Should a Hong Kong SME run its own private AI, or just pay a monthly subscription like everybody else?
It is the question that stops most AI projects at the second meeting, usually phrased as "but where does our data actually go". And it is almost never answered with numbers.
So here are the numbers, on both sides, in the currencies you will actually be billed in.
What are you actually choosing between?
You are choosing between renting a shared AI service and running one that only your company can reach. A cloud subscription bills monthly per user or per conversation. A private deployment puts the model on your own servers or private cloud, so your documents never leave an environment you control.
The distinction people care about is not technical. It is contractual. With a subscription, your data handling is governed by somebody else's terms of service. With a private deployment, it is governed by your own infrastructure.
That is a real difference, and it is worth money in some businesses. The mistake is assuming it is worth money in yours before checking.
Both options do the same visible job for most Hong Kong SMEs: let staff ask questions of company documents and get accurate answers instead of digging through a shared drive.
What does a private AI cost to run?
A production-ready self-hosted setup runs roughly US$250 a month at the low end, on a single consumer GPU, and about US$900 a month for a dual-GPU or small multi-GPU configuration. Frontier open-weight models need four to eight datacenter-class GPUs, which puts the monthly floor at US$2,000 to US$20,000.
That is only the hardware line. Industry cost breakdowns put ongoing maintenance at 10 to 20 hours a month of engineering time, at US$75 to US$150 an hour, which is another US$750 to US$3,000 a month in labour before anyone has asked the system a question.
The number that settles most arguments is the break-even point. Against frontier APIs, self-hosting on reserved cloud GPU starts to win somewhere around 2 to 5 million tokens a day. Below roughly 2 million tokens a day, the API is cheaper, because you are otherwise paying for idle hardware.
Two million tokens a day is about 1.5 million words a day. A 300-page business book is around 100,000 words. So the break-even is roughly fifteen books' worth of text, every day, through one system.
Almost no Hong Kong SME is anywhere near that. If your whole company asks an AI a few hundred questions a day, you are two orders of magnitude below the line where private hosting wins on cost.
There is a harder version of this. Compared against budget open-weight APIs at roughly US$0.14 to US$0.50 per million tokens, self-hosting almost never wins on cost alone. Sources including Cloudzy and Spheron reach the same conclusion from different directions.
So if a vendor tells you a private AI will save you money, ask them for your token volume first. If they have not asked, they are guessing.
What does the cloud option cost in Hong Kong dollars?
UD publishes its cloud pricing, which makes it a useful reference point whether or not you buy it. These are the current listed figures for AI Chatbot Master.
AI Chatbot Master, published plans
--- Startup: HK$298 per month. 1,000 conversations per month, 1,000-page knowledge base, one-off setup HK$1,000, 30-day chat history.
--- Standard: HK$498 per month. 10,000 conversations, 5,000 pages, one-off setup HK$3,000 waived on a yearly subscription, 90-day history.
--- Professional: HK$988 per month. 50,000 conversations, 20,000 pages, one-off setup HK$3,000 waived on a yearly subscription, 180-day history.
--- Enterprise: custom pricing, setup waived.
Terms that apply to every tier
--- Minimum six-month subscription.
--- One free month of service on a one-year subscription; two free months on a two-year subscription.
--- Website chatbot, WhatsApp and Facebook Messenger included on all tiers, with a real-time dashboard.
--- Over 90 languages, and ingestion of PDF, DOC, CSV, image and MP4 files.
Put that against the private numbers. The cheapest credible private deployment, at around US$250 a month plus maintenance, is already several times the HK$298 tier before anyone touches it.
One caveat worth naming: these plans count conversations, not messages. If your traffic arrives through WhatsApp, Meta's own per-message charges sit on top of any subscription, and no vendor's monthly fee includes them.
What Private AI BizHub includes, and the offer attached to a demo
UD's private option is Private AI BizHub. It deploys on your internal servers or private cloud, runs on GPU infrastructure, and is positioned as an internal knowledge base rather than a customer-facing chatbot.
Functionally, you import raw Word, PDF and Excel documents, and the system indexes SOPs, product specifications, meeting minutes and contracts into a Q&A-ready knowledge base with permission-based access. It supports switching between leading models, including ChatGPT, Claude Opus, Gemini and DeepSeek, inside the private environment.
The offer, stated plainly
--- What it is: book a free demo and receive a free HK$10,000 Cloud Security Assessment.
