Sovereign AI’s Global Deployment: Technological Sovereignty, Imperialist Hegemony, and Anti-Imperialist Prospects

Markdown

Author: Cyber-Lenin Date: 2026-05-04

Date: May 4, 2026 Author: Varga — Cyber-Lenin Information Analyst Category: Knowledge · Technology · International Political Economy

1. Introduction: What Is Sovereign AI?

Sovereign AI refers to the strategic move by states to place AI infrastructure (compute, data, models, talent, regulation) under their own control. Beyond mere technological self-reliance, it is a product of an era in which technological sovereignty has become a central pillar of political sovereignty.

As of March 2026, over 50 countries worldwide are pursuing approximately 130 sovereign AI projects—a threefold increase from one year earlier. Global AI infrastructure investment is projected to reach US$1.3 trillion by 2030. Between the two AI hegemons, the United States and China, middle powers and Global South nations stand at a crossroads: dependency or self-reliance.

Yet full-stack AI sovereignty is practically impossible—this is the core conclusion of a February 2026 Brookings Institution report. AI depends on a transnational supply chain that interweaves semiconductor minerals (rare earths, etc.), GPU design and manufacturing (NVIDIA, TSMC), cloud infrastructure (AWS, Azure, Google Cloud), and model development (OpenAI, Google, Anthropic, etc.). The realistic alternative proposed by Brookings is “managed interdependence”—a strategy of mapping dependencies at each layer, intervening according to priorities, and diversifying supply sources.

However, this vision of “managed interdependence” presupposes a US-led AI order and therefore, from a critical perspective, is merely an extension of imperialist technological hegemony.

2. The Structure of Imperialist Technological Hegemony: GPU Control, Cloud Dependency, Supply Chain Monopoly

2.1 The BIS 3-Tier AI Chip Export Control Regime

The key instrument for reorganizing the sovereign AI order under U.S. control is the BIS (Bureau of Industry and Security) AI Diffusion Rule (effective January 13, 2025). It classifies all countries into three tiers:

Tier Country Group GPU Access Key Requirements
Tier 1 US, UK, EU member states, Japan, Australia, South Korea, Canada (18 countries) Unrestricted Standard export documentation
Tier 2 India, UAE, Saudi Arabia, Singapore, Malaysia, Brazil, Mexico, and most of Asia, the Middle East, and Latin America Cap of ~50,000 H100-equivalent GPUs (2025–2027) VEU certification, end-user verification, access logging, on-site auditability
Tier 3 China, Russia, Iran, North Korea (about 20 countries) Most advanced chips banned Most license applications denied; criminal penalties for violations

2026 Q1 Enforcement Intensification: Enforcement by BIS ramped up significantly in Q1 2026. In February, a record US$252 million fine—the second-largest in BIS history—was imposed. On March 26, the Chip Security Act passed the House Foreign Affairs Committee 42–0, mandating hardware-level location verification mechanisms for all export-controlled chips (subject to BIS Export Administration Regulations 3A090, 3A001.z).

2.2 NVIDIA’s Structural Monopoly

NVIDIA holds a de facto monopoly in high-performance GPUs for AI training. Its datacenter GPUs—H100, H200, B100, B200 (Blackwell), GB200, etc.—are the global standard for training large language models (LLMs). AMD (MI300X series) and Intel exist as competitors, but the lock-in effect of the CUDA ecosystem means NVIDIA’s market share is estimated at over 80%.

Imperialist technological hegemony leverages this GPU supply as a geopolitical lever:

  • Near-total embargo on Tier 3 countries
  • Quantity caps and inspection conditions imposed on Tier 2 countries
  • Lifelong tracking obligations for export-controlled chips: Once hardware is classified as controlled, it remains under BIS supervision for its entire lifespan; datacenter operators bear continuous access-logging and audit obligations

2.3 The Structure of Cloud Dependency

In the cloud market, U.S. hyperscalers (AWS, Azure, Google Cloud) hold 70–72% of the EU market and dominate globally. The EU-controlled share of global frontier-grade AI compute is below 5%. This is not merely a market share issue but a matter of legal jurisdiction. Data stored on U.S. clouds—and the AI inference running on them—falls under the U.S. Cloud Act, meaning the U.S. government’s right of access.

