Eldric AI Datacenter — flagship product

Eldric AI Datacenter. The complete operating system for the AI datacenter you run yourself.

The chat, the agents, the retrieval, the inference. One platform.
A single operating system across HPE, Eviden, Dell, NVIDIA, mixed hardware. Your rack, your choice.
On-prem-first. EU-built. No mandatory cloud calls. Yours to run.

Eldric AI Datacenter — three bands: customer surfaces above, the Eldric kernel and modules in the centre, hardware below band · customer surfaces tenant slices agent stack plugin market band · eldric ELDRIC kernel & modules controller router inference data edge agent swarm science media comm iot one platform · one configuration grammar band · your hardware GPU accel. CPU x86 · ARM NET storage Eldric AI OS running the complete AI datacenter operating system
§1 — The shape of an AI datacenter

Three bands. One operating system.

Shipping today

An AI datacenter is not just GPUs in a rack. It is the operating layer that turns compute, storage, and network into an AI service — chat, agents, retrieval, training, observability, multi-tenancy, billing, audit, and a plugin host — all under one configuration grammar. Eldric is that layer. It does not replace your hypervisor, your storage array, your firewall, or your identity provider. It does provide every layer those tools cannot.

What this page is. The flagship product framing: Eldric as the complete operating system for the AI datacenter you run yourself — the full datacenter story, from cluster modes through tenancy, plugins, install, and where it fits. For audience-segmented landings see For service providers, For enterprises, For research institutions.
Three-band stack — customer surfaces at the top, the Eldric kernel band in the centre, hardware at the bottom band · customer-facing surfaces chat shell /chat tenants isolation agents 15 types plugins marketplace API OpenAI-compat band · eldric kernel & modules ELDRIC one operating system, all the modules controller router inference data agent edge swarm science media comm training iot one package · one configuration grammar · one upgrade path band · the hardware you bought GPU accelerator CPU x86 · ARM RAM memory storage local · NFS · S3 network VLAN · WAN the AI OS shipping today
Plate 01 The complete AI tier. Above any conventional datacenter stack. Below the customer surfaces that turn it into a service.

§2 — What Eldric runs

One platform. All the pieces.

Shipping today

Each Eldric module is a named capability with a clear job. Compose the ones you need; skip the ones you don't. They all ship in the same package. They all speak the same configuration grammar. They all upgrade together.

controller

The cluster brain.

Orchestrates the fleet. Pushes topology, distributes the license, handles register and heartbeat, runs the rolling-upgrade orchestrator. See clustering & HA.

router

Smart load balancing.

Intent classification, AI-aware routing, tool-required substitution. Sends each request to a worker that can actually answer it. How it works.

inference

Models on your hardware.

CPU and CUDA workers. Native GGUF loading without external runtimes. Structured-ML xLSTM workloads alongside the conventional stack. All features.

data & RAG

Storage and retrieval.

File storage with multi-tenant isolation. Vector retrieval. Matrix memory. Replication between data nodes. NFS for shared filesystems. Using RAG.

agent

Agentic reasoning.

Multi-tenant agent runtime. Session memory, query decomposition, multi-agent execution, named workflows. Advanced retrieval.

edge

The public face.

TLS termination, authentication, rate limiting, the embedded chat shell, plugin host. Customers see this; nothing behind it. For developers.

swarm

Multi-agent orchestration.

Topologies for fanned-out reasoning: hierarchical, peer-to-peer, ring, star, mesh, hybrid. For deeper agent collaboration than one chat turn permits.

training

Fine-tuning & federation.

LoRA, QLoRA, DPO, RLHF. Training chains with latent-reasoning techniques. Federated rounds across the worker fleet without moving the data.

science

A research-data surface.

One registry across 16 categories of scientific sources — astronomy, particle physics, genomics, climate, medical, materials, more. Pluggable.

comm

Messaging & calls.

Email, SMS, WhatsApp, Signal, Teams, XMPP, voice over SIP. A unified message surface; AI-assisted responses with operator approval.

media

Audio & video.

Speech-to-text, text-to-speech, transcription, voice chat, multimedia retrieval. Local engines and cloud APIs, side by side.

iot

Things & sensors.

