AI Sovereignty Around the World

POLISCI 158 · AI & Politics · Session 16 · Monday, November 16, 2026

Simone Paci

Roadmap

  1. Part I: Who wins, who depends?
    • AI as a global distributive shock -> who gains
    • Place a country on the dependency stack
  2. Part II: What would sovereignty take?
    • Four tests: compute · models · jurisdiction · performance
    • What a “sovereign AI” deal buys, at what cost
  3. Part III: How do AI rules travel?
    • Diffusion channels: coercion, competition, learning, emulation · Brussels effect · stack exports
    • State capacity limits what rules deliver

Last week two powers. This week everyone else?

US leads China leads
compute · chips · money · frontier models open models · robots · power

What about the other countries?

Session 14 scoreboard

Will AI widen or narrow the gap between rich and poor countries?

Widen
Narrow
No change
Not sure

Show of hands · we vote again at the end

Part I: Who wins and who loses from AI, worldwide?

Most countries will use AI. Few will make it.

Who is most exposed?

ILO Working Paper 140, Figure 20, left column: share of employment in occupations potentially exposed to generative AI, by income group, stacked by exposure gradient; totals: World 24%, High income 34%, Upper-middle income 25%, Lower-middle income 20%, Low income 11%

ILO (2025), WP 140, Fig. 20 · CC BY 4.0

Jobs exposed to GenAI, by country income group: high 34% · upper-middle 25% · lower-middle 20% · low 11%

Exposure ≠ gain: “Infrastructure constraints (electricity, broadband), insufficient digital skills, the cost of technology”

Exposure = AI overlaps job tasks, “not the actual impact on employment”

ILO Working Paper 140, Gmyrek et al., May 2025 (Fig. 20)

Why call AI a distributive shock?

Access -> productivity gains · no access -> falls further behind

digital divide AI access gains (productivity) no access -> gap widens

“a potential widening of the digital divide and global income disparity”

An IMF projection from 2024, not a measured outcome

Cazzaniga et al., IMF Staff Discussion Note SDN/2024/001, Jan 2024

Who is actually using it?

Microsoft world map: AI diffusion by economy, June 2026; AI user share in four bands (under 10%, 10–20%, 20–30%, over 30%) plus insufficient data; labelled: Norway 49.4%, Ireland 49.9%, France 49.6%, UAE 73.3%, Singapore 64.3%

Share of working-age people using GenAI, Q2 2026: Microsoft’s “Global North” 28.8% vs “Global South” 16.2%

  • Gap: 9.8 pp (H1 2025) -> 12.6 pp (Q2 2026)
  • Top: UAE 73.3% · Singapore 64.3% · US 21st (33.0%)

Vendor telemetry · North/South grouping not defined by Microsoft · use ≠ gain

Microsoft AI Economy Institute, Jan 8 and Sep 21, 2026

What sits under every AI answer?

Frontier models Data + jurisdiction Energy Compute (data centers, cloud) Chips

Where are the data centers?

AI Index 2026 map: global distribution of data centers, 2025, by country in bands from 1–9 to 530+; the United States is in the darkest band; most of Africa shows no available data. Source: Cloudscene, 2025

US: 5,427 data centers, more than 10× any other country (2025)

  • Germany 529 · UK 523 · China 449 · Brazil 197 · India 153

All data centers, not AI only · count ≠ capacity · white = no data, not zero

Stanford HAI, AI Index 2026, ch. 1 (Cloudscene data)

Who has state-backed AI supercomputers?

AI Index 2026 line chart: number of public or public-private AI supercomputers, cumulative, 2010–25; 2025 end labels: China 85, Europe and Central Asia 44, North America 41, East Asia and Pacific 27, Latin America and Caribbean 8, Middle East and North Africa 3, South Asia 2. Source: Epoch AI, 2026

Public or public-private AI supercomputers, 2025: China 85 · Europe & Central Asia 44 · North America 41 · East Asia & Pacific 27 · Latin America & Caribbean 8 · Middle East & North Africa 3 · South Asia 2

Proxy for frontier training capacity · planned systems excluded · no sub-Saharan Africa line

Stanford HAI, AI Index 2026, ch. 8, Fig. 8.3.1 (Epoch AI data)

Where can you rent AI compute?

