> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # What Is Sovereign AI? Who Pays for a Country’s Own AI Compute - URL: https://www.sotkn.com/explained/sovereign-ai-explained/ - Published: 2026-10-18T09:55:17.000Z - Updated: 2026-10-18T09:55:17.000Z - Description: Sovereign AI is AI computing power, data and models a country keeps inside its borders and under its own laws. Governments rarely run the GPUs; regional neoclouds and telcos do. Nvidia is paid when the chips ship. The operator is paid only when someone signs for the hours. - Author: Sean - Tags: Mechanism, #desk-explained EXPLAINED · Mechanism · Last reviewed: 18 Oct 2026 · Next review: after the EU AI Gigafactory selection in early 2027 or Nvidia's next quarterly results Sovereign AI is AI computing power, data and models that a country keeps inside its own borders and under its own laws. In practice it usually means GPUs in local data centers, run by a local operator, for users the government chooses or pays for. **In short.** Sovereign AI is a buying decision more than a technology. A government decides that some AI work must run at home, under its laws, and then pays for it in one of four ways: buying compute time as an anchor customer, subsidising GPU-hours, buying GPUs for a national centre, or handing land and power to a local operator. Governments rarely run the GPUs. Regional neoclouds and telcos do, and Nvidia said in August 2026 that its sovereign business, mostly sold through those neoclouds, grew 35% in a quarter. Nvidia is paid when the chips ship. The operator is paid only when someone signs for years of GPU-hours, and that signed buyer is the stuck step. Why an investor cares: sovereign AI is now a demand line in Nvidia’s results and in neocloud prospectuses. The useful question is who signs, because a national plan is not revenue until a buyer commits to the hours. If the terms are new, start with [what a neocloud is](https://www.sotkn.com/explained/neocloud-explained/) and [how GPU-as-a-service works](https://www.sotkn.com/explained/gpu-as-a-service-explained/). ## What is sovereign AI? Sovereign AI is about control. A country wants some AI work, such as public services, health and defence data, and models trained on its own languages, to run on machines inside its borders, under its laws, with data that does not leave. The term bundles three things: local compute, local control of data, and local models. Compute means GPUs in a data hall at home. Data control means prompts, outputs and logs stay in the country; Singtel describes its sovereign token service in Singapore as keeping inference, embeddings and logs in Singapore. Models means the national language models governments now fund. India’s government is supporting 12 organisations and consortia to build sovereign foundation models, and Korea’s partners, including NAVER Cloud, LG AI Research and SK Telecom, are building Korean ones. The chips are usually imported. Korea’s plan, announced on 30 October 2025, is for more than 260,000 Nvidia GPUs, and Nvidia calls itself “a neutral partner to every sovereign and neocloud”. Sovereignty here is about who controls the machines and the data, not who designed the chip. ## What is the difference between sovereign AI and sovereign cloud? Sovereign cloud is the older idea: ordinary cloud services, such as storage and databases, run under local law with data kept in the country. Sovereign AI applies the same rule to GPU computing and models. The difference matters for investors because a sovereign AI operator usually owns the GPUs, so far more capital sits on each megawatt, as the Singapore example below shows. | | Sovereign AI | Sovereign cloud | Foreign cloud region | | ----------------------- | ----------------------------------------------------------------- | ----------------------------------------------------------- | -------------------------------------------------- | | **What it is** | GPU computing, data and often models kept at home under local law | Ordinary cloud services kept at home under local law | A global cloud’s data centers built in the country | | **Who usually runs it** | A regional neocloud, telco or national centre | A local provider, or a global cloud through a local partner | The global cloud itself | | **What stays local** | Chips, data, logs and the model weights | Data and the operating staff | Data location, but control sits abroad | | **Typical buyer** | Ministries, research bodies, regulated firms, local AI labs | Governments, banks, health providers | Anyone | *Our summary of how the terms are used; definitions vary by country.