OverviewSolutions & TiersForge

Mid Enterprise.

20–200 users. Centralized AI server. Software pods.

Architecture


              AI  Center  Server
  ┌────────┐  ┌────────┐  ┌────────┐  ┌────────┐
  │ A6000  │  │ A6000  │  │ A6000  │  │ A6000  │
  │ 48 GB  │  │ 48 GB  │  │ 48 GB  │  │ 48 GB  │
  └────────┘  └────────┘  └────────┘  └────────┘
         Unified GPU Pool: 192 GB VRAM

     SteraOS  AI  Center
     ┌───────────┐ ┌───────────┐ ┌───────────┐
     │  Model    │ │ Inference │ │   Pod     │
     │  Depot    │ │  Engine   │ │  Roster   │
     └───────────┘ └───────────┘ └───────────┘
     ┌───────────┐ ┌───────────┐ ┌───────────┐
     │  Joules   │ │ Training  │ │   RBAC    │
     │  Economy  │ │  Engine   │ │  + SSO    │
     └───────────┘ └───────────┘ └───────────┘
                        │
         ┌──────────────┼──────────────┐
         │              │              │
  ┌──────┴───┐   ┌──────┴───┐   ┌─────┴────┐
  │ Terminal │   │ Terminal │   │  Mobile  │
  │  (desk)  │   │  (desk)  │   │  (field) │
  │  €99-149 │   │  €99-149 │   │   pad    │
  └──────────┘   └──────────┘   └──────────┘
         │              │              │
  ┌──────┴───┐   ┌──────┴───┐   ┌─────┴────┐
  │ Finance  │   │ Legal    │   │ Support  │
  │ Dept.    │   │ Dept.    │   │ Dept.    │
  └──────────┘   └──────────┘   └──────────┘
            

The fundamental shift

In the sole & small tier, each machine has its own GPU and runs its own models. This works for individuals and small teams. But at 20+ users, a different architecture becomes more efficient: one powerful centralized server with multiple high-end GPUs, running all AI workloads for the organization.

The pod becomes software. Instead of buying a hardware box for each AI role, you deploy a software pod — a model configuration with an occupation definition — onto the central server. A financial analyst pod, a legal researcher pod, a customer support pod. Each runs on the shared GPU pool. Adding a new AI capability means deploying a new pod, not purchasing new hardware.

Workers access the AI Center through thin terminals — silent, fanless devices at each desk that cost a fraction of a full machine. They have no GPU of their own. All computation happens on the server. The terminal is just an interface. Field workers use Stera Mobile pads. Remote workers connect through the web. The intelligence is centralized. The access is everywhere.

Software pods


  Pod  Roster

  ┌─────────────┐  ┌─────────────┐
  │ Developer   │  │ Analyst     │
  │ Pod         │  │ Pod         │
  │             │  │             │
  │ Model: 34B  │  │ Model: 14B  │
  │ VRAM: 22 GB │  │ VRAM: 10 GB │
  │ Dept: Eng.  │  │ Dept: Fin.  │
  └─────────────┘  └─────────────┘

  ┌─────────────┐  ┌─────────────┐
  │ Support     │  │ Researcher  │
  │ Pod         │  │ Pod         │
  │             │  │ Pod         │
  │ Model: 8B   │  │ Model: 72B  │
  │ VRAM: 6 GB  │  │ VRAM: 48 GB │
  │ Dept: Supp. │  │ Dept: R&D   │
  └─────────────┘  └─────────────┘

  Total VRAM allocated: 86 / 192 GB
  Pods active: 4
  Available for new pods: 106 GB
            

A software pod is a self-contained AI worker: a model, an occupation configuration, and a VRAM budget. It is deployed to the AI Center like a new hire is onboarded to a team. It has a role, a department, and a scope of responsibility.

Pods are co-developed with your organization. We work with your teams to understand the work, select or fine-tune the right model, define the occupation, and calibrate the pod for your specific operations. This is not off-the-shelf software. It is a purpose-built AI worker trained on your domain.

Joules economy


  Joules  Dashboard

  Engineering  ████████████████░░░  78%
  Finance      ████████░░░░░░░░░░  42%
  Legal        ██████░░░░░░░░░░░░  31%
  Support      ████████████░░░░░░  58%
  R&D          ██████████████████  95%

  Total compute this month: 847,000 Joules
  Budget remaining: 153,000 Joules
            

Joules is Stera’s internal compute currency. Every GPU-second consumed by a pod costs Joules. Departments get budgets. Managers see exactly where AI compute goes. There are no surprise cloud bills — just predictable, transparent, internal cost accounting.

The CFO sees a dashboard, not an invoice from a cloud provider. When a department needs more compute, you allocate more Joules — not more budget to a third-party API.

On-premise training

Enterprise customers fine-tune pods on their own proprietary data — entirely on-premise. Select a base model, upload your training data, configure parameters, and the GPU Pool allocates training compute. Your data never leaves the building. The trained model deploys directly to the Pod Roster.

A legal firm trains its contract analyst on 10 years of its own contracts. A hospital trains its medical scribe on its own clinical documentation style. A manufacturing company trains its quality inspector on its own defect library. The model becomes expert in your specific domain — not a generic approximation of it.

What you get


  AI Center Server
  ─────────────────────────────────────────
  Hardware        Multi-GPU rack server
  GPU Pool        4× A6000 48 GB = 192 GB (configurable)
  Operating Sys.  SteraOS AI Center
  Modules         GPU Pool · Model Depot · Inference Engine
                  Pod Roster · Joules Economy · Training Engine
  Security        RBAC · SSO · audit logging
  API             Enterprise API Gateway (HTTPS + WebSocket)
  Power           ~1,200W

  Per Desk
  ─────────────────────────────────────────
  Stera Terminal  Silent thin client, €99-149
  Stera Mobile    Voice-operated field pad

  Software
  ─────────────────────────────────────────
  Software Pods   Co-developed with your org
  SteraLink       Desktop / mobile app (free)
  Dashboard       Full AI Center management

  Pricing
  ─────────────────────────────────────────
  AI Center       from €15,000
  Terminals       €99-149 each
  Monthly fees    €0
  Per-token costs €0
            

Scale up or start here

Need multi-site federation with data sovereignty per region? That’s the large enterprise tier. Starting smaller? Individual machines work just as well.