Your machine, or a vendor’s cloud.
Your documents stay in your office
- Cloud AI
- No
- Your machine
- Yes
Questions and answers never reach a vendor
With an open model on the machine. A cloud model is yours to switch on.
- Cloud AI
- No
- Your machine
- Yes
No vendor staff can read your prompts
Cloud vendors allow staff and contractors access for support and abuse checks.
- Cloud AI
- Partly
- Your machine
- Yes
You decide how long anything is kept
Cloud APIs keep logs for 7 to 30 days. In 2025 a US court ordered one vendor to keep deleted chats.
- Cloud AI
- Partly
- Your machine
- Yes
Out of reach of foreign orders on the vendor
US vendors answer to the US CLOUD Act, even for data stored in Asia.
- Cloud AI
- No
- Your machine
- Yes
Your model is never retired or repriced
Popular models left one chat app in Feb 2026. One model’s input price doubled at launch in Jun 2025.
- Cloud AI
- No
- Your machine
- Yes
No charge per question or per GB out
Data out of a cloud region in Asia costs about USD 0.12 a GB.
- Cloud AI
- No
- Your machine
- Yes
Keeps working when a cloud goes down
A cloud region failure on 20 Oct 2025 took popular AI apps down for about 15 hours.
- Cloud AI
- No
- Your machine
- Yes
The largest frontier models
An open model runs on the machine. A cloud model is one setting away.
- Cloud AI
- Yes
- Your machine
- Partly
More capacity in minutes
- Cloud AI
- Yes
- Your machine
- No
No hardware to house or power
- Cloud AI
- Yes
- Your machine
- No
Failover across data centres
One machine. If it stops, you restore from backup.
- Cloud AI
- Yes
- Your machine
- No
Cloud AI facts are dated and public. They describe the category, not any one firm’s contract.
What the rules make easier.
Korea
A model inside your network needs no exception to the FSC’s network-separation rules. SaaS became a named exception on 20 Apr 2026, but not for unique identifiers or personal credit data, and generative AI is not yet one.
FSC network separation · Apr 2026
Hong Kong
Records kept on your own premises sit outside the SFC’s rules for external data storage, which ask for an undertaking from the storage provider. The 2024 AI circular still applies: human review and testing for research.
SFC circulars · Oct 2019 and Nov 2024
Singapore
Running the model yourself removes the outsourcing due diligence and exit plan for an AI vendor. Technology risk management still applies: recovery targets, patching and access.
MAS outsourcing guidelines and TRM · Jan 2021
Choose your hardware.
- NVIDIA DGX Spark
- A desktop machine with 128 GB of memory and a 240 W supply. In a client’s office, Qwen3.8 27B starts answering in 0.17 s and writes 8.5 tokens a second for one person, 62 across eight
- Your NVIDIA GPU server
- An x86 server your IT team already runs. Core installs with its GPU profile, sized with us
- Your own model server
- Any server that speaks the OpenAI API, such as vLLM or llama.cpp, on your network
- A cloud model, if you choose
- Claude, GPT or Gemini with your key, for work that needs a frontier model. Off until you switch it on
Hardware is quoted in your proposal, or Core runs on a machine you already own.
What runs on the machine.
The model
Qwen3.8 27B on vLLM for answers, and bge-m3 for search, both on the GPU.
Search and the map
A vector index, a keyword index and a graph database for the map of your documents.
Core
The app, its workers and the web page, from one release bundle with a bill of materials and a scan per image.
The record
Your documents, every answer and the answer log, in a database on the machine’s disk.
Everything that can leave the machine.
- A cloud model
- Your question and the passages that answer it, to Anthropic, OpenAI or Google
- Virtual analyst
- Its task and browser screenshots, to Anthropic
- Prices and exchange rates
- Ticker symbols and currency codes, four times a day
- Market data add-on
- Requests for movers, claims and the calendar, to AlphaLenz under your key
- Filings
- Company identifiers, to SEC EDGAR
- Microsoft 365 and Google
- A read-only sync, if you connect them
- Our relay
- Encrypted traffic from phones and your PMS vendor, passed through unread. Leave it out and nothing reaches Core from outside
- Updates
- A check for new releases, unless you upgrade from a disk
Security and compliance lists every connection, with how to turn it off.
Install, upgrade and roll back.
- 01
One bundle
Images with a signed checksum list, a bill of materials and a vulnerability scan each.
- 02
Checked upgrades
Each upgrade is verified against its signature, and an older version is refused.
- 03
Roll back
The previous release stays on the machine. One command returns to it.
- 04
From a disk
No connection needed: upgrade from a bundle we send.
What we set up for you.
Sign-in for each person
Personal accounts with two-step sign-in, and every sign-in recorded.
Scheduled backups
The library, the map and the answer log, backed up to a disk or share you choose.
Your own web address
HTTPS on your firm’s name, with the certificate renewed for you.
Health checks
A health check and a resource view, run after every upgrade.
What you take on.
- One machine
- No failover. If it stops, Core stops until it restarts or is restored from backup
- Power and space
- A DGX Spark sits on a desk. A GPU server needs a rack, power and cooling
- Model size
- A DGX Spark runs a model of about 30B parameters, or a larger mixture-of-experts model at 4-bit. The largest models need a cloud model or a bigger server
- Speed for one person
- About 8 tokens a second on the 27B model: slower than a cloud model for one long answer, and steady for a desk at once
- Your IT team
- A fixed address on your network, a DNS name and someone on the day we install. Upgrades are planned with you
Questions.
What do we need to provide?
A machine, or room for one we supply, a network port, a fixed address on your network, and someone from IT on the day we install.
Can it run with no internet at all?
It is built to: an open model on the machine, and upgrades from a disk. Features that reach the web, such as prices, filings and the virtual analyst, stay off.
How fast is it?
On a DGX Spark in a client’s office, Qwen3.8 27B starts answering in 0.17 s and writes about 8.5 tokens a second for one person, 62 across eight people at once.
Can we still use a cloud model?
Yes, if you choose. Switch on Claude, GPT or Gemini with your key for the work that needs it. The rest stays on the machine.
Who keeps it running?
We do, as part of your partnership: upgrades, model changes and support from engineers who know your setup.