* ## How does a country pay for sovereign AI? In four ways, and each one puts a different party at risk. Most national plans mix them. | How the state pays | Example | What the state commits | Who carries the build risk | | ---------------------------------------- | -------------------------------------------------------------------- | -------------------------------------------------------------------------------- | ----------------------------------------------------- | | **Buys compute time as anchor customer** | EU AI Gigafactories (call opened 30 Jul 2026) | Joint EU and member-state purchase of compute access time from up to seven sites | The private consortium that builds and runs each site | | **Subsidises each GPU-hour** | IndiaAI compute (25 Mar 2026) | 38,231 GPUs from 14 providers offered at about Rs 65 per GPU-hour on average | The providers, who own the GPUs and need users | | **Buys GPUs for national centres** | Korea, Ministry of Science and ICT (30 Oct 2025) | Up to 50,000 GPUs over several years, starting with 13,000 | Mostly the state budget, through local cloud partners | | **Allocates land and power** | Nvidia’s description of regional clouds; Malaysia’s Budget 2027 plan | Sites and grid capacity for a local operator | The operator, which still has to find paying users | *Figures as each source states them. The last column is our read. Sources under References.* The first is to buy computing time as an anchor customer. In the EU’s AI Gigafactories call, opened on 30 July 2026, EuroHPC and member states will jointly buy compute access time from up to seven privately built sites. The call says the public money will act as an anchor customer and should unlock more than €20 billion of private investment. The second is to subsidise each GPU-hour. India said on 25 March 2026 that 38,231 GPUs from 14 providers were offered to start-ups and academia at an average of about Rs 65 per GPU-hour, except some high-end chips. The providers own the GPUs; the state lowers the price. If all 38,231 GPUs were rented around the clock for a year at that rate, users would pay about Rs 21.8 billion (our arithmetic: 38,231 × Rs 65 × 8,760 hours). That is a ceiling: GPUs are never fully booked, and the statement does not say how much of each hour the state pays. The third is to buy GPUs for national centres. Korea’s Ministry of Science and ICT plans up to 50,000 over several years, through its National AI Computing Center and three local clouds, starting with 13,000. The fourth costs no cash: land and power. Nvidia put it plainly on 26 August 2026: “A country or region can allocate land and power directly to a regional cloud partner in ways it never would to a foreign hyperscaler.” Where grid capacity is rationed, that can be worth more than a subsidy, but the operator still has to find users who pay. ![Sovereign AI explained: how a national AI plan gets paid, from the plan to the signed buyer, the chips and the powered hall](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/sovereign-ai-how-it-gets-paid.png) Stuck step: the signed buyer. Second Order. ## Who gets paid in sovereign AI? In order: the chip maker first, the operator second, the landlord and utility along the way. The government gets control rather than cash. The chip maker is paid when the GPUs ship. Nvidia said on 26 August 2026 that its sovereign AI business, “primarily through the regional neoclouds”, grew 35% on the quarter and more than tripled on the year. It named Yotta and Neysa in India, Firmus in Australia, GMI Cloud in Taiwan and YTL AI Cloud in Malaysia among them. Nvidia does not give a sovereign dollar figure; it reports it inside a wider group of non-hyperscaler buyers, including enterprises, that brought in $40 billion in the quarter. Our [map of GPU cloud providers](https://www.sotkn.com/explained/gpu-cloud-providers/) sets these next to the US neoclouds. Nvidia is also starting to share demand risk. On the same call it described giving a neocloud a take-or-pay commitment on part of a facility’s capacity, a revenue floor lenders can underwrite, in return for a share of revenue above that floor. “We get paid twice,” Nvidia said. Our read: that makes Nvidia a partial anchor customer for the neoclouds it sells to. [Take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/) explains why lenders want that floor. The operator is paid only when the hall is powered and the hours are sold. Singapore gives the clearest numbers. Singtel’s RE:AI, launched in October 2024, sells GPU computing to government, research and enterprise customers. Its first pilot of 512 H100 GPUs was fully used. At its Investor Day on 27 August 2026, Singtel said the first 6 megawatts of its larger build were fully contracted, with about S$600 million of revenue on three-to-five-year terms, against about S$300 million of capex. | Singtel RE:AI, Singapore | First 6 MW | Next 6 MW | | --------------------------------- | ---------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | **Status** (27 Aug 2026) | Building; fully contracted | Under negotiation | | **Long-term contracts** | About S$600 million, 3–5 years | More than S$500 million, 3–5 years | | **Capex** | About S$300 million | About S$300 million | | **Timing** | Deployed by Q4 FY27 (Jan–Mar 2027); revenue 6–9 months after deployment; payback by FY30 | From Q4 FY27 to FY28; starts only once more than 80% is contracted | | **Per megawatt** (our arithmetic) | About S$100 million contracted, or S$20–33 million a year; about S$50 million of capex | — | *Company figures from Singtel’s Investor Day deck. The last row is our arithmetic. Sources under References.