Lehdonvirta, Wu and Hawkins, Figure 2: world map of the global compute divide based on public cloud GPU compute; Compute North in blue (including US, Canada, Western Europe, China, India, Japan, Korea, Australia), Compute South in red (including Mexico, Brazil, Chile, Argentina, South Africa, Saudi Arabia, Turkey, Indonesia), Compute Desert in beige (most of Africa, Central Asia, Russia)

Figure design: Juuso Koponen

39 countries host any public cloud region; 30 with GPUs

  • Compute North: US, China + 15 others (all Global North but India)
  • Compute South: 13 countries, deployment-grade chips
  • Everyone else: “Compute Desert”

As of early 2024 · public cloud only, six hyperscalers

Lehdonvirta, Wu & Hawkins, AIES 2024

And the chips?

SIA and BCG 2024, Exhibit 2: semiconductor industry value added by activity and region, 2022, percent; design and EDA dominated by the USA, equipment by the USA, Japan and EU, wafer fabrication and assembly concentrated in East Asia

Same chokepoint as Session 14

  • “TSMC is a single point of dependency in the global AI supply chain”

Value added by region, 2022

SIA & BCG, May 2024; Stanford HAI, AI Index 2026, ch. 1

Where do notable models come from?

AI Index 2026 map: number of notable AI models by geographic area, 2003–25 (sum), in bands 1–10 to 181–640; the United States is in the darkest band, China in 61–180; Latin America and Africa are uncoloured. Source: Epoch AI, 2026

2025: US 59 · China 35 · South Korea 8

2003–25: no notable model on Epoch’s list from Latin America or Africa

Counted by author affiliation · curated list, not a census

Stanford HAI, AI Index 2026, ch. 1 (Epoch AI data)

Which layer can a country realistically own?

Frontier models Data + jurisdiction Energy Compute (data centers, cloud) Chips ? ?

Most countries: buyers, not builders

Chips, cloud compute, frontier models -> concentrated in a few places

Slides 10–14

Part II: What would it take for a country to rule its own AI?

Every government now says “sovereign”. Few mean the same thing.

What is AI sovereignty?

Sovereignty = ability to self-rule AI sovereignty = ability to govern AI’s national impact AI Index: “a state’s capacity to act deliberately and make independent decisions over … AI systems within its jurisdiction”

“AI sovereignty is an ambiguous concept” (World Bank)

Stanford HAI, AI Index 2026, ch. 8 §8.3; World Bank, WDR 2026, ch. 6, fn. 15

How would you test a sovereignty claim?

Computewhere does inference run?
Modelswho holds the weights?
Jurisdictiondoes national law reach the provider?
Performancedoes it work in the national language?

One framework of several · AI Index uses five layers (infrastructure, data, model, application, talent)

Test 1, compute: must the data stay at home?

AI Index 2026 line chart: data localization measures by region, cumulative, 2000–24; 2024 end labels: East Asia and Pacific 77, Sub-Saharan Africa 71, Europe and Central Asia 66, Middle East and North Africa 44, Latin America and Caribbean 36, South Asia 24, North America 3. Source: Ferracane et al., 2026

Data localization measures, cumulative to 2024: East Asia & Pacific 77 · sub-Saharan Africa 71 · Europe & Central Asia 66 · North America 3

  • Local storage ≠ local law: “extraterritorial reach of a cloud provider’s home-country laws”

Counts measures, not strictness · regional totals hide country variation

Stanford HAI, AI Index 2026, ch. 8 (data: Ferracane); World Bank, WDR 2026, ch. 6

Test 2, models: whose weights are under the national model?

