Not to a consumer chatbot, and not to any AI service until you know where the data is processed, how long it is kept and who can be compelled to hand it over. For a Singapore bank or other financial institution, confidential customer information sits under Banking Act secrecy, MAS outsourcing rules and the PDPA, and MAS's new Guidelines on AI Risk Management make the institution accountable for AI it buys as well as AI it builds. Enterprise AI APIs can be used with the right contracts and controls. As of November 2026, though, none of the major vendors processes its flagship business AI in Singapore. Running an open-weight model on hardware you control, such as a DGX Station GB300, keeps processing inside your own environment, at the cost of running it yourself.
Which rules apply to financial data and AI in Singapore?
Several overlap. None of them bans AI. Together they decide what you must check before data leaves your systems.
| Rule | What it says that matters here | Status |
|---|---|---|
| Banking Act 1970, section 47 | "Customer information must not, in any way, be disclosed by a bank in Singapore or any of its officers to any other person except as expressly provided in this Act." Fines of up to S$125,000 and/or 3 years' jail for individuals, and up to S$250,000 otherwise | In force |
| MAS Notice 658 and Guidelines on Outsourcing (Banks) | Outsourcing register, notification of sub-contractors, due diligence, and the right to terminate if a provider fails to safeguard customer information | Effective 11 Dec 2024 |
| MAS Technology Risk Management Guidelines | Confidential data at rest, including on cloud storage, "should be encrypted"; third-party services must be managed | Issued Jan 2021 |
| MAS Guidelines on AI Risk Management | Apply to "all FIs and all forms of AI technologies"; "FIs remain accountable for AI used in the services they deliver"; cover generative AI and agents | Final 7 Oct 2026; effective 7 Oct 2027, Sections 5–6 by 7 Oct 2028 |
| PDPA section 26 and PDP Regulations 2021 | Personal data sent overseas needs a "comparable" standard of protection, usually through contracts or certifications | In force |
| PDPA breach notification | Notify the PDPC within 3 calendar days of assessing a breach as notifiable; 500 or more affected individuals counts as significant scale | In force |
Sources: Singapore Statutes Online (Banking Act s47; PDPA ss24, 26, 26B–26D; PDP Regulations 2021; Notification of Data Breaches Regulations 2021); MAS Notice 658 and Third-Party Risk Management pages; MAS TRM Guidelines (Jan 2021); MAS media release on the AI Risk Management Guidelines (7 Oct 2026). Accessed 5 Nov 2026. This is a summary for orientation, not legal advice. Confirm your position with compliance and counsel.
What do the AI vendors actually do with your data?
Their business tiers are far better than consumer chat apps. None of the four major vendors trains on business API or enterprise data by default. The differences are in retention and in where processing happens:
- OpenAI API: no training by default; abuse-monitoring logs are kept for up to 30 days, and Zero Data Retention needs OpenAI's approval. Singapore is one of ten data-residency regions, but OpenAI's documentation says only the US, Europe and the UAE offer regional processing. Singapore residency covers storage at rest.
- Anthropic API: no training on commercial inputs and outputs by default, with deletion within 30 days. Inference location can be set to "us" or "global" only, and stored data stays in the US. There is no Singapore option.
- Google (Vertex AI, now documented as Gemini Enterprise Agent Platform): regional endpoints keep processing within the chosen jurisdiction, and Singapore (asia-southeast1) is listed as a processing location. Which Gemini models are available there should be checked per model. Global endpoints give no residency guarantee.
- Microsoft 365 Copilot: prompts and Graph data are not used to train foundation models. Microsoft's in-country processing programme for Copilot does not include Singapore, and queries from outside the EU "may be processed in the US, EU, or other regions".
Two further points matter for financial institutions. First, residency is not the same as control: a US provider can be compelled under the US CLOUD Act (18 U.S.C. §2713) to produce data in its "possession, custody, or control" wherever it is stored. Our post Data Residency vs Data Sovereignty explains the difference. Second, terms change. Anthropic's newest top-tier models require 30-day retention, so an existing zero-retention arrangement does not cover them.