* Per megawatt, that is about S$100 million of contracted revenue, or S$20 million to S$33 million a year depending on the term, and about S$50 million of capex (our arithmetic). Contracted revenue is roughly twice the capex, and Singtel targets payback by FY30, its year to March 2030. Compare a landlord. Bitdeer’s Tydal hall lease in Norway, announced on 4 August 2026, earns about $2.4 million a year per IT megawatt. Even after converting Singapore dollars, which are worth less than US dollars, a GPU cloud books several times more per megawatt than a hall lease, because it owns the chips and sells the computing (our read). It also carries far more capital and a shorter contract. [What Nxera does](https://www.sotkn.com/stack/ai-data-centers/what-does-nxera-do/) covers Singtel’s data center side, the landlord next to RE:AI. ## Where is the stuck step? At the signature. A national plan creates GPUs to fill, but someone still has to commit to paying for the hours, for years. Until then the operator waits, and the plan is a speech. Singtel builds this into its plan. Its next 6 megawatts, with more than S$500 million of customer contracts under negotiation, will only be deployed once more than 80% of the capacity is contracted. The power is secured; the GPUs wait for the signatures. Headline numbers also bundle different buyers. Of Korea’s 260,000 GPUs, the government’s share is up to 50,000, about 19% (our arithmetic). The rest are AI factories for Samsung, SK Group and Hyundai Motor Group, and NAVER Cloud’s own expansion. Those companies are real buyers, but they are buying for themselves. ![Sovereign AI in Korea: of more than 260,000 Nvidia GPUs announced on 30 October 2025, up to 50,000 are for the Ministry of Science and ICT and the rest for NAVER Cloud, Samsung, SK Group and Hyundai Motor Group](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/sovereign-ai-korea-gpus-by-buyer.png) Korea’s GPUs by buyer. Second Order arithmetic from Nvidia’s 30 Oct 2025 figures. Neocloud filings show the same gap. Nscale, which pitches itself as a sovereign-capable platform, said in its 18 September 2026 filing that it expects “a meaningful and growing portion” of revenue from sovereign customers. But its largest customer supplied 52% of revenue in the first half of 2026 and 73% in 2025, and it says hyperscale customers are currently the largest source of demand. It also disclosed that OpenAI withdrew from its Stargate UK and Stargate Norway partnerships in April 2026, with Microsoft taking the extra Norway capacity. Our read: sovereign demand is real, but much of the money paying neoclouds today still comes from a hyperscaler or a lab. Our [Nscale Glomfjord Wire](https://www.sotkn.com/wire/nscale-glomfjord-data-center-six-months-late-35b-ipo-wire/) tracks what that means on the ground. Some plans have no buyer named at all yet. Malaysia’s Budget 2027 speech in October 2026 said a sovereign AI cloud was being planned to keep Malaysians’ data under national control, but gave no project cost, location, operator or timetable. Our [Malaysia note on booked but unused data centre power](https://www.sotkn.com/markets/malaysia/malaysia-data-centre-power-use-october-2026/) shows why the operator choice matters there. Public money can also arrive late. The EU says the bulk of its public funding for the gigafactories will come from its next long-term budget, which runs from 2028 to 2035\. The anchor customer exists on paper; much of its cash comes later. ## How long does sovereign AI take? Years, not months, from the speech to the first billed GPU-hour. The clocks differ by route. | Programme | Started | Where it stands | First compute | | ------------------------------- | ----------------------------- | ------------------------------------------------------------ | -------------------------------------------------------------------------------- | | **EU AI Gigafactories** | InvestAI announced 9 Feb 2025 | Formal call 30 Jul 2026; bids close 12 Nov 2026 | Selection early 2027, then operations within 18 months, so 2028 (our arithmetic) | | **Singtel RE:AI**, Singapore | Launched Oct 2024 | 512-GPU pilot fully used; first 6 MW fully contracted | 6 MW by Jan–Mar 2027; revenue 6–9 months later | | **IndiaAI compute** | Mission launched 2024 | 38,231 GPUs onboarded by 25 Mar 2026 | Live, sold by the hour | | **Korea national GPUs** | Announced 30 Oct 2025 | Initial 13,000 GPUs through NAVER Cloud, NHN Cloud and Kakao | Up to 50,000 over several years | | **Malaysia sovereign AI cloud** | Budget 2027, Oct 2026 | No cost, site, operator or timetable given | Not stated | *Dates as each source states them. Sources under References.