AI Sovereignty Around the World

Session 16 · Mon, Nov 16

POLISCI 158 · AI & Politics · Stanford · Fall 2026

Roadmap

  1. In this session
  2. Questions we’ll tackle
  3. Before class
  4. After class

In this session

Most countries will use AI that is built, hosted and governed somewhere else. We treat AI as a global distributive shock and map where countries sit on a dependency stack of chips, compute, energy, data and frontier models. We then ask what “AI sovereignty” would take, using four tests: where inference runs, who holds the model weights, whether national law reaches AI providers, and whether models work in the national language. Cases from Côte d’Ivoire, the Gulf, India, Europe, Southeast Asia and Brazil show that each “sovereign AI” deal buys some of these and none buys all. In class, groups audit one country each. Last, we look at how AI rules travel across borders, through learning, emulation, competition and coercion, the Brussels effect, and US and Chinese efforts to export whole AI stacks, and why state capacity limits what any of them deliver.

Questions we’ll tackle

  • Why does AI work as a global distributive shock, and will it widen or narrow the gap between rich and poor countries?
  • Which layers of the AI stack does a given country depend on others for: chips, compute, energy, data and the law that covers it, or frontier models?
  • What would AI sovereignty require in practice, and when a country signs a “sovereign AI” deal, which of the four tests does it actually pass, and at what cost?
  • Through which channels do AI rules and AI stacks travel from one jurisdiction to another, and has the EU AI Act travelled?
  • Why do rules on paper depend on a government’s capacity to use, audit and enforce them?

Before class

Read

  • Priority: World Development Report 2026: The Promise of Artificial Intelligence, World Bank (2026). Start with the overview and the discussion of AI sovereignty in Chapter 6; note why the report argues that control by a few companies creates dependency risks, “but pursuing ‘AI sovereignty’ comes at a high cost and with uncertain benefits.”
  • Policy and Governance, The 2026 AI Index Report, Stanford Institute for Human-Centered AI (2026). Look at the maps of national AI strategies and the chart of state-backed AI supercomputers by region, and ask who has the infrastructure behind sovereignty claims.
  • The Token Tax: Systematic Bias in Multilingual Tokenization, Jessica M. Lundin, Ada Zhang, Nihal Karim, Hamza Louzan, Victor Wei, et al. (2026), AfricaNLP 2026 Workshop. Read Figure 1 for how far the tested models’ accuracy drops from English to African languages, and why the authors link it to tokenization.

Explore

  • Data on GPU Clusters, Epoch AI (ongoing). A public dataset of large AI supercomputers, with location and ownership; filter by country and compare public with private owners.
  • Government AI Readiness Index, Oxford Insights (2025 edition). Country scores on six pillars, from policy capacity and AI infrastructure to governance and public sector adoption; pick one country outside the US and China and see where it falls short.

After class

Key takeaways

  • AI works as a global distributive shock, and most countries are buyers rather than builders. The 2026 AI Index counts 5,427 data centers in the United States, more than ten times any other country, and 85 public or public-private AI supercomputers in China against 8 in Latin America and the Caribbean. Language is part of the gap: in one benchmark study of 2024-era models, African languages trailed English by about 30 percentage points on average.
  • AI sovereignty is the ability to govern AI’s national impact. Four tests make a sovereignty claim concrete: where inference runs, who holds the model weights, whether national law reaches AI providers, and whether models perform in the national language. Each case we read targets some tests and none covers all; the World Bank warns that pursuing sovereignty comes “at a high cost and with uncertain benefits.”
  • Sovereignty labels can hide new dependencies. Côte d’Ivoire’s “sovereign national data center” is part of a $170 million package financed by the US Export-Import Bank, which approved $66.1 million for the data center itself, and built by a US firm.
  • AI rules travel through learning, emulation, competition and coercion. The EU AI Act’s reach looks partial so far, while the United States and China each promote exports of whole AI stacks. National AI strategies are spreading fast (in 2024 more than half of newly adopted strategies came from emerging economies), but what they deliver depends on state capacity to use, audit and enforce.

Go further

About this deck

Adapted from POLISCI 157 Session 7B (Aug 5, 2026) for the dependency stack and the four sovereignty tests; the cases, the sovereignty audit and policy diffusion are new for 158.

Curated by Simone Paci, Stanford University. Drafted with AI assistance (Claude) from the instructor’s course materials and reviewed by the instructor before publication; every external link is checked before each update.