Implications of Artificial Intelligence on Geopolitics

The rapid development of Artificial Intelligence (AI)—and, in particular, of the more powerful and transformative Artificial General Intelligence (AGI)—is upending political, social, economic, and military activity both within nations and among them. We are already witnessing significant geopolitical churn, as the asymmetrical distribution of these advanced capabilities reshapes inter-state relations and creates new external challenges, particularly for the developing countries of the Global South, including India.

These countries are keen to harness the promise of AI for their economic transformation, but they are equally anxious to avoid a discriminatory regime of the Nuclear Non-Proliferation Treaty (NPT) type—dominated by the United States and China—that would perpetuate an asymmetry of power based on controlled access to a new and formidable technology, reverse the trend toward a multipolar world, and push the Global South to the geopolitical margins. The nature of that challenge, its impact on geopolitics, and the strategies the Global South must adopt to avoid the periphery are the subject of what follows.

AI and AGI are not interchangeable terms, and any serious analysis must begin by distinguishing them.

AI refers to systems that perform specific tasks using data and algorithms; it enhances capacity within particular domains. In a 2025 RAND Corporation/Perry World House report on the AGI race and international security, AI policy researcher Miles Brundage argues that “by allowing AI systems to learn how to reason through their own experience rather than from humans telling them how to do so, this paradigm has the potential to significantly exceed human intelligence in any domain where a source of feedback is available.”

AGI, by contrast, is a system capable of performing valuable, important, or relevant tasks at or beyond the level a human being could achieve, across a general—rather than narrow—scope, and with a degree of autonomy.

Artificial Super-Intelligence (ASI) remains a hypothetical stage in which machine intelligence far exceeds human cognitive abilities across the board. ASI may be thought of as an all-powerful “general-purpose strategic technology” that acquires additional capabilities in an accelerating, cumulative process. This stage is also described as the technological singularity: a point at which the growth of intelligence becomes so rapid and so transformative that human institutions, laws, and even geopolitics may be rendered irrelevant.

While AI is already pervasive—through programs such as ChatGPT and Co-Pilot from the United States, or DeepSeek and Qwen from China—we remain some distance from fully autonomous AGI systems. Their feasibility, however, is no longer in doubt.

AI has already changed the way countries relate to one another, and the nature of diplomacy itself. The character of warfare has undergone a sea-change, with AI-embedded weapons already in use in the ongoing war in Ukraine, and AI reportedly used for target selection during the recent Iran war.

AI dominance flows through physical ‘chokepoints’—primarily advanced chips, the data centers that house them, the power required to run them, and the water needed to cool them. For advanced chips, the key metrics include the processing node, measured in nanometers: the smaller the node, the more densely the chips can be packed, and the greater the Total Processing Performance (TPP) they deliver. Interconnection bandwidth determines chip-to-chip communication speed—a crucial factor, since training large language models (LLMs) requires linking thousands of chips together.

Data centers perform three related functions. First, they carry out “compute,” which encompasses both the training of models—adjusting billions of parameters over weeks—and inference, the running of a trained model to answer queries. This depends on “tokenization,” the process by which a language model breaks text into the chunks it can process. A token may be a single short word such as “cat,” or a fragment of a longer one: “invaluable,” for instance, breaks into “in-valu-able,” or 3 tokens. The number of tokens a model can handle at once is known as its “context window.” Tokens are the foundation of AI communication.

Second, data centers store training datasets. Third, they enable the extremely high-bandwidth interconnections between chips and between racks that training demands, since models require constant synchronization across several thousand Graphics Processing Units (GPUs). This, too, can become a chokepoint—even where advanced chips are available.

Data centers are often rated by the power they can draw and the heat they can dissipate. The grid capacity a center can draw upon continuously and in a stable manner, and its ability to cool the facility through access to large volumes of water, may therefore become major constraints. Estimates vary, but global data-center power demand stood at around 104 gigawatts in 2025, and may rise to 132 gigawatts in 2026 and reach 290 gigawatts by 2030—an annual increase estimated at between 14 percent and 27 percent. The United States accounts for approximately 43 percent of this consumption. To grasp the scale involved, India’s total power capacity at present is 592 gigawatts. It is power availability, rather than chip supply, that is the major constraint on adding new capacity.

Less discussed is the significant demand for water to cool the data centers. The International Energy Agency (IEA), in its 2025 “Energy and AI” report, estimates that a 100-megawatt data center in the United States withdraws about 2 million liters of water daily. This is a further bottleneck, especially in water-stressed countries such as India.