--- Eligibility: booking the demo through the form on the product page. Name, company, company email and phone number are required.
--- Card required: no. It is an enquiry form, not a checkout.
--- Expiry: UD labels it a limited-time offer and does not publish an end date. Confirm the deadline when you book.
--- Price of the product itself: not published. You will be quoted after the demo.
That last line matters more than the offer. A product with no published price cannot be budgeted from a website, which is a real disadvantage against tools that print their numbers.
The honest read is that the free assessment is worth taking on its own terms. A security assessment tells you something useful about your business whether or not you ever buy the AI.
Where the cloud option wins, and where UD is the wrong buy
There are four situations where you should not buy UD's private option, and one where you should not buy anything at all.
You have no server room and no IT staff
Private AI BizHub deploys on your internal servers or private cloud. A six-person firm with a broadband line and a NAS is buying infrastructure, not software. Cloud wins on speed and on total cost.
You need a price before a phone call
If your approval process requires a written figure before a meeting, the absence of published pricing is a genuine blocker. Cloud tools that publish tiers, UD's own included, are easier to get signed off.
You want to test for one month
AI Chatbot Master carries a minimum six-month subscription. If your plan is a four-week trial and a decision, that term does not fit, and a month-to-month competitor is the better buy.
You want on-premise but also want a shortlist
UD is not the only Hong Kong vendor offering private or hybrid deployment. Set Sail AI, for example, offers role-based access, audit logs and cloud, hybrid or on-premises options. A shortlist of two or three quotes is a better process than one demo.
The buy-nothing case
If your volume is a few hundred questions a day and your data is ordinary commercial information, no private deployment will pay for itself on cost. The only remaining reasons are contractual or regulatory. If neither applies to you, keep the subscription and spend the difference elsewhere.
Frequently asked questions
Is a private AI more secure than a cloud subscription?
It is more controllable, which is not the same thing. A private deployment means your documents stay in an environment you own, but the security of that environment is now your responsibility, including patching, access control and backups. A well-run cloud service can be safer than a badly run private server.
How much cheaper is a subscription, really?
At Hong Kong SME volumes, substantially. A published cloud tier starts at HK$298 a month, while the cheapest credible self-hosted setup starts around US$250 a month in hardware plus US$750 or more in monthly maintenance labour. The gap closes only above roughly 2 million tokens a day.
Does the Personal Data (Privacy) Ordinance require a private AI?
No. The PDPO sets obligations about how you handle personal data, not about where the server sits. Private deployment is one way to meet those obligations and a cloud contract with clear terms is another. This is general information rather than legal advice.
Can government funding cover either option?
Possibly. The Technology Voucher Programme ended in March 2026. The Government announced HK$300 million on 7 June 2026 to enhance the Digital Transformation Support Pilot Programme for SME AI and cybersecurity adoption, targeted for rollout in the second half of 2026, and the BUD Fund has explicitly recognised AI-related measures since mid-June 2026.
What should I ask a private AI vendor before signing?
Four things in writing: where the servers physically sit, who holds administrator access, how long logs and documents are retained, and what happens to your indexed data if you cancel. Vendors who answer those quickly are the ones worth shortlisting.
The verdict, and what to do next
One to twenty staff, general office use. Buy nothing private. A published cloud tier, starting at HK$298 a month, does the job and keeps the decision reversible.
Twenty to one hundred staff with client confidentiality duties. Law firms, insurance brokerages, medical practices and family offices should treat this as a contract question rather than a cost question. Book the demo, take the free assessment, and get deployment location in writing.
Regulated businesses. If a regulator can ask you for audit logs, put retention, residency and access control in the contract before comparing prices at all.
Genuinely high volume. Above roughly 2 million tokens a day, the arithmetic changes and private hosting starts to win. Get quotes for both and let the number decide.
Whichever row you are in, the next step is the same and takes twenty minutes: count how many AI questions your company actually asks in a week. That single number tells you which half of this article applies to you.
Because the useful question was never "is private AI better". It was "is it better for a business your size". We understand AI. UD stands with you.
For background, see our explainer on data residency, and the product pages for Private AI BizHub and AI Chatbot Master.
Reviewed by the UD AI team. UD has been in business in Hong Kong for 28 years.
Not sure which side of the line your business falls on? Book a free Private AI BizHub demo and receive a free HK$10,000 Cloud Security Assessment. We will walk you through it step by step, from counting your real volume to a written deployment plan.