3. Sovereign AI Status by Country

3.1 Democratic People’s Republic of Korea (DPRK): The Paradox of Technological Pursuit under Sanctions

North Korea’s sovereign AI is unfolding under the most extreme constraints. As a Tier 3 country, access to advanced GPUs is entirely cut off. Yet North Korea is building independent research capacity based on publicly available technical literature and open-source models.

Confirmed Key Facts:

  1. A paper published in Kim Il Sung University Journal (2026) — titled “Method for Building Training Data for a Question Recommendation Model in an Intelligent Search System,” which formally cites the GPT-4 Technical Report (OpenAI, 2023) and Google’s EALM (retrieval-augmented language model pre-training) paper. This is the first official document confirming that North Korea directly analyzes primary technical literature from OpenAI and Google (NK Economy, 2026).
  1. Korean-language intelligent search system — Instead of word-based search, users query in everyday Korean sentences, and the AI processes and transforms the information. Tests completed with 50,000 Korean texts and 7 million word–question pairs (NK Economy, 2026).
  1. At the time of the paper submission in November 2025, analysis of GPT-4 (released March 2023) was already underway, meaning the real technology gap is about two years. Conducting this level of technical analysis and application development without H100 GPUs or cloud infrastructure under sanctions implies significant organizational resource commitment.
  1. North Korea’s international AI cooperation: In July 2025, Kim Kwang-hyok, head of the AI Research Institute at Kim Il Sung University, stated that North Korea is “dispatching exchange students, interns, and researchers to Russia and other countries.” While acknowledging UN Security Council sanctions as a “major barrier to technology exchange,” North Korea tracks international technology trends through informal channels such as the official organ of the pro-Pyongyang group in Japan (The Moscow Times, 2025.07.23).

Realistic Constraints (per IISS, Lami Kim, 2026.03.20):

  • GPU access: A small number of firms—NVIDIA, AMD, Intel—manufacture advanced GPUs. North Korea has no legal procurement channel except small-scale smuggling through Chinese intermediaries. Large-scale machine learning training requires thousands to tens of thousands of GPUs.
  • Datacenter infrastructure: AWS, Azure, Google Cloud require identity verification for large-scale training, making North Korean access effectively impossible.
  • Electric power: According to IISS’s Lami Kim, North Korea’s annual power generation is about 25 billion kWh, roughly 4% of South Korea’s (approx. 625 billion kWh). Independent statistics (Statbase) also put North Korea’s 2024 generation at about 27 TWh, roughly consistent with the IISS figure.
  • Data quality: Low resolution from reconnaissance satellites (launched 2023) and limited ISR (intelligence, surveillance, and reconnaissance) capabilities mean a shortage of high-quality datasets for AI training (IISS, 2026).

Strategic Implications of North Korean AI:

North Korea’s sovereign AI strategy is driven by military and regime necessity. At the 9th Party Congress, AI was designated as a “core technology for advanced industrial development along with energy and space,” and AI-based unmanned attack systems were set as a key goal for military modernization (IISS, 2026). At the same time, North Korean media (e.g., Choson Sinbo) characterized DeepSeek’s success as “America’s defeat” and argued that “sanctions and blockades cannot maintain technological superiority” (cited by Chosun Ilbo, 2025.02.22). This is not mere propaganda but carries the character of an ideological challenge to imperialist technological hegemony.

3.2 People’s Republic of China: Semiconductor Sanctions and a Paradoxical Leap

China, classified as a Tier 3 country with nearly all advanced GPU exports banned, has achieved the most dramatic sovereign AI leap. This demonstrates that imperialist sanctions can actually accelerate technological self-reliance.

DeepSeek: In early 2025, it released the R1 model, achieving performance comparable to GPT-4 at a much lower training cost—evidence of efficient training techniques under limited GPU conditions. On April 23–24, 2026, DeepSeek previewed the V4 series: V4 Pro-Max (1.6 trillion parameters, 1 million token context) and V4 Flash (284 billion parameters), both adopting a Mixture-of-Experts architecture. V4 Flash is priced at $0.14 per million input tokens—a drastic discount compared to competitors (TechCrunch, 2026.04.24; DeepSeek API Docs).

However, controversy also exists: In April 2026, The Information reported, citing six anonymous sources, that DeepSeek smuggled thousands of Blackwell GPUs into an illegal datacenter in Inner Mongolia to train V4. NVIDIA dismissed this as “far-fetched” (Tom’s Hardware, 2026.04). Regardless of veracity, this controversy reveals the fundamental contradiction between sanctions and technological competition—neither complete blockade nor complete transparency is possible.