Industrial OPC-UA, Modbus, MQTT. Consumer HomeKit, Matter, Netatmo. Predictive maintenance from inference; closed-loop control via structured-ML policy.

Every module exposes a documented API. Every module participates in the same multi-tenant guard. Every module emits OpenTelemetry. The boundary between "platform" and "infrastructure" stays clean: we own the AI surface, you own the metal.


For the operator, this is three jobs the platform makes possible — and one job the platform replaces.


§3 — For the datacenter operator

Sell AI services. Run your customers' inference. Federate across sites.

Shipping today

A datacenter that runs Eldric is not a datacenter that runs models. It is a datacenter that runs an AI service — a thing customers buy, with isolation, billing, branding, and operator control. Three jobs the platform makes possible.

job · 01

Sell AI services.

Multi-tenant by design. Per-tenant isolation enforced by the kernel — data, sessions, retrieval, and audit all stay inside their own boundary. Per-tenant theming and branding so each customer sees their own product. License tiers built for resale: the operator buys a tier; tenants get a slice.

See the resale model →
job · 02

Run their inference.

Any model, any backend. Ollama, vLLM, TGI, NVIDIA Triton, llama.cpp, native GGUF — local. OpenAI, Anthropic, xAI, DeepSeek, Groq — cloud. All under one unified abstraction. Customers ask for a model; the router finds it; the worker serves it. The operator sees one billing surface.

See the backend list →
job · 03

Federate across sites.

One HQ datacenter and ten branch sites. Vienna and Salzburg. Primary and warm-DR. A peer mesh that gossips state, routes around outages, keeps regional data resident where the customer's contract says it must stay. Cross-WAN, cross-AZ, cross-subnet.

See clustering & HA →

The three jobs compose. Sell to enterprises whose data must stay on-prem. Run their inference on the GPUs you already own. Federate the service across the racks you operate — and bill all of it from one place.


§4 — What you don't have to build

The substrate, not the application.

Twelve to eighteen months of platform engineering, before the first customer signs up. The pieces a serious AI service needs that nobody actually wants to write from scratch — the substrate beneath the product, the layer that has to be done well or nothing else works. Eldric ships that substrate. Your team builds the product on top.

Honest framing. The list is not "things you cannot build" — every item on it has been built before by someone. It is "things you will spend the first year of the company building, instead of the product the customer is paying for".

Build a chat application. Build a research agent. Build a private AI service for your industry. We will not build the application for you — but we will spare you a year of substrate work to get there.


§5 — Hardware-agnostic

We don't sell hardware. We make whatever you bought sell AI.

An AI platform that demands a specific vendor's hardware is the wrong shape for the datacenter. Eldric runs on Linux on x86 or ARM, with or without GPUs, on whatever your team already procured. We are not trying to lock you to a chipset, a server vendor, or a network fabric. We are trying to give you a platform that turns those things into an AI service.

What we ask of the rack. A modern Linux distribution. Network reachability between cluster nodes. Storage you can mount. Optional GPUs for serious local inference. Nothing else.
HPE — ProLiant, Apollo, GreenLake Eviden — BullSequana, Smart Management Center NVIDIA — DGX, HGX, RTX, Grace Dell — PowerEdge, PowerScale Supermicro — SuperServer, SuperBlade Lenovo — ThinkSystem, ThinkAgile AMD — EPYC CPU, Instinct GPU Intel — Xeon, Gaudi accelerators Arm — Neoverse, Ampere Mixed — whatever the rack already holds

A note on Eviden and the European AI-factory posture: BullSequana XH3000 + Smart Management Center is an exceptionally clean fit for the platform — the management plane is Redfish-clean, the supercomputer-class density is real, and the Atos partnership posture aligns with how Eldric ships in Europe. The same platform runs on top of HPE's GreenLake estate, NVIDIA reference architecture, Dell PowerEdge, mixed Supermicro builds, and the rack of whatever you already bought. We sell the AI layer above the iron, not the iron.

If your hardware runs Linux and reaches the network, it runs Eldric. The platform's job is to make the rack productive, not to dictate what the rack contains.


