The Commodification of AI Sovereignty: Lessons from the Fight for Sovereign Oil
Rui-Jie Yew, Kate Elizabeth Creasey, Taylor Lynn Curtis, Suresh Venkatasubramanian
cs.CY
2026-01-17
NVIDIA, AWS, Azure and Google now sell "sovereign" factories, clouds and models. A Brown-Mila paper maps this onto oil concessions and asks who actually absorbs the technology.
Governments now write "sovereignty" into AI policy. The EU wants a sovereign stack. India and Singapore are building national language-model ecosystems. Africa's AI declaration includes sovereign compute. At the same time, NVIDIA sells "sovereign AI factories," AWS, Azure, and Google Cloud sell "sovereign clouds," and several vendors sell "sovereign" language models. A contested political value is being priced as a product.
The practical risk is that vendors get to define the term. Whoever defines sovereignty also decides whether the buyer receives data residency, an encryption control, or actual technology absorption. This cs.CY paper from Brown's computer science and history departments and Mila unpacks that vocabulary by setting it against Anglo-American oil concessions in the twentieth-century Middle East, and asking which older vulnerabilities the buyer is reopening.
The authors split sovereignty with Krasner's four-part scheme, then walk the AI stack against oil history.
The corpus is NVIDIA blog posts tagged "sovereign" from 2019 through December 2025, sovereign-cloud material from AWS, Azure, and Google, and multilingual-model narratives from Meta, Cohere, and Hugging Face. The oil comparison stays with firms rooted in the US and Britain before nationalization: Aramco when it was still an American company, and the Anglo-Persian Oil Company, later BP.
Two analogies do the work. Oil and AI are both bound to growth and national ambition. Concession contracts once gave foreign firms control over territorial resources; "sovereign AI" contracts now give vendors control over the stack. The authors mark the mismatch themselves: oil companies went to buy a resource in place, while AI firms present themselves as offering one, and oil is geographically bound in a way AI is sold as something every country can "grow." That posture is the point of the paper. Offering a resource can hide how data, energy, and labor are extracted from the buyer.
This is not a benchmark paper. The yield is a stack-level map plus a handful of checkable figures.
| Layer | What is sold | Oil parallel | Vulnerability named |
| Hardware and infrastructure | NVIDIA "sovereign AI factories"; OpenAI for Countries; Stargate Norway | Aramco's "partnership" inside Saudi nation-building | Land and power are opened; chips, models, and training stay vendor-owned |
| Cloud | AWS/Azure/GCP EU sovereign clouds; the three hold about two-thirds of global cloud compute | "The only thing Iraqi about the company was the location of the oil"; Anglo-Persian royalty formulas | Privacy tech is recast as "sovereignty controls"; vendors can still cut off service and remain under the US CLOUD Act |
| Models | Multilingual "sovereign" LLMs; "no language left behind" | Oil labor stratified by race and nationality | Annotation extraction is covered by cultural-preservation talk |
| Item | Figure | ||
| NVIDIA sovereign-AI partners | More than ten countries, including India, Denmark, Thailand, New Zealand, Switzerland, Japan | ||
| India training | Over 100,000 AI developers and over 2,000 NVIDIA Inception companies, many serving on models such as Nemotron-4-Mini-Hindi-4B | ||
| NeMo production license | $4,500 per GPU per year, or $1 per GPU-hour in the cloud, billed by GPU count | ||
| Scale annotation pay | $21.55/hour for German, $1.43/hour for Telugu | ||
| AWS European Sovereign Cloud | Planned €7.8 billion investment; claimed €17.2 billion GDP contribution to Germany through 2040 and 2,800 full-time jobs a year | ||
| Iraq Petroleum profits | Anglo-Persian and Shell together took 43.75% |
NVIDIA writes sovereign AI as "a nation's capabilities to produce artificial intelligence using its own infrastructure, data, workforce, and business networks." Jensen Huang says every country needs it and should "own the production of their own intelligence." It sounds like localization. In India the trained developers and startups sit on NVIDIA models and tooling. In Japan the company has lobbied for more power to "generate intelligence." Aramco's rename to Arabian American Oil Company told the same partnership-and-modernization story, in order to keep the concession.
The cloud layer is closer to the royalty formula. France's digital minister, at a 2025 cloud summit, called US-EU partnerships "sovereignty-washing": data nominally stays in the EU, operations stay with US firms. NVIDIA sells sovereign AI to governments and self-sovereign AI to enterprises as a hedge against government scrutiny. Both products can be sold at once. The collision is talked away.
The model layer has a labor ledger. "No language left behind" needs annotators. Scale AI has cut entire countries off its annotation platform overnight, with no explanation. The paper sets this next to racialized job ladders in Gulf oil camps: locals in labor, Indian migrants in trades, Europeans and Americans in authority.
The two recommendations are rulers, not blueprints. Technology transfer should separate "can operate the vendor's tools" from "can develop independently." After nationalization, Aramco still used an In-Kingdom Value Add Score on local training, R&D, and salaries. NVIDIA's technical guide to sovereign models runs NeMo through the whole pipeline; the stack is open-source to try, licensed per GPU to productionize, which ties clients to a vendor paradigm through chip-market position. Data should be traced as a resource. The 1962 UN resolution on permanent sovereignty over natural resources required that resources serve national development and the well-being of the people. Huang calls data a natural resource and argues it should not be exported, refined, and sold back, but that framing still treats data as belonging to the place it was produced. Data generated by a community using AI services, and the electricity spent to process it, need not serve that community. Differential privacy is used by vendors to weaken dataset-transparency rules; the same methods can count dataset contents under privacy constraints.
Anyone packing a "sovereign" offering for a government or a large buyer can treat this as a checklist, not an architecture. Three questions are enough. Do the people being trained only know one vendor's stack? Is data written into the contract as an asset that keeps being extracted? Are cutoff rights and CLOUD Act-style extra-territorial access on the risk register?
For policy, the paper splits buying a sovereign cloud from having sovereignty. The EU and the UK were once on the exporting side of oil concessions; they are now among the loudest buyers of digital sovereignty. The authors note that those governments' older record on oil and resource sovereignty may be more informative than today's cloud marketing.
This is discourse analysis plus historical analogy, not an engineering result. It will not add a training trick. What it can do is put an old concession question in front of the next "sovereign AI factory": did technology transfer happen, and is data being traced as a resource?
The authors say the analogy is imperfect. Oil companies went to buy a resource in situ. AI companies present themselves as offering one. Oil rights bind to geology; AI is sold as something every country can grow. The relation looks inverted. Extraction can hide inside delivery.
The paper also refuses a recipe and refuses to score whether joining a sovereign-AI program is worthwhile. The recommendations stop at how to measure transfer and how to trace data. There is no audit procedure and no contract clause. The evidence is NVIDIA blogs and cloud marketing, not actual contracts, export controls, or a comparison with homegrown stacks. The French minister's "sovereignty-washing" line is a political charge; the paper does not independently establish who holds operational control inside EU sovereign clouds.
The closing glance at space data centers and blurring firm-state lines, and the reminder that "sovereignty is not a gift," points somewhere. The evidence there is thin. A historical lens opens a space for criticism. It does not replace an audit of a procurement contract.