Pillar guide

Enterprise AI on-premise.
Architecture, deployment, procurement.

Production AI for enterprise environments where the cloud isn't an option. Air-gapped, on-premise, private cloud, or Microsoft Foundry — same code, same engineers, deployed inside your perimeter.

On this page

What is enterprise AI on-premise?

Enterprise AI on-premise is a deployment model where the AI workload — model, runtime, application code, and audit log — runs inside the customer's own infrastructure. The customer controls the network perimeter, the data residency, the audit surface, and the supply-chain dependencies. The vendor ships software (and sometimes hardware) into that environment, but doesn't operate the live system from a remote cloud.

In 2026, the on-premise question is no longer "can it be done?" — it can. The question is "is your AI vendor designed for it, or are you stretching a cloud-first product to live somewhere it doesn't belong?" Most of the AI-vendor market is the second case. Wavenetic is the first.

Four deployment models — and when each is right

Real enterprise environments don't pick one deployment model and run everything there. They have a mix: some workloads can run on a public-cloud SaaS, some have to run inside a private VPC, some have to run on hardware in the building, some are classified and have to live in a sealed enclosure. A serious AI vendor handles all four.

1. Microsoft Foundry / public-cloud SaaS

The lightest deployment surface. Used when the customer already runs on Microsoft and the workload doesn't touch high-sensitivity data. Procurement is fast (existing ELA agreements), governance is by Microsoft's own EU compliance frame, and the customer doesn't have to operate any infrastructure. Wavenetic deploys the Wave product stack on Microsoft Foundry as a Microsoft partner.

2. Private cloud / customer-controlled VPC

For customers with existing AWS, Azure, or GCP commitments who need tighter governance than public cloud allows. The Wave stack runs inside the customer's cloud account, in an EU region, with controlled inbound and outbound traffic, customer-managed identity, and customer-controlled audit pipeline.

3. On-premise — customer hardware

The customer owns the servers; Wavenetic ships software. Standard for regulated enterprise (banks, insurers, healthcare) and TSO-class operators. The customer's IT team operates the cluster; Wavenetic supports the runtime and the application layer. WaveNode appliances are an option here — recommended for performance-bounded workloads and air-gap-capable deployments.

4. Air-gapped — sealed appliance, zero outbound network

The strictest deployment. WaveNode hardware ships with signed firmware, attested boot, and offline-only update channels. No phone-home, no telemetry, no CDN model pulls, no license-server check. Standard for defence, classified environments, and parts of critical infrastructure that cannot be assumed to have outbound connectivity. The customer audits the box before deployment and at any point afterwards.

Reference architecture

Every Wavenetic deployment, regardless of which of the four models above it lands in, has the same logical layers. The deployment target changes; the architecture doesn't.

Layer 1 — Hardware

Either customer-provided servers (with documented requirements: GPU type, RAM, network, storage), or Wavenetic-shipped WaveNode appliances (S, M, L, or clustered). For air-gap deployments, WaveNode is the default — it's the only path that meets the trust requirements out of the box.

Layer 2 — Wave Runtime

The orchestration layer that runs models, agents, and tools. Inspectable end-to-end. No black boxes. Open-source orchestrator (WaveCode) available on GitHub for the agent layer; the runtime is governed by an enterprise license.

Layer 3 — Wave product applications

WaveOps (notebook workspace — chat, cited briefings, FAQs, audio overviews; one switch flips the workspace between cloud LLMs and on-prem Ollama), WaveFlow (finance / accounting AI), WaveStorm (marketing AI), and any custom-built application on the same runtime. The same product code that runs on the public web runs inside the customer's perimeter.

Layer 4 — Identity + billing

WaveID for single sign-on across products, WaveCredits for unified credit-based billing. Both launching on May 15, 2026. In on-premise deployments, both are deployable inside the customer's perimeter.