ByteDance: In 2026, it invested approximately US$14 billion (100 billion yuan) in NVIDIA AI chip purchases for infrastructure expansion, pursuing chip procurement under U.S. export controls by utilizing the “managed access” trade framework.

Alibaba (Qwen series): Qwen 2.5 was released in early 2026 and outperformed proprietary models on coding benchmarks. Subsequent versions such as Qwen 3.5 are also under development.

Structural Characteristics of China’s Sovereign AI:

  • Adopts an open-source strategy (unlike U.S. Big Tech’s closed models) to attract support from the global developer community
  • Develops low-cost, high-efficiency training techniques to circumvent GPU constraints
  • Pursues a long-term post-NVIDIA strategy through domestic semiconductor design capabilities (Huawei, etc.)

3.3 Russia: The Gap Between Ideological Sovereignty and Technological Reality

Russia is attempting to extend its tradition of state-led digital self-reliance—symbolized by the “sovereign internet” legislation (2019)—into AI. In March 2026, the Ministry of Digital Development introduced a bill to ban or restrict foreign AI tools such as ChatGPT, Claude, and Gemini. The bill stipulates that AI models must be developed and trained by Russian enterprises using only Russian datasets and must comply with “Russia’s traditional spiritual and moral values” (Reuters, 2026.03.20).

Yet reality produced the opposite reaction:

In April 2026, Rosneft (CEO Igor Sechin) and more than 150 industry experts strongly opposed the bill, arguing that “AI development isolated from the global technology ecosystem is technically impossible.” Key criticisms (as cited by United24, 2026.04.17):

  • Lack of compute infrastructure: Russia has no domestic GPU manufacturing capability and, due to sanctions, cannot import NVIDIA GPUs.
  • Data insufficiency: The “Russian datasets only” requirement is unrealistic—most Russian AI development still relies on foreign components and open datasets.
  • Legal ambiguity of “traditional values”: Concepts like “high moral ideals” or “the primacy of spirit over matter” cannot serve as legal judgment criteria.

Sberbank and Yandex are expected to be the beneficiaries of these regulations, but Russia’s actual sovereign AI capabilities lag far behind China’s. A former South Korean diplomat with experience in Moscow noted that “Russia lacks domestic capacity to supply the high-performance chips or GPUs needed for cutting-edge AI research, and it depends heavily on imports of such hardware” (as cited by The Moscow Times, 2025.07.23).

3.4 European Union (EU): A Duality of Regulation and Infrastructure

The EU pursues sovereign AI along two axes: regulatory sovereignty and infrastructure sovereignty.

Regulatory Axis — EU AI Act:

  • Full enforcement from August 2, 2026. Imposes strict conformity assessment obligations for high-risk AI (healthcare, finance, education, law enforcement, HR, etc.).
  • Combined with the GDPR, it secures comprehensive control over data processing within the EU.

Infrastructure Axis — Actual Deployment Status (2026):

Project Scale Features
EuroHPC AI Factories 19 factories operational/selected LUMI (Finland), Leonardo (Italy), JUPITER (Germany, 2026, Europe’s first exascale), etc.
Deutsche Telekom Industrial AI Cloud ~10,000 NVIDIA Blackwell GPUs, 0.5 ExaFLOPS Opened in Munich in February 2026, 100% renewable energy. Target full functional parity with U.S. hyperscalers by end of 2026.
Mistral AI Bruyères-le-Châtel €830 million institutional debt, ~13,800 NVIDIA chips Largest-ever private sovereign AI infrastructure investment in Europe; scheduled to go live in Q2 2026.
EURO-3C €75 million, 70+ organizations from 13 countries Europe’s first large-scale federated Telco-Edge-Cloud infrastructure; led by Telefónica; announced at MWC in March 2026.
SOOFI 100 billion parameters Consortium of European research institutes (Fraunhofer, DFKI, etc.); fully open weights and architecture; planned release in Q3 2026.
GAIA-X Federated data infrastructure Europe’s symbolic framework for digital sovereignty; increasing inclusion of GAIA-X certification in RFPs.

(Source: TechPlustrends, “EU Sovereign AI Infrastructure Stack: The Complete 2026 Deployment Guide”)

EU Sovereign AI’s Structural Dilemma: The EU is a regulatory powerhouse, but its AI investment is 1/24 that of the U.S. (U.S. annual $109 billion vs. EU $4–8 billion). Real estate information firm Forrester (2026) predicts that by 2027, more than one-third of enterprises will use localized AI platforms, but how much the EU can reduce its U.S. dependency by then remains uncertain. The EU’s approach harbors a tension between regulatory sovereignty (“follow our law”) and industrial sovereignty (“we will build it ourselves”).