* The EU route is the slowest because it runs through public procurement. InvestAI was announced in February 2025, the formal call opened 17 months later, and the first operations are due about 18 months after selection in early 2027 (our arithmetic). Subsidised hours, as in India, are faster because the providers already own the GPUs, and a telco with secured power, as in Singapore, moves fastest once it has signatures (our read). After the signature, the usual data center clocks apply: a powered hall, delivered GPUs, then commissioning. [How long it takes to build an AI data center](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/) covers that part. In hot climates such as Singapore’s, efficiency rules are part of the clock too; see [PUE](https://www.sotkn.com/explained/pue-explained/). ## What sovereign AI does not tell you A sovereign label says where the computing sits and who controls it, not whether the GPUs are busy. A half-idle national cluster earns half the money; see [GPU utilization](https://www.sotkn.com/explained/gpu-utilization-explained/). It does not settle the depreciation clock either. If a contract ends before the chips have paid for themselves, the operator carries the gap. Singtel’s first 6 megawatts are due by March 2027 and target payback by FY30, about three years later and inside its three-to-five-year terms (our arithmetic); see [GPU depreciation](https://www.sotkn.com/explained/gpu-depreciation-explained/), and [who finances GPU clusters](https://www.sotkn.com/explained/who-finances-gpu-clusters/) for how lenders read the contract. ## What to check next Bids for the EU AI Gigafactories, closing on 12 November 2026, and the selection in early 2027\. Whether Singtel’s first 6 megawatts are deployed by March 2027 and its next 6 megawatts reach 80% contracted. Nvidia’s next quarterly comment on sovereign growth. Updates to Nscale’s filing on how much revenue comes from sovereign customers. Whether Malaysia names an operator, a cost and a date. Our [Singapore note on the 200 MW data centre licence](https://www.sotkn.com/markets/singapore/singapore-data-centre-licence-october-2026/) follows the power side in Singapore. ## Common misconceptions - **People assume sovereign AI means government-owned data centers.** Actually most of it is run by regional neoclouds and telcos; the state buys hours, subsidises them, buys GPUs or allocates land and power. - **People assume sovereign AI means home-grown chips.** Actually the GPUs are usually imported; Korea’s plan is for more than 260,000 Nvidia GPUs. - **People assume a national AI plan is a contract.** Actually the EU’s gigafactory bids only close on 12 November 2026, and Malaysia’s plan names no operator, cost or date. - **People assume sovereign demand is what fills neoclouds today.** Actually Nscale’s largest customer supplied 52% of its first-half 2026 revenue, and it says hyperscalers are the largest source of demand. ## FAQ **What is sovereign AI?** Sovereign AI is AI computing power, data and models that a country keeps inside its own borders and under its own laws. It usually means GPUs in local data centers, run by a local operator, with data and logs that stay in the country. **What is the difference between sovereign AI and sovereign cloud?** Sovereign cloud keeps ordinary IT, such as storage and databases, under local law. Sovereign AI applies the same idea to GPU computing, the data that trains and runs models, and often the models. **Why do countries want sovereign AI?** For control. Governments want sensitive work, and models trained on local languages, to run on machines they can regulate and that a foreign supplier cannot switch off. **Who pays for sovereign AI?** Usually a mix. Governments buy compute time as anchor customers, subsidise GPU-hours, buy GPUs for national centres or allocate land and power. Operators and their lenders fund most of the build. **Does Singapore have sovereign AI?** Singapore’s clearest example is Singtel’s RE:AI, which Singtel describes as a sovereign GPU-as-a-service provider selling to government, research and enterprise customers. At its August 2026 Investor Day Singtel reported about S$600 million of contracts on the first 6 megawatts, on three-to-five-year terms. **How big is Nvidia’s sovereign AI business?** Nvidia does not break out a dollar figure. On its 26 August 2026 call it said the business, mostly sold through regional neoclouds, grew 35% from the previous quarter and more than tripled from a year earlier. It sits inside a wider group of non-hyperscaler buyers that brought in $40 billion in the quarter. **Is this a list of stocks to buy or sell?