Sovereign AI Rules: What India, the EU and Saudi Arabia Ask of Foreign Vendors

Three governments now spend public money in the name of AI sovereignty, and they mean three incompatible things by it. India's compute tender bars user data from leaving the country, even encrypted. The EU caps its own stake in each AI gigafactory at 17% of capital cost and asks that the owners be European. Saudi Arabia is drafting a law that lets foreign governments apply their own law inside Saudi data centres.

The short answer. India requires that subsidised AI compute be sold by an Indian-registered entity from Indian data centres, with user data kept in India. The EU requires that gigafactory consortia be majority owned under its rules and free of high-risk suppliers. Saudi Arabia currently requires nothing comparable of foreign vendors, and instead offers them hosting under their own law.

Sovereign AI rules are three different instruments, not one policy

The phrase "sovereign AI" hides the important variable, which is the legal instrument doing the work. A statute binds everyone in a jurisdiction. A procurement rule binds only those who want the government's money. An ownership rule binds the capital table, not the customer.

Most coverage stacks India, the EU and the Gulf into one trend line labelled "AI sovereignty". That framing is wrong for anyone who sells software. The three regimes put the obligation on different parties, so the same vendor faces a hard rule in one market and an open door in the next.

The earlier generation of this debate was about data residency, meaning where records are physically stored. That question is covered in what EU and Indian buyers demand on AI data residency. This post is about the next layer: who owns the compute, who may sell it, and whose law governs the machine.

The money side, meaning how much governments are actually spending on chips, is a separate question from the rules. It is tracked in the analysis of sovereign AI funds as a buyer of last resort. Here we compare only the obligations.

The sovereign AI rules comparison table

Every row below comes from an official document or from reporting on one. Where we could not open the primary text ourselves, the source column says whose reading we are relying on.

The last row is not a Saudi rule. It is in the table because, for a vendor deciding where to run workloads in the Gulf, it is the constraint with teeth.

India: IndiaAI Mission compute binds the data and the seller

India's instrument is a procurement contract. The government does not own GPUs. It empanels private providers and subsidises access for startups, researchers and public bodies. The rules sit in the eligibility criteria for becoming a provider.

The eligibility rules in the 2024 request for empanelment

MeitY's request for empanelment, issued in August 2024, drew 19 bidders, including Jio Platforms, Tata Communications and Yotta Data Services. To qualify, a bidder had to be a legal entity registered in India and had to deliver all AI services from Indian data centres.

The financial bar was modest by hyperscaler standards. Bidders needed average annual turnover above Rs 100 crore across three years, with Rs 50 crore for lead members of a consortium. They needed at least 1,000 AI compute units on a self-service cloud, or a commitment to reach that within 6 months. Empanelment runs 36 months, extendable by 12.

Notice what is absent. The criteria reported do not bar foreign ownership. A global cloud provider with an Indian subsidiary and Indian regions is not excluded by these words. What it cannot do is serve a mission workload from Singapore.

The 2025 clauses that matter more than the headline

The re-issued tender in 2025 went further, according to MediaNama's reading of the document. Any "enabling services" the provider needs to deliver the AI service must also sit in India. User-uploaded data may not leave Indian territory "in any form, be it anonymous, pseudonymous, or encrypted".

That last clause is the one a B2B vendor should read twice. Encryption is the usual answer to residency objections. Here it is ruled out by name. If your product calls a model endpoint, a logging service or an evaluation tool hosted abroad, and your customer runs on IndiaAI compute, your architecture fails the clause.

The same tender requires an annual third-party audit at the provider's own cost and compliance with evolving CERT-In requirements. It also folds in the Digital Personal Data Protection Act 2023 and the IT Act 2000. In practice that is a running compliance cost, not a one-time certification.

Our view is that India has chosen the most enforceable model of the three. A procurement clause is checked at contract signature and at every audit. It needs no new statute, no court and no treaty. It also binds far fewer companies than a law would.

EU AI gigafactories: sovereignty through ownership, not location

The EU's instrument is a joint undertaking, EuroHPC, which co-funds large compute facilities with member states and private consortia. The rules sit in the regulation that governs it and in the tender for the facilities.

What the January 2026 amendment added

The Council adopted an amendment to the EuroHPC regulation on 16 January 2026. According to Open Future's summary of the act, it formalises a majority ownership rule for AI gigafactory consortia and excludes high-risk vendors from their supply chains. It also gives the EU a legal basis to own a share of computing time and redistribute it to European startups.

The financing is where the sovereignty claim gets thin. The EU contribution is capped at 17% of capital cost, matched by the host member state. The private consortium covers about two-thirds of construction and all operating cost, according to an analysis by University of Amsterdam researchers. The same authors call the EU "a minority tenant in infrastructure using significant sums of public money".

What the call for tenders asks of bidders

The tender itself closes on 12 November 2026. PNO Innovation's breakdown of the call sets out two scales. A medium facility needs a minimum of 25,000 accelerators and a large one 40,000. EU commitment is up to EUR 500 million and EUR 1 billion respectively, delivered mainly as purchased compute time rather than a grant.

Applicants must be consortia or dedicated vehicles with average consolidated turnover above EUR 200 million over three years. The first phase must be running within 18 months of contract signature. Capacity must be open to European researchers, companies, SMEs, startups and public bodies.

Read together, the EU rule binds the owners and the supply chain. It does not, on the documents we could read, require that your SaaS product run inside a gigafactory. A US vendor's EU region is unaffected. A US vendor wanting to co-own a gigafactory is directly affected.

We think the EU has written the most precise rule and attached the least leverage to it. A 17% stake with a majority-ownership condition is a gate on who sits at the table. It is not a gate on what European customers buy.