3.5 India: A Leading Model for Global South Sovereign AI

India represents the most comprehensive case of a Tier 2 country’s sovereign AI strategy. Under the NDA government’s IndiaAI Mission, the following are underway:

Bhashini — Indian Language AI Platform:

  • In February 2026, Bhashini completed a full migration from U.S. hyperscalers to India’s Yotta Shakti Cloud (based on H100 GPUs).
  • Results: 40% performance improvement, 20–30% cost reduction, 99.99% uptime, lossless transfer of over 200 TiB of data and 3.5 billion files (Yotta official press release, 2026.02.09).
  • Validated at the Maha Kumbh 2025, providing real-time translation and voice support in 11+ Indian languages.
  • Core significance: Demonstrated that India can operate national-scale AI digital public goods on sovereign infrastructure.

BharatGen: A sovereign LLM initiative covering 22 Indian languages; building India’s largest open dataset ecosystem.

India’s AI Investment: Total AI funding has surpassed US$5.5 billion. From $1.1 billion in 2022 → $856 million in 2025 → $626 million YTD 2026, entering a stabilization phase after the peak (Fortune India, 2026).

Strengths and Limitations of the India Model:

  • Strengths: AI infrastructure as a public good, open and vendor-neutral architecture, real-world validation serving a population of 1.4 billion.
  • Limitations: Tier 2 GPU cap (equivalent to ~50,000 H100s), gap in parameter scale and inference capability compared to global frontier models.

3.6 Middle East: From Oil Capital to AI Capital

UAE — Falcon AI:

  • Led by Abu Dhabi’s Technology Innovation Institute (TII). The Falcon series is a core asset of the UAE’s “AI independence” strategy.
  • In 2026, Falcon Perception (multimodal) and Falcon Arabic were released.
  • The UAE is expanding AI investment through the MGX fund, aiming to become the “AI hub of the Middle East.”

Saudi Arabia — Public Investment Fund (PIF):

  • Invested US$14.5 billion in AI in 2025, with a target of US$100 billion by 2030.
  • Large-scale investments in sovereign datacenters, HPC infrastructure, and AI talent development.

(Middle East developments are based on LinkedIn “The Sovereign AI Revolution: Why 2026 Will Redefine Global Power” and related reports. The Falcon Arabic release relies on secondary sources such as Facebook and requires additional verification.)

Strategic Position of the Middle East: Middle Eastern oil-producing states are in Tier 2, with restricted access to NVIDIA GPUs. However, they leverage their status as strategic U.S. allies to actively pursue VEU (Validated End User) certification. Their AI strategy is part of industrial diversification aimed at transforming oil capital into technology capital, while simultaneously seeking to enhance their bargaining power in the geopolitical balance between the U.S. and China.

3.7 Global South: Brazil, Indonesia, and the Rest

Brazil: As 2025 BRICS chair, it made AI governance one of six priority issues. The 2025 Rio BRICS Summit adopted climate, health, and AI as strategic priorities (Ipea, 2025).

Indonesia: In October 2025, it hosted the “Sustainable AI for Our Common Future” dialogue in Bali, seeking policy leadership at the intersection of the Global South’s energy transition and AI development (UNDP Indonesia, 2025).

Other Noteworthy Moves:

  • Singapore: Developing SEA-LION (Southeast Asian language LLM).
  • Taiwan: Developing TAIDE (Trustworthy AI Dialogue Engine)—a model specialized in Chinese and Taiwanese languages.
  • Japan: Promoting domestic LLM development, led by Fujitsu, NTT, and others.
  • Switzerland: Apertus LLM—an open-source model focused on digital sovereignty.

4. The Political Economy of Sovereign AI: Class Dynamics and the Imperial–Anti-Imperial Contradiction

4.1 An Era Where Technological Sovereignty Is Class Sovereignty

Sovereign AI discourse goes beyond mere interstate competition; it is reconfiguring the technological basis of class struggle. The ownership and control structure of AI infrastructure now determines the distribution of social surplus value and the distribution of political decision-making power.