India currently has 1.5 gigawatts of power capacity dedicated to data processing, projected to rise to 8 gigawatts by 2030. The Adani Group has a flagship data center planned at Visakhapatnam, rated at 1 gigawatt, to be completed by 2030 at a cost of $15 billion; it will likely require large volumes of water for cooling, given the hot and humid climate of the city. The Reliance Group has a data center planned at Jamnagar in Gujarat, in collaboration with Meta, rated at 3 gigawatts—though in the initial two-year phase it will offer 168 megawatts of capacity for Meta’s use, cooled with desalinated water. These efforts are modest by comparison with those of the United States and China, but they keep India in the AI race.

The United States and China are the undoubted leaders in AI, but they have pursued different development strategies. The United States has adopted a proprietary approach, licensing each generation of AI and raising revenue from its users. China’s strategy is to make its AI offerings open-source, permitting rapid diffusion across the whole spectrum of economic and military activity and enabling productivity gains throughout its economy. If the American emphasis is on cutting-edge research, the Chinese focus is on application. The Chinese “stack” is likely to prove the more attractive and cost-effective proposition for the countries of the Global South, particularly those tied more closely to the Chinese economy through the Belt and Road Initiative (BRI). The U.S. stack, though more advanced, will carry greater currency among America’s network of allies and partners—especially those where the major U.S. technology companies, such as Meta, Google, Amazon, Microsoft, and Apple, are already entrenched. These rival stacks are likely to grow increasingly incompatible with one another and may force the large constituency of developing countries to choose between them. Neither side is likely to grant access to its most sensitive and most powerful versions. The temporary U.S. suspension in mid-2026 of foreign access to Fable and Mythos—the more powerful models created by Anthropic—though subsequently lifted, was a pointer to precisely this logic.

For countries like India, the hope of remaining in the AI game is reinforced by the “fast follower” phenomenon, which requires some explanation. A model’s capability is contained in a massive file of numbers, or “weights,” produced by repeated cycles of training over enormous volumes of data. This file can be copied or emailed and then modified for different uses, and it can be run on hardware far less powerful than that used to create it. Such is the nature of open-source AI. For proprietary versions such as ChatGPT or Gemini, the weights remain with the company: the user has access to an app based on them, but not to the weights themselves. Even proprietary models may be “distilled,” a process that involves gathering data from a cascade of responses sought repeatedly from the app and using it to train a clone of the original—one of lesser capacity but produced at a fraction of the cost of developing a powerful new model from scratch.

ChatGPT’s makers have alleged that the Chinese open-source model DeepSeek is the product of exactly such distillation. The point is that it is very difficult to prevent the rapid diffusion of advanced models, so long as the hardware—advanced chips, data centers, power, and water, the so-called physical chokepoints—is not a binding constraint. The availability of highly skilled AI engineers is a further constraint. This is why the United States has restricted the export of the most advanced chips produced by Nvidia, its premier semiconductor company, and also curtailed the supply of the complex photolithography machines used to manufacture them—the most advanced of which are produced solely by the Dutch firm ASML. China is working to close the gap by concentrating resources and skills on this very technology.

While China has pursued a strategy of embedding AI across its extensive manufacturing chain, the United States is determined to dominate frontier AI and to achieve AGI at an early date. It holds an edge at this stage, and it is the so-called U.S. hyperscalers—Google, Anthropic, Microsoft—that are most prominent in the race toward AGI.

For the present, and for the foreseeable future, it is AI that has become the source of power and influence in the geopolitical arena. Yet the drive among AI leaders to monopolize the technology is constantly undermined by its inherent tendency toward rapid diffusion. That diffusion can be subjected to controls over physical chokepoints—a ban on the supply of advanced chips, for example—but historical experience suggests that technology remains fungible across borders. The problem of non-diffusion is compounded by the fact that, unlike nuclear systems, AI offers few visible indicators of the kind that nuclear power plants, fissile-material stocks, or reprocessing facilities provide.

AI also blurs the boundary between the civilian and military domains, where verification is virtually impossible. Experts describe this as an “ungovernable dead zone” for arms control. The older approach—applying prohibitions, or regulated and verifiable access, to high technology for military use while allowing easier access for civilian purposes—is no longer feasible. The default position becomes one in which the leaders prevent access to high-end technology such as AI across the board. This should be of significant concern to the countries of the Global South.

Current AI governance literature has begun to explore whether lessons from nuclear arms control might be adapted to the management of advanced artificial intelligence. In a 2025 RAND and Perry World House report, political scientists Jane Vaynman and Tristan A. Volpe contend that the Nuclear Non-Proliferation Treaty offers a useful, though imperfect, analogy for AI governance. Their central insight is that effective management of AI may require not only universal restraint or open competition, but also what they call “selective collusion”: cartel-like arrangements among leading AI powers that make responsible behavior more distinguishable, reduce incentives for deception, and create the conditions for more stable competition in both commercial and military domains.