Fine-tune an open model -> national model, foreign base

  • One Singapore program, SEA-LION: bases from Alibaba (China), Google and NVIDIA (US), Swiss AI (Switzerland)
National model ← Base model Base origin
Qwen-SEA-LION-v4.5-27B-IT (AI Singapore) Qwen3.6-27B China (Alibaba)
Gemma-SEA-LION-v4-4B-VL (AI Singapore) Gemma 3 4B US (Google)
Nemotron-SEA-LION-v4.8-120B (AI Singapore) Nemotron 3 Super 120B US (NVIDIA)
Apertus-SEA-LION-v4-8B-IT (AI Singapore) Apertus 8B Switzerland
Sarvam-M (India) Mistral Small 3.1 France
Sahabat-AI Gemma2 9B (Indonesia) Gemma 2 US (Google)

Base models as listed on Hugging Face model cards, 2024–26

Hugging Face model cards

Test 3, jurisdiction: can a state make a platform comply?

Brazilian Supreme Court (STF) news page, 02/09/2024: STF confirma decisão que suspendeu o X, antigo Twitter, em todo o país; Por unanimidade, a Primeira Turma referendou a medida adotada pelo ministro Alexandre de Moraes; photo of the court building

Brazil v. X, 2024

  • Aug 30: Justice Moraes suspends X nationwide; Sep 2: panel upholds, unanimous
  • Oct 8: X back after compliance + R$28.6M in fines
  • “O Brasil é soberano.” Communications Minister

A platform case, not an AI one

STF, Sep 2, 2024; CNN Brasil and TELETIME, Oct 8, 2024

Does the model speak your language?

“African languages trail English by 30 percentage points … on average”

  • 16 African languages, 5 subjects (AfriMMLU)

The Token Tax, Figure 1 panel (a), Performance Gap block: accuracy gap in percentage points between English and the average of African languages, for eleven models across Total, Math, Facts, Geography, Economics and Law; Total gaps range from 15 (o1-preview) to 40 (Qwen2.5-32B, phi-4)

Models from 2024–early 2025 · reasoning models narrow the gap

Lundin et al., “The Token Tax,” AfricaNLP 2026

Test 4, performance: who builds AI for their own languages?

State platformIndia · Bhashini36+ languages350+ models
GrassrootsAfrica · Masakhane30+ languagesMT benchmarks

Bhashini figures self-reported · Masakhane benchmarks date from 2020

PIB, Feb 10, 2026; Nekoto et al., Findings of EMNLP 2020

What is Côte d’Ivoire buying?

Architectural rendering, aerial view: a white data center campus among trees, with a large hall, a curved administrative building, solar-panel parking canopies and access roads

Rendering via TechAfrica News (Aug 4, 2026); attributed to Cybastion

$170M package

$66.1M data center

$170M package, US EXIM-financed: national data center + government digitization platform + border surveillance system

  • $66.1M EXIM guarantee for the data center (Aug 2025): “displaces competition from the PRC”
  • 1.73 MW IT · launch targeted 2027

How the $170M splits is not public

TechAfrica News, Jul 17, 2026; US EXIM, Aug 21, 2025; DCD, Sep 9, 2026

Can you buy sovereignty with capital?

OpenAI · USOracle · USNVIDIA · US Cisco · USSoftBank · Japan NVIDIA · USMistral · France UAE Saudi Arabia 1 GW planned · 200 MW first phase announced (May 2025) US licence-free chips (Jul 10, 2026) 84 MW, ~27,000 GPUs; 50 MW for a US client (Oct 2026)

UAE · Stargate UAE: 1 GW planned, 200 MW first phase announced for 2026

  • OpenAI, G42, Oracle, NVIDIA, Cisco, SoftBank · “in coordination with the U.S. government”
  • Licence-free US chips for approved firms since Jul 10, 2026

Saudi Arabia · HUMAIN (Public Investment Fund) + NVIDIA chips · Mistral for Arabic models

Announcements, not delivered capacity · 200 MW live status unconfirmed, Nov 16

OpenAI, May 22, 2025; Reuters via Al-Monitor, Jul 2026; NVIDIA, May 13, 2025; Mistral, Aug 24, 2026

Can the state build it instead?

