AI Act obligations are real

GPAI transparency, logging, and human oversight are now legal requirements.

Where the prompt goes

Your application

The prompt never leaves your network unencrypted.

Your service calls a private endpoint inside your VPC. There is no public URL to leak, no shared hostname, no key that also works somewhere else.

Request path

1/5

  1. Your VPC
  2. Encrypted link
  3. Private instance
  4. GPU pool
  5. Response + trace

The request starts inside your own network boundary.

The link

Private peering, not the open internet.

Traffic crosses a dedicated link or a VPC peering attachment. TLS 1.3 in transit, mutual authentication, and an allow-list at both ends.

Request path

2/5

  1. Your VPC
  2. Encrypted link
  3. Private instance
  4. GPU pool
  5. Response + trace

Private peering or dedicated interconnect, TLS 1.3, mutual auth.

The instance

One tenant per deployment.

Your models run on GPUs allocated to you. No neighbour batch job sets your latency, and no other customer's prompt shares your memory.

Request path

3/5

  1. Your VPC
  2. Encrypted link
  3. Private instance
  4. GPU pool
  5. Response + trace

Single-tenant inference. Your weights, your memory, your latency.

The GPUs

Hardware we operate, in regions we can name.

Amsterdam and Frankfurt, with the GPU model fixed per contract. Capacity is not oversubscribed, and the allocation is visible in your console.

Request path

4/5

  1. Your VPC
  2. Encrypted link
  3. Private instance
  4. GPU pool
  5. Response + trace

Allocated capacity in Amsterdam or Frankfurt, fixed per contract.

The return

The answer comes back inside the same boundary.

Response, usage record, and trace return through the same path. Nothing was routed to a region you did not approve.

Request path

5/5

  1. Your VPC
  2. Encrypted link
  3. Private instance
  4. GPU pool
  5. Response + trace

Completion, usage record, and audit trace return the same way.

Compare

Three isolation levels, one operator

Shared catalogDedicated instanceAir-gapped
GPU allocationPooledExclusiveExclusive
Your own weights—✓✓
NetworkPublic endpointVPC peeringNo egress
Data residencyEUEU, region pinnedYour premises
Ops burdenNoneNoneShared with your team
Typical lead timeMinutes2 weeks4–6 weeks

All three use the same OpenAI-compatible API, so moving up a tier does not change application code.

Hardware

Capacity we actually operate

NVIDIA H200 · 141 GB

Frontier reasoning and long-context work

Cluster

NVIDIA H100 · 80 GB

Large models at production throughput

Instance

NVIDIA A100 · 80 GB

Established models, best price per token

Instance

NVIDIA L40S · 48 GB

Vision, embeddings, and mid-size models

Instance

Availability varies by region and is confirmed at scoping. If a specific GPU is required for contractual reasons, say so early.

Security

The controls that get audited

  • Single-tenant GPU allocation — no shared batch queues, no neighbour traffic.
  • Encryption in transit (TLS 1.3) and at rest (AES-256), keys held in your KMS where required.
  • Network isolation: private VPC peering, IP allow-listing, optional zero-egress configuration.
  • Role-based access with SSO, plus an audit log of every administrative action.
  • Data processing agreement, sub-processor list, and AI Act documentation on request.
  • Prompt and completion retention configured per contract — including immediate deletion.

Honest note

We are a young company. We can show you the architecture, the contracts, and the access controls today; formal certifications such as ISO 27001 are on our roadmap, not in our current claims.

Onboarding

Four weeks, no surprises

  1. Week 1

    Scoping

    Models, throughput, latency target, residency and retention requirements, network topology.

  2. Week 2

    Provisioning

    GPU allocation, private networking, key management, and the observability stack.

  3. Week 3

    Validation

    Load test against your real prompt mix; we tune batching and quantisation to the target.

  4. Week 4

    Handover

    Runbook, dashboards, escalation path, and a documented rollback plan.

Pricing

Predictable pricing, three shapes

Dedicated instance

from €1,850/mo

One GPU instance for steady production workloads.

Managed cluster

Custom

Multi-node, high availability, fine-tuning pipeline included.

Air-gapped

Custom

Isolated VPC or on-premises deployment with no egress.

All prices exclude VAT. Dedicated capacity is billed monthly with a 30-day notice period. GPU model and region are fixed per contract; usage-based overage is agreed in advance.

FAQ

Is the hardware really exclusive to us?

Yes. The GPUs are allocated to your deployment and are not used to serve other tenants. You can see the allocation in the console, and it is a term of the contract.

Can you deploy into our own data centre?

Yes, for the air-gapped tier. We ship the inference stack and the gateway as containers, install them with your team, and hand over the runbook. Support continues remotely under the same SLA.

What does the AI Act require from us?

It depends on your role and use case. As an infrastructure provider we supply the technical documentation, logging, and data-handling evidence; you remain responsible for your own risk classification and human-oversight arrangements.

How fast can we start?

A single instance typically goes live within two weeks of signature. Cluster and air-gapped deployments depend on network changes on your side; four to six weeks is realistic.

What happens to our data if we leave?

We export logs and configuration, then delete everything within the period agreed in the DPA. There is no exit fee and no data-retention hostage.

Let's design your private AI stack.

Bring your compliance and performance requirements — we will map them to a concrete deployment.

Request a quote