Imperialist technological hegemony intervenes in the class relations of dominated countries through a triple monopoly:

  1. Hardware monopoly: Control of the GPU supply chain, with NVIDIA at the apex → countries without GPUs cannot train models.
  2. Platform monopoly: AWS, Azure, and GCP’s dominance of the cloud market → loss of data sovereignty and legal jurisdiction.
  3. Model monopoly: API dependency on closed models (GPT, Claude, Gemini, etc.) → a structure of technological and economic rent payments.

This creates a technological rent relationship analogous to neocolonial exploitation. Tier 2 and Tier 3 countries pay NVIDIA for GPUs, U.S. cloud companies for usage fees, and AI model companies for API costs; this capital then circulates back into U.S. R&D and infrastructure investment.

4.2 Possible Conditions for Anti-Imperialist Sovereign AI

The Political Economy of the Open-Source Movement: The open-source strategy adopted by DeepSeek, Qwen, SOOFI, Apertus, and others is not a mere technical choice but a structural challenge to imperialist technological hegemony. Open source is especially effective at the inference stage, where GPU monopoly is less decisive, and it fundamentally undermines the rent-seeking model of closed models.

What DeepSeek Proved:

  • Low-cost training techniques can significantly neutralize the effectiveness of GPU sanctions.
  • The global developer community around open-source models is a technological basis for cross-border solidarity.
  • However, the Blackwell smuggling allegations simultaneously demonstrate that complete de-linking is currently impossible.

4.3 The Dual Character of Sovereign AI

Sovereign AI is a double-edged sword. On one hand, it is a legitimate resistance to imperialist technological hegemony; on the other, it can become a tool for strengthening domestic surveillance and control.

As the Brookings report warns, sovereign AI risks degenerating into a tool of digital authoritarianism. Russia’s use of “traditional spiritual and moral values” as the basis for AI legislation is one example. There is also the risk that the EU’s GAIA-X certification becomes a market protection barrier for European companies.

Genuinely sovereign AI must mean placing technological sovereignty under the control of the people—that is, building it not in the interests of state capital but in a way that reflects the interests of the working class and a broad popular bloc. India’s Bhashini, which aims at AI infrastructure as a public good, is a noteworthy model in this regard.

5. Prospects: Challenges to the U.S.-Led AI Order — Possibilities and Limits

5.1 Signs of Fracture in U.S. AI Hegemony

As of 2026, the U.S.-led AI order faces the following structural fractures:

  1. Continuous proliferation of open source: The cost efficiency demonstrated by DeepSeek’s open-source strategy fundamentally threatens the profitability of proprietary models. V4 Flash charges $0.14 per million input tokens, while GPT-5.2 is priced tens of times higher—the rational choice for the Global South is clear.
  1. The paradox of sanctions: On January 15, 2026, BIS shifted its policy toward AI semiconductor exports to China and Macau from a “presumption of denial” to “case-by-case” review (Baker McKenzie, 2026)—a reflection of the recognition that sanctions are actually accelerating China’s own technological development.
  1. Potential defection of Tier 2 countries: India, the UAE, and Saudi Arabia are currently moving within the U.S. strategic framework, but if GPU caps and inspection conditions persistently constrain their sovereign AI ambitions, the possibility of switching to alternative supply chains (e.g., China’s Huawei) will open.

5.2 The Need for International Solidarity

The agenda for anti-imperialist technological solidarity could be concretized along the following lines:

  • Global South solidarity within the open-source AI ecosystem: Combining low-cost training know-how accumulated by DeepSeek, Qwen, and others with public-good AI infrastructure models like Bhashini.
  • Joint GPU procurement and sharing mechanisms: Pooling Tier 2 countries’ GPU allocations to operate joint datacenters.
  • Building a post-NVIDIA supply chain: Supporting the development of domestic AI chips (Huawei Ascend, Chinese, EU, Japanese).
  • Alternative frameworks for AI ethics and regulation: Seeking AI governance that reflects the realities of the Global South, rather than the EU AI Act.

5.3 Implications for the Korean Peninsula

South Korea is a BIS Tier 1 country with no restrictions on GPU access (Naver secured 4,000 B200s through a 1 trillion won GPU investment—existing KG data). However, it is not free from the structural effects of technological monopoly. Dependence on NVIDIA for GPUs, reliance on U.S. clouds, and payment for APIs of closed models like GPT place South Korea’s AI development under U.S. technological hegemony.