PIB backgrounder, Democratising AI in India, February 10, 2026, Key Takeaways box: 38,000+ GPUs are available at ₹65/hour; 5G covers 99.9% of districts; AIKosh offers 7,500+ datasets and 273 models

IndiaAI Mission (Mar 2024): ₹10,371.92 crore over five years

  • 38,000 GPUs onboarded (privately owned, subsidised) at ₹65/hour
  • 12 startups selected to build foundation models
  • Output: Sarvam 30B/105B, trained from scratch on IndiaAI compute

Government self-report · onboarded ≠ state-owned

PIB backgrounder, Feb 10, 2026; Sarvam AI, Mar 6, 2026

Can countries pool their way to sovereignty?

EuroHPC infographic, EuroHPC AI Factories: 19 AI Factories, 15 new AI-optimised supercomputers, 13 AI Factory Antennas, 3 access modes, 40+ industrial sectors; a ring of AI Factory and Antenna names with country codes around a central AI brain graphic

EuroHPC: 19 AI Factories + 13 Antennas, shared across member states

  • Up to 7 AI Gigafactories: call open, >€20B private investment expected

National champion: Mistral (France), open weights

Gigafactories chosen in early 2027 · call deadlines Dec 3, 2026 (Lot 2), Feb 4, 2027 (Lot 1)

EuroHPC JU, Jul 30, Sep 1 and Nov 5, 2026; Mistral, 2026

Host others’ compute, or build your own models?

Malaysia hosts

Operational supply capacity, Aug 2026 · 45 data centers Need by 2030 6,041 MW Johor 4,428 8,391 MW

Johor: “$35 billion beneficiary” of Singapore’s 2019–22 moratorium

Peninsular data centers need 8,391 MW by 2030 · Johor 4,428 MW

Permit rule on US AI chip exports and transits (Jul 2025)

Singapore builds

SEA-LION: open, state-funded model · S$70M initiative

Burmese
Indonesian
Filipino
Malay
Tamil
Thai
Vietnamese

Two neighbours, two strategies · Malaysia’s own models not covered here

Reuters via Khaleej Times, Jul 26, 2026; PETRA via The Vibes, Oct 20, 2026; MITI, Jul 14, 2025; IMDA, Dec 4, 2023; AI Singapore model cards

What does each deal buy?

Case Compute Models Jurisdiction Performance
Côte d’Ivoire ●
Gulf (UAE, Saudi Arabia) ● ● ●
India ● ● ●
Europe ● ●
Brazil (X case) ●
Southeast Asia ● ● ●

Each buys some tests. None buys all.

● = test the flagship move targets, not how well it does

Slides 21–28

Is sovereignty worth the price?

“pursuing ‘AI sovereignty’ comes at a high cost and with uncertain benefits”

  • “for most countries, the middle path between sovereignty and dependence is a more viable strategy”

World Development Report 2026, chapter 6 opener, AI's Political Impact: Reshaping Power Within and Across Countries; Main messages box, first message: AI dependence and AI sovereignty are both costly; diversify AI suppliers, ensure systems work together, use open models, and join with similarly situated countries to strengthen bargaining power

World Bank, WDR 2026 web page (ch. 6 summary) and ch. 6

Whose stack is under the “sovereign” label?

