The Sovereign AI Guide

What is sovereign AI? A practical guide to data sovereignty for AI

Where your AI runs now decides which laws apply to it. This guide covers what sovereign AI means, where AI data actually ends up, the difference between data residency and data sovereignty, how foreign laws reach data abroad, the rules in each market we work in, and how to choose models, deployment and compute that keep AI under your control.

What sovereign AI is

Sovereign AI means the data an AI system uses, the model that processes it and the computers it runs on all sit under a jurisdiction you chose, operated by a company you can hold to account there. The term started with governments building national AI capacity. It now applies just as much to a bank in Jakarta, a law firm in Singapore or a hospital group in Riyadh deciding where its AI workloads should run.

A windowless data-centre building at dusk behind a steel perimeter fence, vehicle barrier, bollards, security cameras and a lit gatehouse
Sovereignty starts with a physical place: who can get through the gate, who runs the building, and which law governs both.

For a business, sovereignty comes down to three questions:

  • Where is the data stored and processed? This covers prompts and documents, plus embeddings, logs, backups and fine-tuned model weights.
  • Which country's law can compel access to it? This follows the ownership of the provider, not just the location of the server.
  • Who operates the hardware? That means who has physical and administrative access, and whom your contract and regulator can reach.

AI makes these questions sharper than ordinary cloud hosting did. A language model reads the full text of everything staff and customers send it. A model fine-tuned on your data keeps traces of that data in its weights. Request logs and vector stores can reconstruct the source documents. Every one of these sits somewhere, under someone's law, and most organisations have never been asked to account for them.

It helps to separate the two ways the phrase is used. National sovereign AIis about a country owning enough compute, data and models that its economy doesn't depend on foreign platforms: Singtel's RE:AI GPU cloud, the ILMU language model running in YTL's Johor AI data centre and Indonesia's plan for a national cloud in sovereign data centres are all examples. Organisational sovereign AI is narrower and more practical: making sure your own AI workloads run where your regulators, clients and board expect them to. This guide is mostly about the second, though the first is what is building the infrastructure you can now choose from.

Why sovereign AI is on the agenda now

For most of the last decade, the default answer to "where should this run?" was the nearest region of a large US cloud. Three things have changed that.

Spending is moving. Gartner expects worldwide spending on sovereign cloud infrastructure to reach about US$80 billion in 2026, up 35.6% on 2025, and to pass US$110 billion in 2027. The fastest growth is not in Europe, where the debate is loudest, but in the Middle East and Africa and in mature Asia-Pacific markets.

Worldwide sovereign cloud IaaS spending, US$ billion
04080120$59.3B2025$80B2026+35.6%$110.6B2027forecast

Source: Gartner, 9 February 2026. Fastest 2026 growth: Middle East and Africa +89%, mature Asia-Pacific +87%, Europe +83%.

Buyers are saying it out loud.In a Gartner survey of Western European CIOs published in November 2025, 61% said geopolitics would increase their reliance on local or regional cloud providers. In June 2025, a Microsoft France executive told the French Senate under oath that the company could not guarantee French customers' data would never be handed to US authorities without French consent, while noting that no such request had been made. That single answer did more to move procurement conversations than years of white papers.

Regulation is catching up.The EU's proposed Cloud and AI Development Act, published in June 2026, sets out assurance levels for public procurement, with the highest requiring immunity from non-EU law. Across Asia and the Gulf, new and amended data-protection laws have arrived almost every year since 2020, and sector regulators have added their own expectations for AI.

Different regions, different drivers

In Europe, buyers talk about foreign legal reach. In Southeast Asia, the Gulf and India, the documented driver is in-country hosting and local ownership, often still on US hardware.

That difference matters when you plan. A Saudi ministry or an Indonesian bank will rarely mention the CLOUD Act; they will ask which data classification applies, where the data centre is and who operates it. A German manufacturer is more likely to ask who owns the provider.

Where AI data actually lives

Most data-protection reviews look at the main database. AI systems scatter data much more widely, and each copy needs its own answer to the three questions above.

Where data lives inside one AI system
AI model+ serving stackPrompts & uploadsVector storeRequest logsFine-tuned weightsBackupsSupport access

Six places your data ends up, beyond the model itself.