§6 — Sovereignty & EU compliance

On-prem-first. No mandatory cloud calls. Data residency by design.

Shipping today

Sovereignty is not a marketing badge; it is an architectural property. Eldric runs entirely on the operator's hardware. There is no phone-home requirement. Licenses verify offline. Telemetry is opt-in. Cloud LLM backends are an option you choose, not a dependency you accept. Data residency is enforced by the federation layer: a tenant's region pins where their data lives, full stop. The AI plane is fully EU-built — kernel, modules, license server, vendor — so residency is a property of the architecture, not of a policy document.

Built in Vienna. [core] Informationstechnologie, Vienna — the EU-based vendor behind the platform. EU-resident license server. No CLOUD Act exposure on the AI plane. No US-jurisdiction supply chain in the kernel.

Honest framing: where formal certifications are appropriate for your sector (ISO 27001, BSI C5, sectoral schemes), those are customer-side audit work — the architecture is built to pass, the paperwork follows deployment.
On-prem-first today EU-built Offline license verification Air-gap installs Hash-chained audit Per-tenant regional residency

For organisations that legally cannot ship their data across borders — healthcare under national rules, finance under regulatory mandate, defence under operational rules, public administration under sovereignty policy — the platform's posture is the platform's product. Eldric runs where your contract says it must.


§7 — Cluster modes

The same architecture, four scales.

Available today

Datacenter buyers do not deploy a "small cluster" the way they deploy a "big cluster" — the operational realities differ. Eldric uses the same software in every mode and changes only what is asked of it. A pilot starts on a single rack. The production rollout adds quorum and HA. The next office adds federation. None of the steps require a re-platform.

Honest scope. Automatic controller failover (Raft) is code-complete and final-verifying in an upcoming 5.0.x patch. Today's 5.0 ships clustered workers, gossip discovery, and rapid operator-driven controller swap.
single rack
Compact

One to three U of compute. Pilot deployment, single team, single tenant. Available today.

3-node HA
Quorum

Three nodes give you any-single-node tolerance with a clean majority vote. The most common production posture.

10–50 nodes
Mid-cluster

Workers split across racks or availability zones. Leader stays in a different zone. Rack-level failure is survivable.

50+ federated
Multi-site

HQ plus branches plus warm DR. Per-site autonomy; federation layer keeps directories and policy in sync.

Cluster mode scales — single rack, three-node HA, ten-node mid-cluster, multi-site federation single rack pilot 1–3 nodes 3-node ha quorum 3 nodes · any-1 tolerance mid-cluster multi-rack rack A rack B 10–50 nodes · zone-aware federated multi-site HQ DR 50+ nodes · cross-site one binary · node count is configuration
Fig. 02 Four common postures, one binary. Compact rack pilots fit on a single chassis; three-node HA is the production sweet spot; mid-cluster spreads work across racks for zone tolerance; federation joins sites without merging them into one giant cluster. See Clustering & HA for the full failover sequence.

The same RPM that runs on a developer's spare server runs on a fifty-node enterprise cluster. Scale is a configuration choice, not a procurement event.


§8 — Tenancy & licensing

One cluster. Many tenants. Clear quotas.

Today, with caveats below

An ISP offering "private AI" to enterprise customers, an MSP running shared infrastructure for several clients, or an internal IT department serving multiple business units — all need the same primitives: per-tenant capability tokens, quotas the cluster enforces, role-based access, and a license model that does not punish them for growing. Eldric ships all of this in one config schema.

Honest scope. Tenant isolation, capability tokens, and license tiers are available today. Full role-based access control is partial — Standard tier and above. Per-tenant matrix-memory privacy is in flight for an upcoming 5.0.x patch.
Tier Scale What it buys
Free 1 controller · 1 router · 2 workers Evaluation, proof-of-concept, small team. All daemons unlocked at a single-node scale; cluster features sized for pilots.
Standard 1 · 2 · 3 RAG, embeddings, agents, prompt database, custom models. Multi-tenant. The "we are using this for real" tier.
Professional 2 · 4 · 10 Dashboard, metrics, security extras, full RBAC, dual-controller. The mid-cluster production target.
Enterprise 5 · 10 · 50 High availability, orchestration, AI-decision routing, priority support, federation, named engineer.
Custom unlimited Source-code access, on-site deployment assistance, custom compliance certifications, air-gap installs.