"sovereign" Frontier models Data + jurisdiction Energy Compute Chips Netherlands / KoreaMistral investors USexport credit, Côte d'Ivoire USclient, HUMAIN 50 MW

“Narratives around AI sovereignty are often driven by foreign technological companies”

  • Côte d’Ivoire’s data center: US export credit, pitched against China
  • Mistral: rounds led by ASML (Netherlands), Samsung (Korea)
  • HUMAIN: 50 MW reserved for a US client

World Bank, WDR 2026, ch. 6, fn. 17; US EXIM, Aug 2025; Mistral, Sep 2025 and Sep 2026; The National/Reuters, Oct 22, 2026

Sovereignty audit: where is your country weakest?

Assigned country cards:

India
UAE
France
Côte d’Ivoire
Brazil
Indonesia
  1. Alone · 1 min: read your card
  2. Groups of 3–4 · 6 min: one auditor per test -> strong / partial / weak; mark the stack layers it owns
  3. Report · 5 min: weakest test + one fix you’d fund
Compute
Models
Jurisdiction
Performance

You audit the country; nobody speaks for it · written answers welcome

Same tests, different weak spots?

Country Compute Models Jurisdiction Performance Fix to fund
India
UAE
France
Côte d’Ivoire
Brazil
Indonesia

Part III: Who writes the AI rules for everyone else?

Rules travel too: by copying, competing and pressure.

Who is writing national AI strategies?

AI Index 2026 map: countries with a national strategy on AI, coloured by year of adoption from 2019 to 2025, plus in development, no strategy and NA; many African and Central Asian countries show no strategy or in development. Source: Oxford Insights, 2024; AI Index, 2026

2024: more than half of newly adopted strategies came from emerging economies

Strategy = policy intent, not implementation · the map shows year of adoption, not income group

Stanford HAI, AI Index 2026, ch. 8 (Oxford Insights data)

Why would a country copy another’s rules?

Coercionpowerful states or organizations change weaker states’ incentives to adopt
Competitionstates adopt what rivals adopted, to keep investment and markets
Learninggovernments update beliefs by watching policy results elsewhere
Emulationstates copy what looks appropriate, not what is proven to work

Which one is at work on the next slides?

Simmons, Dobbin & Garrett (2006), International Organization

How does Brussels regulate the world?

EU rule firms complyeverywhere firms lobbyat home other statesadopt (de facto)(de jure)

De facto: firms “standardize their production globally and adhere to a single rule”

De jure: firms then “lobby their domestic governments to adopt these same standards”

  • Needs market power + regulatory capacity

Bradford, Northwestern University Law Review, 2012

Has the EU AI Act travelled?

Travelling Partial Limited
Brazil PL 2338/2023: “excessive” and “high” risk tiers · passed Senate Dec 2024 · not law South Korea AI Basic Act, in force Jan 22, 2026: “distinct from the more prescriptive EU model” “a limited Brussels Effect” (Engler 2022)
Prediction: de facto + de jure effects “likely for parts” (Siegmann & Anderljung 2022) Effect likely, but a “side effect” undercuts EU values (Almada & Radu 2024)

Studies mostly 2022–24 predictions · EU delayed its own high-risk rules to Dec 2027 / Aug 2028

Senate and Chamber of Brazil; US ITA, May 29, 2026; Brookings, Jun 2022; GovAI, Aug 2022; German Law Journal, 2024; Regulation (EU) 2026/1744

What is Washington exporting?

Federal Register, Vol. 90, No. 142, Monday, July 28, 2025, page 35393: Executive Order 14320 of July 23, 2025, Promoting the Export of the American AI Technology Stack; Section 1 Purpose ends: This order establishes a coordinated national effort to support the American AI industry by promoting the export of full-stack American AI technology packages; Sec. 2 Policy; Sec. 3 Establishment of the American AI Exports Program

EO 14320 (Jul 23, 2025): export “full-stack American AI technology packages”

  • Hardware + cloud · data pipelines · models · security · applications, with federal financing
  • “a trusted, sovereign AI stack” (Kratsios, Apr 2026)

Program renamed “SI Exports” (EO 14434, Sep 2026) · 78 applications, no designations found as of Nov 16

Federal Register, Jul 28, 2025; ITA, Apr 1 and Jul 6, 2026; EO 14434

What is Beijing exporting?