Hover or tap a store. Each one sits in a location, under someone's law, and needs its own answer.

  • Prompts and uploads carry the full text of contracts, medical notes, customer emails and source code.
  • Vector stores hold embeddings of your documents for retrieval. Embeddings are not anonymous: with access to the model, much of the original text can be recovered.
  • Request logs are often switched on by default for debugging and kept by a third-party observability tool, sometimes in another country.
  • Fine-tuned weights are a compressed form of your training data. A model trained on support tickets can repeat parts of them.
  • Backups and disaster recovery copies frequently land in a second region chosen by the provider, not by you.
  • Support access is the one most often missed: the engineer who can log in to fix an outage may sit in a different jurisdiction from the server.

A useful exercise is to draw this diagram for one real AI system in your organisation and write a country and a company next to each box. Gaps usually show up within an hour.

Data residency vs data sovereignty

The two terms are often used as if they were the same. They are not, and the difference is where most compliance surprises come from.

Data residency is about location: the data is stored in a particular country. Many regulators and customers ask for it, and most cloud providers can offer it by letting you pick a region.

Data sovereignty is about control: which laws apply to the data and which governments can require it to be handed over. That depends on who owns and operates the infrastructure as well as where it is.

Same server location, different reach
SingaporeUnited StatesYour dataParent HQLegal order to hand over dataResident ✓ Sovereign ✗

Simplified illustration. Whether a foreign law reaches a provider depends on its group structure and is a question for legal review.

The key distinction

A US company's data centre in Singapore gives you residency in Singapore. It does not, on its own, take the data out of reach of US law.

There is a third term worth knowing. Data localisation is a legal requirement that certain data stay in a country. Residency is what you choose; localisation is what a law or regulator forces. Localisation rules are where residency stops being optional, and they are usually narrower than people assume.

For many workloads residency is enough. For regulated data, government-adjacent work or clients who ask pointed questions about foreign access, sovereignty is what they actually need.

The CLOUD Act and foreign legal reach

The US Clarifying Lawful Overseas Use of Data (CLOUD) Actof 2018 lets US authorities require service providers to produce data in their "possession, custody, or control", regardless of where in the world it is stored. In practice this covers the large US cloud providers wherever their data centres are.

The reach doesn't stop at US headquarters. Law-firm analysis expects courts to apply the "legal right or practical ability" tests used in civil discovery, which means a foreign subsidiary can be reached through a US parent that controls its data, and a non-US company with enough US presence can be in scope itself. Several AI infrastructure companies have added US listings, US investors or US offices in the last two years, so the operator's group structure matters as much as its address.

The US is not unique. China's national security and intelligence laws require organisations to support state intelligence work, which gives similar reach over Chinese-owned providers wherever they operate. Other countries have their own lawful-access powers. The practical question for any provider is the same: who can make them hand over my data, and have I accepted that?

The large US providers have responded with sovereign offerings. Microsoft has committed to contest in court any order to suspend its European operations and works with partner-operated clouds in France and Germany; AWS launched a European Sovereign Cloud in Germany in January 2026. These improve operational separation and protect against service being switched off. Critics point out that ownership still traces to the US parent, so they address suspension risk more than lawful data demands. Whether that matters depends on your requirement, which is why the next section starts with the rules.

Data-protection rules by market

Data-protection law has tightened quickly across Asia, the Gulf and beyond. Most of the laws that shape where AI workloads can run took effect after 2020.

When the main data-protection laws took effect
20102014201820222026South Korea PIPA · Sep 2011Singapore PDPA · Jul 2014EU GDPR · May 2018Brazil LGPD · Sep 2020China PIPL · Nov 2021UAE PDPL · Jan 2022Thailand PDPA · Jun 2022Saudi PDPL enforced · Sep 2024Indonesia PDP Law · Oct 2024Malaysia PDPA transfer reform · Apr 2025India DPDP Rules · Nov 2025Vietnam PDP Law · Jan 2026

Dates are when the core obligations or the cited reform applied. Sources are listed in the references below.