Layer 5 — Audit + compliance

Audit logs are not a feature — they are the storage substrate every other feature is built on. Every event (query, response, agent decision, tool call, error) lands in the audit store, which is then accessible to the customer's compliance tooling. Architecture documentation, conformity-assessment files, and human-oversight gates ship with the deployment.

Compliance frame — what regulated enterprise actually has to satisfy

Four regulatory frames matter for on-premise enterprise AI in the EU in 2026:

Procurement model

Two procurement paths, both designed for enterprise contracts:

Subscription (SaaS-like, but on your hardware)

Annual licence for the Wave product stack, plus on-call engineering support. Hardware optional. Best for customers who want continuous platform improvement and managed delivery, without operating the cluster themselves.

One-time purchase (you own everything)

Wave product licence and WaveNode hardware delivered as a finished appliance. Annual support contract optional. Best for fully air-gapped environments and customers who require asset ownership end-to-end.

Both paths come with full architecture documentation, conformity-assessment-ready files, and named-engineer support contacts. Talk to procurement for a quote.

Frequently asked questions

What is enterprise AI on-premise?
Enterprise AI on-premise means deploying AI models, runtime, and applications inside the customer's own infrastructure — physical servers, private datacenter, or sealed appliance — rather than relying on a third-party cloud API. The customer controls the network perimeter, the data residency, and the audit surface.
How is on-premise AI different from "private cloud" AI?
Private cloud AI typically means the AI workload runs in a customer-controlled VPC on a public cloud (AWS, Azure, GCP). The data stays inside the customer's account, but the underlying infrastructure is still operated by a non-EU hyperscaler. True on-premise AI runs on hardware physically located in a building the customer controls — datacenter, branch office, or operational site.
Why deploy AI on-premise instead of using a cloud API?
Three drivers: (1) Compliance — the EU AI Act classifies AI in critical infrastructure, defence, and law enforcement as high-risk, requiring conformity assessments, audit logs, and human oversight that cloud APIs do not natively provide. (2) Data sensitivity — many enterprise workloads (banking, healthcare, defence, regulated infrastructure) cannot legally or commercially send data to a non-EU vendor. (3) Operational independence — on-premise deployments do not depend on a vendor's uptime, pricing, or roadmap.
What is air-gapped AI?
Air-gapped AI is a deployment with zero outbound network connectivity. The AI runtime, model weights, and inference all happen inside the customer's perimeter, with no phone-home, no telemetry, no CDN model pulls, and no license-server check. Updates are delivered through signed offline artifacts. Air-gap is the default deployment for defence, classified, and many critical-infrastructure use cases.
Does Wavenetic ship hardware?
Yes. WaveNode is a sealed AI appliance with signed firmware and attested boot, available in three sizes (S / M / L) plus a clustered configuration. The customer audits the box before deployment. Updates are delivered through signed offline channels rather than the internet.
How does Wavenetic handle the EU AI Act?
The EU AI Act classifies AI used in critical infrastructure, defence, and law enforcement as high-risk, requiring conformity assessments, technical documentation, audit logs, and human-oversight gates. Wavenetic ships those artifacts with every deployment. The architecture documents, audit-log specifications, and conformity-assessment-ready files are part of the product, not a separate billable service.
Can the same code run on Microsoft Foundry as on a sealed appliance?
Yes. The Wave product stack (WaveOps, WaveFlow, WaveCode, Wave Runtime) deploys across the same code path on Microsoft Foundry, on customer-controlled private cloud, on customer on-premise hardware, and on a sealed WaveNode appliance. The customer chooses based on their compliance posture; the same engineers ship to any of the four targets.
How long does an on-premise deployment take?
A typical pilot with the Wave product stack — one process, one customer environment — goes from approval to live in 30 days. Hardware deployments using WaveNode appliances ship within four weeks of order. Full multi-site rollouts depend on the customer's integration scope; a typical regulated-enterprise rollout runs 90–180 days from contract.

Ready to deploy?

Try our AI products on the public web first. Then talk to us about deploying the same code inside your perimeter.