Enterprise AI That Never Leaves Your Building
A complete appliance that reads your documents, reasons over them locally, and writes the result into the systems you already run. Open-weight models run on the appliance, and no data leaves your network unless you connect your own cloud model account for a task your policy allows.
4U rack or tower · Document OCR and language reasoning on separate accelerators · From $28,700
THE PROBLEM
Why Your Team Is Pasting Contracts Into A Public Chatbot
Recorded on a working deployment.
01
The work is faster with AI
Summarising a 300-page tender, extracting terms from a contract, drafting a purchase order from an invoice, these take an afternoon by hand and four minutes with a language model. Your people know that.
02
So they use one anyway
Unredacted contracts, patient records, and financial statements go into public chatbots because the alternative is doing the work manually. IT policy does not change that arithmetic; it just moves the behaviour out of sight.
03
Chat boxes do not produce enterprise output
A knowledge worker does not want to prompt an empty text box eleven times to get a structured purchase order. They want a form that already knows what it is asking for.
04
And building it yourself rarely works
GPU driver conflicts, VRAM sharding, container orchestration, vector indexing, retrieval tuning. Most in-house attempts stall before production because the specialist MLOps skills are not on the team.
ARSA SOVEREIGN AS THE SOLUTION
What The On-Premise AI Appliance Does
01
Reads documents
Scanned PDFs, handwritten annotations, dense multi-page layouts, schematics. Full-page layout parsing on a dedicated accelerator so document work never competes with reasoning for memory.
02
Reasons locally
A 27B-class open-weight vision-language model running unquantised on the primary accelerator, with the context headroom to hold an entire tender in working memory.
03
Builds the interface it needs
State a goal in plain language and the system generates the form, table, and validation steps to complete it. Not a chat box, a wizard built for that specific task.
04
Writes into your systems
Connectors for Odoo, SAP, and Zoho. Query inventory, draft orders, update records, with an approval gate before anything commits.
05
Works while you are not there
Multi-hour batch jobs run in the background. Start a thousand-document audit at 7pm, get a structured report and a notification when it finishes.
06
Runs entirely on premises
No outbound connection required: inference, documents and models stay on site, and the appliance can be air-gapped. Connecting your own cloud model account is optional.
INNOVATIVE UI
Forms, Not Chat Boxes
Most enterprise AI fails at the interface. A blank prompt box asks the user to know what to ask. ARSA Sovereign generates the interface for the task instead.
01
State the goal
“Audit this vendor contract against our standard NDA terms and draft the purchase order.”
02
The system builds the wizard form
A step-by-step wizard with the fields, tables, validation rules, and approval steps that task actually needs, generated for this job, not hardcoded.
03
Adjust in plain language
A persistent side panel lets you change any field by asking. “Change payment terms to Net-30” re-renders the affected step immediately.
INTEGRATIONS
Connected To The Systems You Already Run
Document intelligence is worth little if the result has to be retyped into an ERP. ARSA Sovereign connects directly, using the Model Context Protocol.
Bring your own cloud model. Open-weight models do the work on site by default. Where your policy allows a task to use a frontier model, Sovereign connects to your own Claude, Gemini or OpenAI account for that task only. ARSA never hosts or proxies that model, and the account, the contract and the data terms stay yours.
Odoo
Query inventory, draft sales orders, update customer records, generate invoices.
SAP
Contextual queries against ledger balances, supply chain status, and requisitions, respecting your existing role-based access controls.
Zoho
CRM, Books, and Inventory record updates driven directly from processed documents.
SYSTEMS AND PRICING
Two Appliances Or A Software-Only Licence
Both run the same Cognitive OS. The difference is model precision, document throughput, and how much context the system holds at once. Neither is sold as hardware, the price covers the appliance, the operating system, the models, the workflows, and the support that keeps them current.
Software-only licence. If you already run GPU servers, the same Cognitive OS, models and connectors install on your own hardware under a perpetual licence, with three years of support and model updates included and Care from year 4. Your servers need at least an 8-core CPU, 32 GB of RAM, a 256 GB SSD for the system and a GPU with 24 GB of VRAM; we recommend 1 TB of SSD or more for your documents and a GPU with 48 GB of VRAM or more.
ARSA Sovereign 48
For teams starting with document review and internal workflows
- Single AI accelerator, 48 GB
- Language model runs quantised
- Document OCR shares the accelerator
- 128 GB ECC system memory
- 8 TB enterprise storage
Start from $28,700
International list price, USD. Exact configuration is quoted. Excludes import duty, VAT, customs clearance, and rack installation.
ARSA Sovereign 96
For heavy document volumes and long-context work.