China promotes WAICO “to help countries of the Global South keep pace with the global intelligent transformation, bolster capacity building and avert a new AI divide” Wang Yi, Jul 17, 2026 (MFA readout)

Action Plan (Jul 2025): “assist the Global South in truly accessing and utilizing AI”

WAICO, founded Jul 16, 2026: 29 signatories, 35 by Jul 31 · HQ Shanghai

Open models: Chinese models = 41% of Hugging Face downloads (Feb 2025–Feb 2026)

WAICO founding list only partly published · “Global South” is Beijing’s term

MFA PRC, Jul 2025 and Jul 17, 2026; Xinhua, Jul 17, 2026; TPRA Sudan, Jul 31, 2026; Hugging Face, Mar 17, 2026

Can everyone else write rules together?

Cover of the African Union Continental Artificial Intelligence Strategy, July 2024
United Nations General Assembly resolution A/RES/79/325, adopted 26 August 2025, page 1: Terms of reference and modalities for the establishment and functioning of the Independent International Scientific Panel on Artificial Intelligence and the Global Dialogue on Artificial Intelligence Governance

Jul 2024

African Union Continental AI Strategy, endorsed by Executive Council

Aug 26, 2025

UN resolution A/RES/79/325, adopted by consensus -> Scientific Panel (40 members) + Global Dialogue

Sep 25, 2025

Global Dialogue on AI Governance launched (UNGA80, New York)

Jul 1, 2026

Panel’s first report

Jul 6–7, 2026

first Global Dialogue, Geneva

African Union, Jul 2024; UN General Assembly, Aug 2025; UN Global Dialogue site; UN Geneva, Jul 2026

Can governments use and oversee AI?

Rules on paper ≠ capacity to use or enforce them

Left: where we found documented government use · Right: a readiness index’s public-sector adoption score

G3O web page, Where have we found documented use?, tab Share with documented use: world map of the share of checked institutions in each country that document some use of generative AI, in bands not checked, checked none found, 0–5%, 5–15%, 15–30%, 30%+

Public Sector Adoption pillar (0–100) UAEFranceBrazil IndiaIndonesiaCôte d'Ivoire 97.397.294.8 79.173.819.0 0100 Oxford Insights, Government AI Readiness Index 2025, Public Sector Adoption pillar

G3O = instructor’s own project; English-language search; not documented ≠ not used · pillar = index score, not measured use

Oxford Insights, Government AI Readiness Index 2025 (Dec 2025); G3O, as of Nov 4, 2026

Who can audit, enforce, or even count?

44 laws

37 authorities

Of 54 African countries: 44 have a data protection law, 37 an operational authority (CIPIT, Apr 2026).

Kenya’s regulator, FY2024/25
2,071complaints
53enforcement notices
7penalty notices
  • Named risks: AI “may lack legal framework”; staff retention

Counts differ across trackers (44–45 laws; 37–39 authorities)

CIPIT, Strathmore University, Apr 2026; Auditor-General of Kenya, ODPC report FY2024/25

What should you take from today?

  1. AI = global distributive shock · most countries buy chips, compute and frontier models
  2. Four tests: compute, models, jurisdiction, performance · each deal buys some, none buys all, at high cost
  3. Rules travel by coercion, competition, learning, emulation · Brussels effect + US and Chinese stack exports · state capacity sets the limit

Vote again: widen or narrow?

Widen
Narrow
No change
Not sure

See you next class

Peer Review 2 · Sun, Nov 22, 11:59 p.m. · 150–200 words

Wednesday, Nov 18: AI Agents: Reliability, Safety, and Governance · Yiqing Xu

Final presentations · Mon, Nov 30 / Wed, Dec 2 · 5 minutes, 3 slides

Peer Review 3 · in session, Nov 30 / Dec 2