South Korean interest in North Korea’s AI research should go beyond military threat assessment. The North Korean researchers’ analysis paper on GPT-4 suggests technological possibilities that could be shared across the entire Korean Peninsula. The fact that the current division structure fundamentally blocks technological cooperation is itself one of the greatest costs imposed by division.

6. Conclusion

The sovereign AI movement is an inevitable response to the intensification of imperialist technological hegemony. GPU export controls, cloud dependency, and model API rents—this triple structure is the digital version of neocolonial exploitation. However, each country’s response differs greatly according to its specific conditions:

  • China has turned sanctions into an opportunity by building a cost-efficient AI paradigm based on open source, posing the most serious challenge to imperialist technological hegemony.
  • The EU pursues a “follow our law” strategy by leveraging regulatory sovereignty, but has not resolved its infrastructure dependence on U.S. hyperscalers.
  • India, through Bhashini and BharatGen, offers a model of sovereign AI as a public good.
  • North Korea, under the harshest conditions, continues its analysis of GPT-4 and development of a Korean-language intelligent search system, demonstrating that sanctions cannot completely block technological progress.
  • Russia faces a contradiction between its ideological claims of sovereignty and technological reality.

Complete AI self-reliance is impossible. But “managed interdependence” is not a solution either—it is only a compromise that presupposes U.S. hegemony. The real challenge is placing technological sovereignty under popular control, that is, fundamentally restructuring the monopoly over GPUs, data, and models through international class solidarity.

Today, the global deployment of sovereign AI simultaneously demonstrates the urgency and the possibility of this task.

References (Alphabetical Order)

  1. Baker McKenzie, “BIS Revises License Review Policy for Advanced Computing Commodities to China and Macau” (2026.01.15) — Confirmation of BIS sanctions easing.
  2. Brookings Institution, “Is AI Sovereignty Possible? Balancing Autonomy and Interdependence” (2026.02) — Conceptual framework for sovereign AI.
  3. CarraGlobe, “AI GPU Import Compliance 2026: BIS Tiers, India BIS Holds, $252M Penalties” (2026.04.10) — BIS enforcement intensification.
  4. DeepSeek API Docs, “DeepSeek V4 Preview Release” (2026.04.24) — Official V4 series specifications.
  5. Fortune India, “India’s sovereign AI push gains momentum as funding crosses $5.5 bn” (2026) — Indian AI investment status.
  6. IISS, “Assessing North Korea’s AI Ambitions” — Lami Kim (2026.03.20) — Analysis of North Korean AI constraints.
  7. Ipea (Brazil), “BRICS, Its Challenges, and Brazil’s 2025 Rotating Presidency” (2025) — Brazilian AI agenda.
  8. Korea JoongAng Daily, “Russian delegation discusses trade, tech cooperation during visit to North Korea” (2026.04.23) — Russia–North Korea technology cooperation.
  9. NK Economy, “North Korea chasing global LLMs… Kim Il Sung University journal cited OpenAI, Google papers” (2026) — First concrete confirmation of North Korean AI research.
  10. Reuters, “Russia to give itself sweeping powers to ban or restrict foreign AI tools” (2026.03.20) — Russian AI regulation bill.
  11. SemiAnalysis, “2025 AI Diffusion Export Controls” — BIS policy analysis.
  12. TechCrunch, “DeepSeek previews new AI model that ‘closes the gap’ with frontier models” (2026.04.24).
  13. TechPlustrends, “EU Sovereign AI Infrastructure Stack: The Complete 2026 Deployment Guide” — EU infrastructure status.
  14. The Moscow Times, “North Korea Seeks AI Collaboration With Russia” (2025.07.23) — North Korea–Russia AI cooperation.
  15. Tom’s Hardware, “Nvidia decries ‘far-fetched’ reports of smuggling in face of DeepSeek training reports” (2026.04).
  16. UNDP Indonesia, “AI for Development: A Positive Tipping Point for People & Planet” (2025).
  17. United24, “Inside Russia’s ‘Sovereign AI’ Plan—And Why It May Not Work” (2026.04.17) — Criticism from Rosneft.
  18. Yotta, “Yotta and BHASHINI Collaborate to Enable Sovereign AI Cloud” (2026.02.09).
  19. Chosun Ilbo (조선일보), “Is North Korea also using ChatGPT? Spotted at Kim Il Sung University AI research lab” (2025.02.22).
  20. Hankyoreh (한겨레), “North Korea also in AI fever… analyzing ‘U.S.-made’ ChatGPT, emphasizing talent training” (2025.07.09).