Humain and Saudi Arabia: sovereignty offered as an export

Saudi Arabia inverts the question. India and the EU ask what foreign vendors must do to serve them. Saudi Arabia is asking what it must offer foreign governments and companies to bring their workloads to Saudi soil.

The draft Global AI Hub Law

The Communications, Space and Technology Commission released a draft Global AI Hub Law on 14 April 2025, with consultation open until 14 May 2025. Reed Smith's reading of the draft describes three models. A private hub hosts data for one foreign state under a bilateral agreement and operates under that state's law.

An extended hub lets private data centre operators host services for the guest country, under agreements with both the operator and the competent authority. A virtual hub lets a private service provider offer hosting under the rules of a designated foreign state. The Council of Ministers may terminate any agreement to protect "the Kingdom's safety, national security, and sovereignty", with privileges surviving up to 120 days.

The draft takes effect 60 days after publication in the Official Gazette. We found no record of final publication during research for this post. Treat it as a stated intention, not a rule.

Humain's customers are the policy

Humain, the Public Investment Fund company building Saudi Arabia's AI infrastructure, is where the intention becomes commercial. Its chief executive, Tareq Amin, told EnterpriseAM in October 2026 that the company is targeting 1.7 GW of data centre capacity in 2027 and has signed xAI and Amazon Web Services alongside several open-weight model companies.

Amin's own description of the model is the clearest statement of Saudi policy available: "Global off-takers [are] coming to Saudi to really export AI tokens to the rest of the world." That is sovereignty as a product line. The Kingdom keeps the energy, land and ownership. The customer keeps its law.

The binding constraint is not in Riyadh. In November 2025 the US Commerce Department approved exports of up to 35,000 GB300-equivalent chips each to Humain and the UAE's G42, conditional on "rigorous security and reporting requirements". A tenant of Humain inherits those conditions, whatever the hub law eventually says. The wider US licensing picture is covered in the analysis of AI export controls and the Fable suspension.

What sovereign AI rules ask of a foreign B2B vendor

If you sell software rather than compute, the three regimes reach you in different ways. The question is not "is this country sovereign-minded". It is "which party does the instrument bind, and am I that party or downstream of it".

In India, you are downstream. The rule binds the compute provider, but its data clause travels with the workload. If an Indian startup builds on subsidised IndiaAI capacity and buys your tool, your tool's data flows are now inside the clause. Expect that to arrive as a contract question, and prepare an answer covering every subprocessor. The clauses that usually carry this are covered in the guide to AI clauses in data processing agreements.

In the EU, you are probably outside the gigafactory rules entirely. They bind owners and suppliers of the facilities. Your exposure is the separate stack of EU product law, chiefly the AI Act, which is mapped in the breakdown of GPAI enforcement and downstream obligations. Do not confuse the two in a sales conversation.

In Saudi Arabia, you are a potential customer of the sovereignty, not its subject. The offer is cheap power, local capacity and, if the hub law passes, legal insulation for your home jurisdiction. The cost is that you sit inside US export conditions you did not negotiate.

The bottom row is the concession most sovereignty commentary leaves out. All three instruments reach only the slice of the market that touches public money or state-linked capacity. That slice is growing, and it is still a slice.

Where this comparison is weakest

We would rather state these limits than have a reader find them.

Two of three rule texts were read second-hand

We could not open the IndiaAI tender PDF, the EU regulation in the Official Journal or the gigafactory tender specification during research. Every requirement above rests on a named secondary reading of a primary document: MediaNama for India, Open Future, Policy Review and PNO Innovation for the EU, Reed Smith for Saudi Arabia. Those are credible readers. They are not the text.

The specific gap that matters most is the EU majority ownership rule. Our sources state the rule exists. None we could read defines precisely who must hold the majority or how "high-risk vendor" is determined. Those definitions will decide whether a US cloud company can be a minority partner, and they should be read in the regulation itself before any decision.

The Saudi law may never be enacted in this form

A draft opened for consultation in April 2025 is not a law in October 2026. Drafts change, and the data embassy idea depends on bilateral agreements that may never be signed. A fair critic would say we have compared two live instruments with a proposal. That critic is right, and it is why the Saudi column of the matrix is mostly mid-tone.

There is also a competing view that India's rule is the weak one. Procurement binds only those who bid, and global clouds can simply decline mission work. On that reading the EU rule, which shapes who owns the continent's largest compute, has more long-run reach. We think enforcement beats reach over a three-year horizon. Reasonable people can weight it the other way.

Signals to watch through 2027

Each regime has one event that would change the analysis above. None of them is a headline GPU count.

For India, watch the next compute tender. If the "any form" data clause is extended to model APIs or software services bought with mission subsidies, the rule moves from compute providers to every vendor in the stack.

For the EU, watch the award of the gigafactory bids. The consortium line-ups will show in practice what the majority ownership rule permits. If US hyperscalers appear as minority partners, the rule is a governance condition. If they are absent, it is an exclusion.

For Saudi Arabia, watch two things: Official Gazette publication of the hub law, and the next US chip licence decision. The first creates the legal product. The second decides how much of it can be sold. Utilisation economics for a tenant are a separate test, set out in the GPU break-even utilisation model.

Frequently asked questions

Where to start

Pull your customer list for India and mark every account that has received IndiaAI compute allocations. For each, map where your product sends their data, including logging, model calls and analytics. Any hop outside India is the conversation you will have at their next renewal, so have it now.

If your roadmap includes Gulf hosting, ask your counsel one question before signing anything: which US export licence conditions flow down to us as a tenant. The answer will tell you more about your obligations than the hub law will. For the full vendor-side checklist, use the AI procurement checklist in reverse, as the questions your buyers will now ask you.

Related reading