The question is not whether an AI vendor is trustworthy. It is whether you can show a regulator where every piece of customer data went, and why.
Where do financial teams actually want AI?
Mostly on documents that are confidential by nature: credit memos, KYC and onboarding files, transaction narratives for suspicious-activity reviews, internal audit working papers, policy and procedure manuals, and board or committee papers. These are text-heavy, repetitive and high-volume, which suits large language models. They are also the documents most likely to contain customer information covered by banking secrecy. That is why the GovTech-built Pair assistant, approved for government data up to "RESTRICTED \ SENSITIVE NORMAL", is set up so that conversations are kept private from model providers, while public ChatGPT is limited to public data. Financial institutions face the same classification question.
How does a locally hosted model help?
It removes the transfer from the picture for that workload. If the model runs on a DGX Station in your own premises or data centre, prompts and documents are never sent to an external AI provider. There is no overseas processing to justify under PDPA section 26, no additional sub-contractor to add to an outsourcing register for the model itself, and no vendor retention window to manage. A DGX Station GB300 has 748 GB of coherent memory and can serve open-weight models such as gpt-oss-120b to many users at once. StorageReview measured 5,023 output tokens per second at 128 concurrent long-prompt requests on the ASUS ET900N G3.
It does not make a bank compliant by itself. MAS's AI guidelines still expect an AI inventory, risk-materiality assessment, testing, human oversight and monitoring for in-house AI. TRM controls still apply to the Station and its storage: encryption, access control, patching and logging. If the Station connects to the internet for updates, that path needs the same scrutiny as any other system. Our post on isolating your most powerful computer discusses air-gapped operation. And open-weight models trail the best closed models on broad benchmarks, so some tasks may still need a commercial API under a negotiated enterprise agreement.
What should a financial institution do before using any AI on confidential data?
- Classify the data. Decide which categories (customer information, personal data, market-sensitive information) may go to which kind of AI service.
- Map the processing path. For each AI service, record where inference happens, what is retained and for how long, and which sub-processors are involved.
- Route by sensitivity. Send public or low-risk work to commercial APIs if that is cheaper or better. Keep secrecy-covered work on infrastructure you control, or on a vendor arrangement your outsourcing framework has approved.
- Put a gateway in front of everything. One control point for keys, logging and data-loss rules makes audits possible. See Can DGX Station Become the Foundation of Your Private AI Cloud?
- Build the AI inventory now. The MAS guidelines take effect on 7 Oct 2027, and an inventory of AI use is one of the expected controls.
For broader context, see Where Does Company Data Go When Using AI?, Your Employees Are Using ChatGPT. But Is Your Company Data Safe? and Sovereign AI: Why Should Your Company's Intelligence Live on Someone Else's Server?
Buying in Singapore
SourceIT lists the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. HP also offers a headless ZGX Fury configuration with self-encrypting drives, which suits encryption-at-rest requirements. For encrypted local backups, the Synology RackStation RS822RP+ is S$2,086.80 (listed 5 Nov 2026). Browse the range in our NVIDIA DGX Spark and DGX Station collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can a Singapore bank use ChatGPT or Claude with customer data? Not through consumer apps. Enterprise APIs may be usable if the arrangement satisfies Banking Act secrecy, MAS outsourcing requirements and PDPA transfer rules. Check where processing happens: OpenAI's Singapore residency covers storage only, and Anthropic offers US or global inference only.
When do the MAS AI Risk Management Guidelines take effect? MAS published them on 7 Oct 2026. They take effect on 7 Oct 2027, with Sections 5 and 6 by 7 Oct 2028.
Does running AI on a DGX Station make us compliant? No. It keeps prompts and documents off external AI services, which simplifies the transfer and outsourcing questions for that workload. You still need governance, testing, access control and logging.
Do AI vendors train on business data? OpenAI, Anthropic, Google and Microsoft all say they do not train on business API or enterprise data by default. Retention periods and processing locations differ, so read each service's terms.
Is data stored in Singapore safe from foreign access? Not necessarily. Under the US CLOUD Act, US providers can be required to produce data in their control wherever it is stored. Residency and sovereignty are different questions.