Every license is Ed25519-signed and verifiable offline against the embedded public key. Workers fall back to the local license file when the controller is unreachable — air-gap deployments do not need a heartbeat to a server. See pricing for cost and for enterprises for procurement.


§9 — Plugin ecosystem

Plug in the rest of your datacenter.

Available today & expanding

A datacenter is not a greenfield. There is already a monitoring stack, a hardware management plane, a switch fabric, and a dozen other systems that need to keep working. Eldric's plugin host lets the LLM call any of them through the same tool surface — and lets the cluster operator add new ones without recompiling. SNMP for the legacy fleet. Redfish for modern servers. GreenLake and iLO for HPE estates. IoT and industrial protocols. Scientific data sources.

Five plugin types. Tool (server-side function), Filter (pre- or post-LLM message processing), Pipe (virtual model / custom backend), Action and Widget (client-side UI). Python on the server, JavaScript on the client.
SNMP · legacy fleet polling Redfish · modern server management HPE GreenLake · estate inventory HPE iLO · out-of-band control Network management · switches & routers OPC-UA · industrial PLCs & SCADA Modbus · legacy industrial MQTT Sparkplug B · industrial IoT Email · SMS · WhatsApp · Signal 140+ scientific APIs · genomics, pharma, climate Your own · server & client SDKs

Plugins are sandboxed Python or JavaScript packages with a manifest, optional valve configuration, and a clear lifecycle. The plugin host enforces rate limits, error budgets, and per-tenant access. The catalog ships with the platform; customer-specific plugins live alongside Eldric's own. The plugin ecosystem is actively expanding — every quarter brings new entries to the catalog.


§10 — Install & setup

From package to working cluster in an afternoon.

CLI available today

No build from source. No container orchestration to learn. No proprietary management plane to license separately. Eldric ships as a signed RPM, available from the Eldric package repository. Install on a Fedora, RHEL, Rocky, or Alma host, point the browser at the setup wizard, drop in your license, and you are running. Adding nodes is a one-line command on each one.

What ships when. The CLI eldric setup command is available today. The browser-based setup wizard with network-hint admin GUI ships fully built in an upcoming 5.0.x patch.

Add the package repository

curl -fsSL repo.eldric.ai/install.sh | sudo bash

A signed, GPG-fingerprint-pinned shell script. Adds the repo.eldric.ai dnf repository on Fedora 42+, RHEL 9+, Rocky, or Alma.

Today

Install the platform

sudo dnf install eldric-aios

One package brings every daemon. The unit file enables and starts the kernel; node roles are configured next.

Today

Run setup

sudo eldric setup

The CLI subcommand waits for local health, activates your license, and prints an install summary. Subsequent nodes join with a cluster secret.

CLI today · GUI in 5.0.x

Browse to chat

http://<host>:8880/chat

The single-node kernel listens on port 8880 by default — that's where the chat shell, controller dashboards, and the OpenAI-compatible API all live until you add a dedicated edge gateway in front (which is when the URL drops the port). First sign-up becomes admin. Configure tenants, models, knowledge bases. The cluster is in production.

Today

For air-gapped installs, the same RPM repository ships on a USB key. The setup flow is identical; only the source of the package changes. See the install guide for the full walkthrough and install troubleshooting for the edge cases.


§11 — Where this fits

Honest about the right deployment.

A platform that tries to be the right answer for everyone usually is not. Eldric is built for a specific kind of deployment: serious infrastructure teams who want a complete AI platform they fully control. That is a real audience, but it is not the only one. The right column below is just as honest as the left.

See how Eldric compares to Ollama, vLLM, NVIDIA NIM, OpenAI and the rest →

Right fit

If you own the rack, the case is strong.