** No. Second Order Explained pages explain how a mechanism works and where the money moves. They do not give buy or sell calls. Related: [neocloud](https://www.sotkn.com/explained/neocloud-explained/); [neocloud vs hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/); [GPU-as-a-service](https://www.sotkn.com/explained/gpu-as-a-service-explained/); [take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/); [GPU cloud providers](https://www.sotkn.com/explained/gpu-cloud-providers/); [AI data center](https://www.sotkn.com/explained/what-is-an-ai-data-center/). ## Stack doors Layer: [Neoclouds](https://www.sotkn.com/stack/neoclouds/). Operators: [Nebius](https://www.sotkn.com/stack/neoclouds/what-does-nebius-do/), [CoreWeave](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/), [Nxera](https://www.sotkn.com/stack/ai-data-centers/what-does-nxera-do/). Also: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/), [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/), [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Markets: [Singapore](https://www.sotkn.com/markets/singapore/), [Korea](https://www.sotkn.com/markets/korea/), [Malaysia](https://www.sotkn.com/markets/malaysia/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when neocloud contracts, GPU deals and national AI plans move. **How we check this** Nvidia’s sovereign AI figures and its description of regional clouds come from its 26 August 2026 earnings call. Korea’s GPU plan comes from Nvidia’s 30 October 2025 release. The EU figures come from the EuroHPC Joint Undertaking’s 30 July 2026 call notice and the European Commission’s AI Gigafactories page. India’s figures come from the government’s 25 March 2026 statement to Parliament. Singtel’s RE:AI figures come from its 27 August 2026 Investor Day deck, Nscale’s from its 18 September 2026 filing with the US Securities and Exchange Commission, and Malaysia’s plan from reporting on its Budget 2027\. Each was re-checked on 18 Oct 2026\. Company figures are the companies’ own. Where we work something out ourselves it is labelled “our arithmetic”, and where we interpret it is labelled “our read”. This page explains a mechanism. It does not give buy or sell calls. **Last reviewed:** 18 Oct 2026 · **Next review:** after the EU AI Gigafactory selection in early 2027 or Nvidia's next quarterly results Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [NVIDIA — Q2 fiscal 2027 earnings call, corrected transcript (26 Aug 2026)](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=sotkn.com) — sovereign AI growth through regional neoclouds; land and power quote; take-or-pay revenue-share structure 2. [NVIDIA Newsroom — NVIDIA, South Korea government and industrial giants build AI infrastructure (30 Oct 2025)](https://nvidianews.nvidia.com/news/south-korea-ai-infrastructure?ref=sotkn.com) — 260,000 GPUs by buyer; ministry share up to 50,000, starting with 13,000 3. [EuroHPC Joint Undertaking — AI Gigafactories call launched (30 Jul 2026)](https://www.eurohpc-ju.europa.eu/eurohpc-joint-undertaking-launches-ai-gigafactories-call-2026-07-30%5Fen?ref=sotkn.com) — anchor-customer model; bids close 12 Nov 2026; selection early 2027; operations within 18 months; 2028–2035 budget 4. [European Commission — AI Gigafactories (timeline page, 2026)](https://commission.europa.eu/topics/competitiveness/competitiveness-coordination-tool-projects/ai-gigafactories%5Fen?ref=sotkn.com) — InvestAI announced 9 Feb 2025; €20 billion; first construction 2027 5. [Press Information Bureau, India — Government supporting organisations to develop sovereign foundational model (25 Mar 2026)](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2245063&ref=sotkn.com) — 38,231 GPUs from 14 providers; about Rs 65 per GPU-hour; 12 sovereign model builders 6. [Singtel — Investor Day 2026, Digital InfraCo deck (27 Aug 2026)](https://cdn1.singteldigital.com/content/dam/singtel/investorRelations/stockExchange/2026/Investor%20Day%202026%20-%20DICo%20%28for%20SGX%29.pdf?ref=sotkn.com) — RE:AI contracts, capex, payback and the 80% contracted rule 7. [Nscale — Form S-1 registration statement (18 Sep 2026)](https://www.sec.gov/Archives/edgar/data/2110365/000119312526395475/ck0002110365-20260918.htm?ref=sotkn.com) — customer concentration; sovereign customer requirements; Stargate UK and Norway changes 8. [EdgeProp Malaysia — Budget 2027: data-centre fund highlighted; TNB to invest RM15b in grid (Oct 2026)](https://www.edgeprop.my/content/1917782/budget-2027-data-centre-fund-highlighted-tnb-invest-rm15b-grid?ref=sotkn.com) — sovereign AI cloud planned with no cost, site, operator or timetable 9. [Bitdeer — $4.7 billion, 16-year AI/HPC lease for Tydal, Norway (4 Aug 2026)](https://www.sec.gov/Archives/edgar/data/1899123/000121390026085655/ea030075201ex99-1.htm?ref=sotkn.com) — about $2.4 million a year per IT megawatt, the landlord comparison 10. [Wikimedia Commons — BSC MareNostrum 5, photo by Steve Jurvetson, CC BY 4.0 (taken 14 Jul 2025)](https://commons.wikimedia.org/wiki/File:BSC%5FMareNostrum%5F5%5F%2854686085845%29.jpg?ref=sotkn.com) — featured image base photo S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)