Must a Singapore clinic self-host its AI model?
Short answer
No. No Singapore law we could find requires a licensed healthcare provider to run a large language model on its own servers. On 4 August 2026, Health Minister Ong Ye Kung told Parliament that data security requirements apply to AI tools that process patient data "whether hosted on third-party cloud services or on-premise", under both the Healthcare Services Act and the PDPA (written answer). So the question isn't where the model runs. It's whether you can show the controls. Self-hosting is one way to make that easier to show, especially if you can't get a vendor to promise it keeps nothing.
28 Oct 2026 · This note is general information, not legal advice.
What the rules actually say
Healthcare Services Act 2020 (HCSA) and the PDPA. MOH's own AI guidelines list both. Under the HCSA, licensed institutions "including those which use AI solutions, must comply with its requirements (e.g. controls on the maintenance and security of medical records)". The PDPA covers "the collection, use, disclosure, protection, and access of personal data used in AI solutions" (AIHGle 2.0, Table 6). Our PDPA note explains why the PDPA tests control, not location.
What public hospitals already demand. In the same answer, the Minister said public healthcare institutions require AI model providers to "give legally-binding commitments that all input and output data are not stored or retained", and that "the AI tools also need to be accessed from secure environments". That is the bar a private provider will be compared against.
The Health Information Act (HIA). An MOH circular dated 6 March 2026 says licensed providers, retail pharmacies and approved NEHR users "must implement reasonable security measures to protect health information and systems" and meet breach-notification duties. The deadlines are phased (MOH HIA circular):
| Batch | Includes | Cybersecurity and data security measures by |
|---|---|---|
| 1 | GP clinics, acute and community hospitals, clinical and radiology labs, nuclear medicine | September 2027 |
| 2 | Specialist outpatient, nursing homes, renal dialysis and others | September 2028 |
| 3 | Outpatient dental, ambulatory surgical centres, assisted reproduction, retail pharmacy and others | March 2030 |
The HIA implementation guide (v2.0, August 2026) lists the baseline. It includes unique accounts for every user, 2FA for admin and remote access, access only to records of "patients under their care", and checking "service providers' certifications and safeguards". An LLM that reads patient notes is a system that holds health information, so plan for it to meet the same list.
AIHGle 2.0 (MOH, March 2026). These are guidelines, not law. They split responsibility between developers, deployers (HCSA-licensed organisations) and users (clinicians). On generative AI, they name "data disclosure" as a specific risk: models "could inadvertently leak sensitive data when users submit sensitive data during use". They ask deployers to red-team and benchmark LLMs, and recommend reducing hallucination with techniques "such as Retrieval-Augmented-Generation (RAG)". Clinical note summarisation by transcription is their own example of human-in-the-loop use: "Doctors retain full control and make final decision on the case notes saved."
So when does self-hosting make sense?
| If this is true for you | Lean towards |
|---|---|
| Your vendor won't sign a binding no-retention commitment for inputs and outputs | Self-hosted or private-tenant model |
| Patient identifiers must be in the prompt (notes, referrals, discharge summaries) | Self-hosted, inside your existing access controls |
| You need clinicians to see only their own patients' data in RAG answers | Self-hosted RAG on your record permissions |
| Use is admin-only with no patient data (rosters, policy Q&A) | Cloud is usually fine with standard PDPA diligence |
| You have no one to patch and monitor a server | Cloud with strong contracts, or a managed private deployment |
Self-hosting doesn't remove obligations. You still own patching, access control, backups, incident response and the AIHGle-style testing. It removes one party, the model vendor, from the chain of custody. That is often the hardest thing to evidence.
FAQ
Is there a Singapore data-residency rule for health data in AI tools?
We did not find one in the HCSA, PDPA, HIA circular or AIHGle 2.0 as of October 2026. The Minister's answer treats cloud and on-premise the same: both must meet the security requirements.
Does AIHGle 2.0 apply to us?
It targets healthcare services licensable under the HCSA, and is aimed mainly at ML and deep-learning tools, including GenAI and LLMs. If your AI is a medical device, HSA's Health Products Act rules apply on top.
When do the HIA security measures kick in?
It depends on your service type: September 2027, September 2028 or March 2030 (table above). Other HCSA licensees not in the table have until September 2028.
Sources
- Parliament of Singapore, written answer by Minister for Health Ong Ye Kung to Assoc Prof Jamus Lim, 4 Aug 2026 (sgparliament.com republishes Hansard): https://www.sgparliament.com/debate/33260
- Ministry of Health. Artificial Intelligence in Healthcare Guidelines (AIHGle) 2.0, March 2026: https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf
- Ministry of Health. Overview and Implementation of the Health Information Act, circular, 6 Mar 2026: https://isomer-user-content.by.gov.sg/377/07ae2200-91ee-4f5b-8d52-009ca9bdce95/Overview_and_Implementation_of_the_Health_Information_Act__HIA_.pdf
- Ministry of Health. HIA Implementation Guide for Healthcare Providers, v2.0, August 2026: https://isomer-user-content.by.gov.sg/377/9b9a298f-3465-4144-b54f-0e91632fae3c/HIA_IG2_Final.pdf
Planning an LLM that touches patient records? We'll map your use case against the HCSA, PDPA and HIA controls and tell you whether it needs to be self-hosted. Start here →
On this site
- Does the PDPA require an on-premise LLM? — the PDPA tests control, not location.
- How an engagement runs — map the workloads, size the hardware, build one workflow, hand it over.
- Notes — title, date, and a one-line short answer.
- Home — who this is for, and how we install models on your hardware.