The table summarises the main regimes in the markets we work in. It is a starting point, not legal advice; your counsel confirms how they apply to you.

SingaporePDPATransfers abroad need protection comparable to the PDPA, usually secured by contract. Sector regulators such as MAS add outsourcing and technology-risk expectations.
MalaysiaPDPA (amended)The whitelist was replaced with a risk-based transfer regime from 1 April 2025, with published transfer guidelines and mandatory breach notification.
IndonesiaPDP Law 27/2022In force since 17 October 2024. Transfers allowed with equivalent protection, safeguards or consent; GR 33/2026 sets detailed transfer rules from 16 January 2027.
ThailandPDPAFully in force since June 2022; detailed cross-border transfer rules took effect on 24 March 2024.
VietnamPDP Law 91/2025Effective 1 January 2026. Storing data with a foreign cloud counts as a cross-border transfer and needs an impact assessment.
UAEFederal PDPLFederal Decree-Law 45 of 2021. Government, health and banking data sit under separate, stricter rules; DIFC and ADGM run their own regimes.
Saudi ArabiaPDPLEnforced since September 2024, with standard contractual clauses for transfers. Government data follows SDAIA classification and CST cloud rules.
IndiaDPDP Act + RulesRules notified 14 November 2025 and phased in. Transfers abroad are allowed except to restricted countries; RBI keeps payment data in India.
JapanAPPITransfers abroad need consent or an equivalent-protection basis, with information duties on the destination's regime.
South KoreaPIPAOverseas transfers need consent or another statutory basis; certification and adequacy routes were added in recent reforms.
European UnionGDPRTransfers outside the EU need an adequacy decision or safeguards such as standard contractual clauses.
BrazilLGPDInternational transfers need adequacy, standard clauses or another legal basis set by the regulator.

Sources for each market are listed in the references at the end of this guide.

Two patterns stand out. First, most general data-protection laws allow transfers abroad with the right safeguards, and several, including Malaysia, Thailand and Saudi Arabia, have recently made transfers easier, not harder. Sovereignty is rarely an outright ban. Second, the strict rules usually come from somewhere else: sector regulators, government data classifications and client contracts. They are the subject of the next section.

Sector rules that bite harder than the general law

When a workload really has to stay in-country, the reason is usually one of these rather than the headline privacy law.

  • Government data. Saudi Arabia classifies government data as public, confidential, secret or top secret, and the classification directly limits which cloud providers can host it. Indonesia requires public-sector electronic system operators to keep data in the country.
  • Banking and payments. The UAE Central Bank expects licensed institutions to keep customer data onshore, and the Reserve Bank of India has required payment-system data to be stored only in India since 2018. Regulators in Singapore and Malaysia set outsourcing and technology-risk rules that shape how banks use any cloud, AI included.
  • Health. The UAE requires health data to be stored inside the country under its 2019 health-data law, and health ministries elsewhere set similar conditions for patient records.
  • On request. Vietnam can require foreign providers to localise data when the Ministry of Public Security asks, and its 2025 personal-data law treats storage with a foreign cloud as a transfer that needs an impact assessment.
  • Client contracts. Large clients, especially governments and multinationals, increasingly write location and operator requirements into their supplier terms. For many service firms this is the strictest rule of all.

The practical upshot is to classify before you architect. A customer-service chatbot that sees order numbers and a document-search tool for a ministry may sit in the same company, but they rarely need the same answer.

Open models and why licences matter

With a closed AI API, the model always sits on the provider's side. Even with strong retention promises, every prompt leaves your boundary. Open-weight models change that: you download the model and run it on infrastructure you chose, so prompts and documents never reach the model's developer.

The quality gap has narrowed sharply. On vendor-reported benchmarks the strongest open models now sit within a few points of the leading closed models for coding and reasoning, and open image, speech and video models are competitive too. For most business tasks, such as answering questions over documents, drafting, classification and transcription, an open model is good enough, and it can be fine-tuned on your own data in a way closed models rarely allow.