  • Enterprises running AI workloads on their own infrastructure. The bigger the team, the bigger the savings against per-token cloud bills.
  • Regulated industries needing data residency. Healthcare, finance, defence, pharma. Data that legally cannot cross borders or leave the building.
  • ISPs and MSPs offering "private AI" to enterprise customers. Multi-tenant by design; license tiers that align with reselling.
  • Sovereign and national-security deployments. Air-gap installs. Verifiable offline licenses. No external dependencies for the AI plane.
  • Organisations replacing per-call cloud-LLM spend with on-prem inference. The platform that pays for itself when you switch from rental to capital.

Wrong fit

If infra is not your business, look elsewhere.

  • Pure consumer AI. If you want a chatbot for personal use and no infrastructure team, a hosted service is simpler.
  • Organisations without a dedicated infrastructure team. Eldric expects an operator who can run a Linux fleet. We do not pretend otherwise.
  • Deployments that need a hypervisor replacement. Eldric sits on top of your virtualisation layer; we do not replace VMware or KVM.
  • Teams that want a per-token API and nothing else. A cloud API will be cheaper at small scale and lighter to operate.
  • Production-grade hard real-time control. Hard real-time stays the PLC's job. Eldric covers soft-real-time and event-driven — and says so.

If you are in the right column, that is a useful answer too. We will tell you when a hosted service or a different tool is the better choice — including in the discovery conversation, before any contract.


§12 — Honest ledger

Today, in flight, on the way.

A three-column ledger of where the platform actually is. Anything not in the leftmost column is not yet running in production — we will not pretend otherwise. The thing we promise is that this table will tell you the truth.

Today

5.0 — shipping

  • One-package install on Fedora, RHEL, Rocky, Alma
  • CPU and CUDA inference workers
  • Multi-backend inference proxy (Ollama, vLLM, TGI, Triton, llama.cpp, OpenAI, Anthropic, xAI, DeepSeek, Groq)
  • Vector retrieval with multi-tenant namespaces
  • Matrix memory with crash-safe persistence
  • Agent runtime with multi-tenant sessions
  • Swarm orchestration topologies
  • Edge gateway with TLS, auth, rate limiting, chat shell
  • Per-tenant kernel isolation
  • Per-tenant theming & branding
  • Rolling RPM upgrades across the cluster
  • Multi-controller Raft cluster — forms, replicates, fails over
  • Multi-site federation surface — cross-branch routing, mTLS gossip, DR panel
  • Install wizard with network-hints GUI
  • Networking-plugin device-fleet GUI
  • Backup, restore, and verify of cluster state
  • OpenTelemetry export
  • Hash-chained audit ledger
  • Offline-verifiable Ed25519 license
  • Plugin marketplace
  • Science source registry (16 categories)

In flight

Next 5.0.x patches — designed & dispatched

  • Leader-aware client endpoint — completes the multi-controller HA story for production bootstrap
  • Three-node controller quorum bootstrap with auth-store unification
  • Per-tenant regional data-residency policy
  • Continuous replication of vector indexes
  • Continuous replication of matrix memory
  • Cluster device registry
  • Native xLSTM inference backend (preview)
  • Federated learning rounds across the worker fleet
  • Cross-tenant guard audit hardening

On the way

Later in 5.0.x — on the roadmap

  • Cross-WAN federation test lab + chaos-engineering harness
  • Warm disaster-recovery replicas
  • Cross-site session continuity for chat
  • Geographic load shaping (latency-aware routing)
  • Hardware-vendor management-plane plugins (HPE iLO, Eviden Smart Management, Dell iDRAC)
  • Browser DevToolbox for agent debugging
  • Tenant-level cost attribution & billing export
  • Long-context structured retrieval over xLSTM

Roadmap timelines change. Anything in the right two columns is genuinely planned and budgeted — but we will not promise a calendar quarter for software we have not finished.


§13 — Next step

Talk to us about the datacenter you want to run.

Eldric is small enough that you can talk to the people who build it. A short conversation will clarify whether the platform fits your deployment, whether the timeline matches your need, and what a pilot looks like. No funnel. No self-service checkout. A real conversation.

For deeper audience-segmented positioning see For service providers, For enterprises, or For research institutions. For the architecture detail see Clustering & HA and How it works.