Deneb
ask about AI that stays inside your own building.
Two things it does. If you are deciding, ask whether private AI fits your business, what it costs and what the rules require. If you are building, it walks you through the setup step by step and reads your errors. It answers from what Altronis has actually done, and it hands you to a person rather than guess.
Loading…
What it does
Someone who has done this before, on both sides of the decision.
Straight answers before you commit
Ask whether this suits a firm your size, what it costs to own, and what changes if your data cannot leave the building. No sales call needed to find out.
It shows its working
Answers come from what Altronis has actually built and measured, and it quotes the source so you can check it. It says plainly when it does not know.
A person stays on the hook
For anything unusual or risky it hands you to an Altronis engineer instead of guessing, and it will never suggest loosening your security to make something work.
Free to try, no sign-in, a few questions a day. Want to go further, or talk to a person? Talk to us.
What the boxes cost today
Singapore prices, read every morning
- AMD Ryzen AI Max+ 395 mini PC, 128 GBS$4,325 to S$5,094
- NVIDIA GB10 box, 128 GB (DGX Spark or ASUS Ascent GX10)S$6,624 to S$8,096
- Apple Mac Studio, M5 Ultra (96 GB base)S$7,999 to S$9,949
- NVIDIA RTX PRO 6000 Blackwell 96 GB (card only)S$28,611
Read 9 Oct 2026 from each seller's own page. Overseas stores converted to SGD with 9% GST at import added, shipping not included. List prices on the day, not a quote. Sellers and links on our private LLM page, measured speeds on TokenMark. The raw feed is public: landscape.json.
Setting up your own box?
Deneb in your terminal
On the box itself, Deneb reads your hardware, recommends a model and prints the setup step by step, then troubleshoots what breaks. The read-only commands work offline with no sign-in and run nothing. When it can fix a fault itself, it asks first and never uses sudo or deletes a file. Open source, MIT licence.
Common questions
Private on-prem AI, in plain terms
What does private, on-premise AI cost?
An entry box with 128GB of unified memory that runs a capable model for a team costs S$4,325 to S$8,096 one-time, as read from Singapore sellers' own pages on 9 Oct 2026 (an AMD Ryzen AI Max+ 395 mini PC at the lower end, an NVIDIA GB10 box such as the DGX Spark higher), plus about S$30 to S$45 a month in electricity. There are no per-seat or per-token fees. Bigger models or heavy concurrency cost more; we size the exact figure with you.
How many people can one AI box serve at once?
A single box comfortably handles a small team's day-to-day use, a handful of people chatting or querying documents at the same time, plus a good volume of automated calls. Because generation speed is shared across concurrent users, dozens of simultaneous users all expecting instant replies needs batching and more hardware. We size it to your real number of users.
Is an on-premise box slower than cloud AI?
For the tasks most firms run, chat, drafting, and question-answering over your own documents, a single modern box is fast enough to feel instant. Speed is capped by memory bandwidth, not by how much memory you buy, so a larger model runs slower on the same box. Very large models are where cloud or bigger hardware wins.
Does my data really stay private with on-premise AI?
Yes. The model runs on your own hardware, in your own building, so your data never leaves the premises. The endpoint is API-key authenticated behind a secure tunnel with no open ports.
On-premise or cloud AI, which is cheaper?
Cloud wins for occasional or unpredictable use and for very large models you would rarely run. On-premise wins for steady daily use and for data that cannot leave the building. Once a monthly cloud or per-seat AI bill would exceed the cost of owning a box, owning usually pays back within months.
What hardware do I need to run private AI?
For most SMEs a single workstation-class box with 128GB of unified memory (an AMD Ryzen AI Max box or an NVIDIA DGX Spark) runs a capable model for a whole team. We run a 125B mixture-of-experts model with vision on one such box every day. The two are within about 7% on memory bandwidth, so they generate text at broadly similar speeds; the difference is prompt-processing and software ecosystem.
Can Deneb help me set up my own box?
Yes. Deneb also runs as a terminal agent on the box you are setting up (NVIDIA DGX Spark, AMD Strix Halo and similar). Its four read-only commands, check, profile, recommend and setup, work offline with no sign-in: setup prints every command for your box and runs none of them. When it has found the cause of a fault it can apply the fix, asking before each change by default, and it never uses sudo, deletes files or touches your firewall. It is open source under the MIT licence.