Licence status of widely used open models, October 2026
PermissiveConditions applyNot for commercial use
DeepSeek V4MIT
GLM-5.2MIT
Qwen 3.6 (open sizes)Apache 2.0
Gemma 4Apache 2.0
Mistral Small 4 / Large 3Apache 2.0
gpt-ossApache 2.0
Kimi K2 familyModified MIT
Llama 4Community licence: 700M MAU cap, EU multimodal exclusion
Qwen-Image 2.1Research licence, non-commercial

Check each release's licence file before building on it; terms change between versions.

Three cautions apply:

  • Licences vary, and they change. Most leading language models now use permissive licences such as Apache 2.0 or MIT, but some popular releases carry user caps or regional exclusions, and image models in particular are drifting towards non-commercial licences with paid tiers. Qwen-Image moved from Apache 2.0 to a research licence with its 2.1 release in September 2026.
  • Origin matters to some buyers. Government suppliers and some multinationals prefer models that are not Chinese-developed. Self-hosting removes the data-exfiltration concern, but procurement rules may still ask for a different base model.
  • Language coverage differs. Models trained mainly on English can be weak in Thai, Bahasa or Vietnamese. AI Singapore's SEA-LION family is tuned for Southeast Asian languages, and its latest release is built on an NVIDIA base model, which gives buyers a non-Chinese option.

In practice that means a shortlist: Gemma, Mistral and gpt-oss as widely accepted defaults, Qwen and DeepSeek where their strengths in code and Asian languages matter, and SEA-LION for regional languages. See our private LLM hosting and Qwen API pages for how we deploy them.

Six ways to deploy, from least to most control

Sovereignty is not a yes-or-no choice. Each step along this spectrum buys more control at the cost of more setup and, at the far end, more capital. Click through to compare.

Six ways to run AI, from least to most control

Local specialist cloud. A locally owned GPU cloud in your jurisdiction. Residency and legal control line up, often at a lower GPU price.

Most organisations end up using more than one. A typical pattern is a public API for low-risk drafting and coding help, a local specialist cloud or the company's own cloud account for anything touching customer data, and on-premises hardware only where a regulator or client demands it. The point is to make the choice deliberately for each workload rather than inherit it from whichever tool a team signed up for first.

Choosing compute and location

Once you know where the data has to stay, the next decision is whose GPUs run the model. There are three broad options, and they differ on price as much as on control.

Price per H100 GPU-hour by buying option, US$
Hyperscaler on-demand$6.88Specialist cloud on-demand (median)$3.39Specialist cloud, 1-year reserved$2.44

Source: Akash Network H100 price survey, Sep–Oct 2026. List prices; negotiated rates differ.

  • Hyperscalers are convenient and integrate with everything you already run, but they are the most expensive for GPU time and, being US companies, the hardest to fit a strict sovereignty requirement.
  • Specialist GPU clouds are usually far cheaper, and some are locally owned and operated in the markets that matter to you. Check the operator's ownership and financial health before signing a long term.
  • On-premises servers give the most control but tie up capital in hardware that loses value as new generations ship. They make sense for steady, high-utilisation or air-gapped workloads. Our AI servers page covers sizing and sourcing.

Physical location is shifting too. Singapore rations new data-centre power: its second capacity call awarded 200 MW to four operators in August 2026, with green-energy conditions attached. Meanwhile contracted capacity in Johor grew from under 0.4 GW at the end of 2023 to more than 2.3 GW at the end of 2025, and new subsea cables put Batam within a few milliseconds of Singapore.

Where new data-centre capacity is going, megawatts
Johor contracted capacity, end-2023≈ 0.4 GWJohor contracted capacity, end-2025> 2.3 GWSingapore DC-CFA2 award, Aug 2026200 MW

Sources: Structure Research (Johor); EDB / IMDA (Singapore, four operators × 50 MW).

For sovereignty this creates a choice that didn't exist five years ago. A Singapore company can keep its contracts, its support and its low-latency workloads in Singapore while running heavy GPU work in Johor or Batam at a lower cost, but that moves the compute into Malaysian or Indonesian jurisdiction. Whether that is acceptable depends on the data, which is why the classification step comes first. Our infrastructure advisory page compares the three locations in detail.