- Primary AI accelerator 96 GB, dedicated 24 GB OCR accelerator
- Language model runs unquantised at full precision
- Document processing never competes with reasoning for memory
- 128 GB ECC system memory
- 20 TB enterprise storage, RAID 1
Start from $29,690
International list price, USD. Exact configuration is quoted. Excludes import duty, VAT, customs clearance, and rack installation.
SECURITY AND CONTROL
Security Built For Audited And Air-Gapped Environments
The Platform
- Hardened Linux host with kernel livepatching: security updates without reboots
- Encrypted storage and encrypted traffic between services
- CIS and DISA-STIG hardening profiles applied at build
- Workloads isolated in containers with dedicated accelerator binding
- Encrypted fleet management channel, or fully air-gapped operation
Your Data
- No outbound connection required at inference time
- No document, query, or result reaches ARSA
- Vector index and document vault on encrypted local storage
- Every ERP write staged for human approval before commit
- Model updates delivered as signed offline packages for closed networks
Who This Is For
This is for
- Law firms and legal teams handling M&A, IP, and client-confidential documents
- Hospitals and clinics processing records under HIPAA, GDPR, or PDPL
- Banks and regulated financial institutions under data residency mandates
- Government and defence organisations requiring air-gapped operation
- Enterprises with cloud API spend that has stopped being an experiment
This is not for
- Teams wanting a frontier-scale model: this runs a 27B-class open-weight model, not a 400B one
- Workloads that are genuinely bursty and occasional; cloud is cheaper for those
- Organisations with no rack space, no power budget, and no appetite for on-premise hardware
- Workloads that must run entirely on a proprietary hosted model; Sovereign can call your own cloud account for permitted tasks, while its core runs open-weight models on site
SCOPE
What You Get
Included in the appliance price
- The appliance (Sovereign 48 or 96), pre-imaged and configured
- The Cognitive OS, the document and language models, and the workflow builder
- Odoo, SAP and Zoho connectors, with approval before every write
- Three years of support, model updates and fleet management
Purchased separately from ARSA
- Care from year 4, covering support, model updates and fleet management
- Connectors and integration work beyond Odoo, SAP and Zoho, quoted after scoping
- Sovereign Cluster, for more users or redundancy, quoted
Not included
- Rack space, power and cooling
- Rack installation
- Network design and structured cabling
- Licences for your ERP and other third-party software
- Import duty, VAT, customs clearance
Exact configuration is confirmed before order. Prices are on the Pricing page.
Questions
ARSA Sovereign Questions
Which model does it run?
A 27B-class open-weight vision-language model under a permissive licence, kept current under your support term. We do not name a fixed version in our documentation because the open-weight field moves faster than product literature does, the model in your appliance is the one that performs best for the workload at the time it ships, and it is updated as part of support.
Can it run completely air-gapped?
Yes. No outbound connection is required for inference. Model updates and licence activation are delivered as signed offline packages, transferred by your own approved media process.
What happens to documents we process?
They stay on the appliance. Text extraction, embedding, indexing, and reasoning all happen locally on encrypted storage. Nothing is transmitted to ARSA or to any third party.
Can an agent write to our ERP without approval?
No. Every connector is read-only by default and every write is staged for explicit human approval before commit. There is no autonomous write mode to enable.
How does this compare to a frontier cloud model?
A 27B open-weight model is not equivalent to a frontier-scale hosted model on the hardest reasoning tasks. It is very capable on the tasks this system is built for, document extraction, structured summarisation, form generation, and ERP workflow execution, and it runs where your data is allowed to be. If a task needs maximum raw capability and your policy allows its data to leave the building, Sovereign can send that task to your own Claude, Gemini or OpenAI account while everything else stays local.
Do we need a data centre?
No. The chassis converts between 4U rack mount and a standalone tower with acoustic damping, and it is designed to run in an office environment.
Who supports it?
ARSA, under a contracted support term included for three years with the appliance. Fleet management, model updates, and remote diagnostics are included; the management channel is encrypted and can be disabled entirely for air-gapped deployments.
How long is the lead time and how does it ship?
Three to four weeks from order. Shipped from Surabaya. Import duty, VAT, and customs clearance are the buyer’s responsibility.
Your Documents, Your Building, Your Hardware
Tell us the documents you process, the number of users and the systems the results must reach, and we confirm which configuration fits and quote it with the delivery terms. Documents stay on the appliance, it runs without an internet connection, and nothing you process reaches ARSA. When a workflow needs a custom build beyond the appliance, our private LLM and AI agent development team scopes it. ARSA already runs AI where data cannot leave the building, such as restricted-area security for the Ministry of Defense and other government and defence sites.
Related reading: stopping confidential documents from reaching public chatbots, on-premise contract review for law firms, on-premise AI for hospital patient records and air-gapped AI for defence and government.