equipment, along with the cooling and power
systems that support them. Such infrastructure is
often housed in data centers and can be accessed
through cloud providers.
15. AI sovereignty is an ambiguous concept, with definitions ranging from the most ambitious form of
complete self-sufficiency to more calibrated
approaches involving control over certain parts of
the AI value chain. Anthony (2025); Pava et al.
(2026); Tanner et al. (2026).
16. Daskal (2018); Schwartz (2018). In part to address
concerns surrounding data localization, cloud
providers now offer various approaches with different degrees of sovereignty. Such approaches
range from software-based controls, to clouds
operated by domestic entities using the hyperscalers’ technologies, to standalone environments
deployed on premises or within secure facilities.
For more information, refer to Meinhardt et al.
(2026).
17. Narratives around AI sovereignty are often driven
by foreign technological companies that operate
under the laws of their own countries and benefit
greatly from the demand that sovereignty generates. NVIDIA has been among its most visible proponents. Jensen Huang, NVIDIA’s chief executive
officer, declared that “every country needs sovereign AI” (Yew et al. 2025). Leading hyperscalers have
developed specialized sovereign cloud solutions. For
example, Amazon Web Services now provides the
European Sovereign Cloud offering.
18. Regardless of whether it is built for economic or
sovereign reasons, a data center tends to generate
domestic political costs wherever it is built. The
clearest illustration is in the United States, where
the build-out has not been to reduce foreign
dependence. Nonetheless, the building of data
centers has mobilized considerable opposition
across political lines and in many types of communities. In 2025, at least 48 data center projects,
worth more than $156 billion, were blocked or
delayed amid local opposition (DePillis 2026, citing
research by Data Center Watch). Commonly cited
concerns include rising utility bills, noise, and
strain on grid capacity. In some cases, opposition
has led to county- or statewide moratoriums, as
well as suspensions of tax incentives. This pattern
is not unique to the United States. Because of
intense local opposition, Meta’s planned
166-hectare data center in the Netherlands was
halted in 2022, and so was Google’s planned $1.1
billion data center in Luxembourg (Roach and
Krukowska 2022). Media coverage in 2026 indicates that planning for Google’s data center in
Luxembourg is advancing despite opposition
(Monaghan 2026).
19. Anand and Mookerjee (2026).
20. Mozur et al. (2025); Pooler (2026).
21. Clayton et al. (2025).
22. For example, refer to Bertrand et al. (2014); Blanes
i Vidal et al. (2012); Blanga-Gubbay et al. (2025);
Kang (2016).
23. 2025 data of Issue Profile: Science and Technology
(dashboard), OpenSecrets, https://www.opensecrets
.org/federal-lobbying/issues/summar y?c ycle
=2025&id=SCI.
24. Despite the dramatic rise, only about 4 percent to
5 percent of firms participating in lobbying were
lobbying on AI-related issues.
25. Csernatoni (2025); Jones (2023); Mukherjee and
Chee (2025).
26. Agência Pública and CLIP (2025).
27. Narechania and Sitaraman (2024).
28. Since 2018, the United States has imposed a number of export controls on advanced semiconductor
chips and the associated supply chain. The measures have compelled AI companies based in the
United States, such as Advanced Micro Devices
(AMD) and NVIDIA, to halt or modify sales to major
Chinese customers, illustrating how state authority can override market incentives in strategically
important sectors (Sutter 2025).
29. For example, in 2021, the Nigerian government
indefinitely suspended Twitter’s operations in
response to disputes concerning content moderation and political expression (Olurounbi 2021). The
ban persisted for seven months and was lifted only
when the company largely agreed to the government’s demands (BBC 2022).
30. A notable example is the disagreements between
Anthropic and the US government in 2026 over
the company’s refusal to relax AI safety guardrails for military use, specifically on autonomous
weapons and mass surveillance. As a result,
Anthropic was classified as a supply chain risk
(Shalal et al. 2026).
31. For instance, in 2025, the US government agreed
to purchase 433.3 million primary shares of Intel
common stock at a discounted price of $20.47 per
share, equivalent to a 9.9 percent stake in the company (Intel Corporation 2025). In recent years,
state-owned entities in China have acquired
golden shares in subsidiaries of major Chinese
firms, such as Alibaba, ByteDance, and Tencent.
32. Reuters (2025).
33. Phartiyal (2026).
34. World Bank (2026).
35. Kobrin (1984); Vernon (1971).
36. Artunç and Saleh (2025); Faccio (2006); Fisman
(2001).
37. Rent seeking refers to the practice of manipulating
public policy or economic conditions as a strategy
to increase profits.
38. Jakarta Post (2026).
39. Akcigit et al. (2023).
40. Refer to Government Requests for User Data
(graphic), Meta, https://transparency.meta.com
/reports/government-data-requests/.
41. Network effects refer to a phenomenon in which
the value of a product or service increases exponentially as more people use it. As the user base
grows, the service becomes more useful, secure,
or valuable, creating a self-reinforcing loop that
drives widespread adoption. Data feedback loops
AI’s Political Impact: Reshaping Power Within and Across Countries
221