Three worked examples

Financial services · Singapore

An internal assistant for relationship managers

Staff want to ask questions over client files and product documents. The data includes personal and confidential client information, and technology-risk and outsourcing rules apply. The answer: an open model fine-tuned on product documents, deployed in the bank's own cloud account in Singapore, with the vector store and logs in the same account, SSO through the existing identity provider and no calls to public AI APIs.

E-commerce · Indonesia

Product imagery at marketplace scale

A retailer wants thousands of product images restyled for different marketplaces. The inputs are product photos, not personal data, so residency matters less than cost and licence. The answer: a permissively licensed image-editing model on a specialist GPU cloud in the region, with every model's licence recorded so the images can be used commercially.

Government supplier · Saudi Arabia

Document search for a ministry contract

The documents are classified government data, so the hosting choice follows the classification rules, not just the privacy law. The answer: in-country infrastructure from an approved local operator, a model the client's procurement rules accept, and access logs the ministry can audit.

Common myths

"If the server is in our country, we're sovereign."
That is residency. Sovereignty also depends on who owns and operates the provider and which laws reach them.
"A non-US provider is automatically safe from US law."
Not if it has a US parent, a US listing or enough US presence. Check the group structure of the contracting and data-controlling entities.
"Our privacy law bans sending data abroad."
Most laws in the region allow transfers with safeguards. The hard rules usually come from sector regulators, government classifications and contracts.
"Sovereign AI means worse models."
Open models are now close to closed ones for most business tasks, and they can be fine-tuned on your data.
"Sovereign AI is always more expensive."
Local specialist clouds often charge less per GPU-hour than hyperscalers. The extra cost tends to be setup and governance, not compute.

A sovereign AI checklist

Work through it for one AI system at a time. Tick items off as you go; nothing is saved or sent.

0 of 10 done

If you want help working through it, we start every sovereign AI engagement with exactly this inventory.

Glossary

Data residency
The physical location where data is stored and processed.
Data sovereignty
Which jurisdiction's laws govern data, including who can compel access to it.
Data localisation
A legal requirement that certain data be stored, and sometimes processed, inside a country.
Sovereign cloud
Cloud infrastructure designed so data, operations and legal control stay within one jurisdiction.
CLOUD Act
A 2018 US law letting US authorities require providers to produce data they control, wherever it is stored.
Open-weight model
An AI model whose trained weights are published, so anyone can run it on their own hardware.
Vector store
A database of embeddings used to retrieve relevant documents for a model; it can reveal the source text.
Fine-tuning
Further training a model on your own data. The resulting weights can retain traces of that data.
Air-gapped
A system with no connection to outside networks, used for the most sensitive workloads.
Sub-processor
Any third party a provider uses to handle your data, such as a backup or support vendor.

Frequently asked questions

Sources and references

  1. Gartner: sovereign cloud IaaS spending forecast (Feb 2026)
  2. Gartner: Western European CIO survey on local cloud providers (Nov 2025)
  3. US Department of Justice: CLOUD Act resources
  4. Skadden: understanding the CLOUD Act's reach
  5. SDxCentral: Microsoft France testimony to the French Senate (June 2025)
  6. Hogan Lovells: the EU Cloud and AI Development Act proposal (2026)
  7. Baker McKenzie: Saudi Arabia data-transfer regulation update (2024)
  8. Chambers: Saudi Arabia cloud computing trends (2026)
  9. Middle East Briefing: UAE data-protection and transfer obligations
  10. Mayer Brown: Malaysia PDPA amendments and transfer guidelines (2025)
  11. Tilleke & Gibbins: Thailand cross-border transfer rules (2024)
  12. Rajah & Tann Asia: Indonesia GR 33/2026 on cross-border transfers
  13. HLC: Vietnam's Law on Personal Data Protection (2025)
  14. Reserve Bank of India: storage of payment system data (2018)
  15. Structure Research: Johor and Batam data-centre report (2026)
  16. Akash Network: H100 rental price survey (2026)

This guide is general information, not legal advice. Laws, licences and providers' ownership change; confirm how they apply to you with counsel.

Put your AI where your regulators expect it

Tell us where your data has to stay. We'll come back with a region, an operator and a cost comparison.