The reality is that the very nature of the technology, and the speed of its development, demand urgent multilateral and collaborative action: to establish common safety and security standards for AI, to build strong and effective auditing and verification mechanisms, to incorporate universally applicable verification provisions, and—most importantly—to secure a global consensus on maintaining human oversight of AI. This is all the more necessary because AI lends itself to use by non-state actors, such as terrorist or extremist groups, capable of inflicting significant damage on societies.

We have already witnessed catastrophic cyber-attacks, disruptive disinformation campaigns, and the use of deepfakes by criminal groups—threats that can only be countered through multilateral collaboration. But such security threats must not become the pretext for an intrusive and restrictive control regime operated by the technology leaders while they themselves remain exempt from its application. The Global South needs to mobilize international opinion in favor of a genuinely multilateral AI regime, universally applicable, and to reject any attempt at cartelization.

Setting the terms for international engagement on AI is both important and urgent, for an even greater challenge will confront humanity if AGI becomes a reality. As defined earlier, AGI would represent a quantum leap in capability, enabling accelerated science and technology, explosive economic growth, and unprecedented military power. It is not clear whether the principle of diffusion we have seen in AI would continue to operate with AGI, since AGI could unleash a rapid, cumulative, and self-reinforcing dynamic of technological change that favors first movers. Nor is it known how this would affect human societies, nation-states, and the relations among them.

Since AGI may function with autonomy, would it pursue objectives that do not align with those of nation-states—including the states in which it originated? Could it become an independent global actor? Could it produce a “wonder weapon” whose possession might render one nation-state truly dominant? We do not know. The question is whether humanity should knowingly cross the Rubicon into a state from which there is no return, and no reversibility. Some scholars warn that we are unprepared for what AGI will unleash, and that the technology is outstripping our capacity to control it. This is a “threshold moment” for our species.

On AGI, only the United States and China stand at the technological frontier. No other country comes close, and their lead widens by the day. They have every incentive to compete—and a concurrent temptation to collude in order to preserve their shared monopoly. The United States is probably ahead, given the unprecedented scale of resources being poured into the sector by the hyperscalers. The Bank for International Settlements (BIS) has drawn attention to the risk of a global financial meltdown should investor sentiment sour and an investment bust follow. The hyperscalers are expected to invest more than $1 trillion between 2025 and 2026, through equity and bond markets, even as the prospect of commensurate returns grows increasingly problematic. The BIS, in the “Progress and Peril” chapter of its 2026 Annual Economic report, pointed to historical parallels in which “a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify” led to “an eventual reversal in investment, inducing economy-wide recession.”

Were such a bust to occur, it would be far more severe than the global financial crisis of 2007 to 2008, and possibly of longer duration, since most major economies no longer have the fiscal space to spend their way out of recession as they did then and during the subsequent COVID pandemic. This would be uncharted territory; but at the very least, the pursuit of AI, and of frontier research in AGI, would be likely to slow, if not stall altogether.

After the dust has settled, China might, by default, emerge as the winner of the race, its more measured strategy of placing a series of smaller bets having proved wiser than a gamble on the holy grail of AGI. China could not escape the effects of a global recession, due to its economy being deeply integrated into world economy, but its recovery might well be faster.

For the countries of the Global South, and particularly its leading members, AI is a technological capability that, if harnessed properly, could drive rapid economic growth. Attempts by the leading AI nations to restrict access must be resisted—and this means, above all, addressing the physical chokepoints of semiconductor chips and data centers described earlier. It may be possible for a coalition of partners, drawn from both the Global South and the developed world—France, Germany, Japan, South Korea—to pursue projects with pooled resources, such as the limited supply of GPUs. This would help counter the drift toward a duopoly.

It is worth noting that international discourse on AI has shifted from an emphasis on safety toward the leveraging of its potential. The framing of risk has been progressively diluted. The recent AI summit in Delhi was deliberately rebranded from a Safety Summit to an Impact Summit, and the Indian government’s official guidance explicitly states that “other things being equal, responsible innovation should be prioritized over cautionary restraint.”

This dilution of risk is a matter of concern, and its consequences are already visible in the growing menace of cyber-fraud, identity theft, and the use of AI tools by predatory states to invade the privacy of their citizens. The increasing reliance on automated systems, and the recourse to AI-driven recommendations in decision-making, carries significant risks both within countries and in their external relations. AI can make mistakes; it may hallucinate, conjuring the perception of a threat where none exists. It may tempt states to use AI tools to enable kinetic action for regime change—as we witnessed during the recent Iran war—while leaving them unable to manage the consequences on the ground.

With its capacity to assemble vast quantities of data and to conduct intrusive intelligence operations remotely, AI has tipped the balance in favor of coercive and kinetic action over patient diplomacy. Automated systems may set an action-and-reaction process in motion, and an escalatory spiral of military action, before political leaderships can weigh in and take control. The speed of response—any response—rather than careful deliberation, has come to define geopolitical success. This is creating a more unpredictable and more dangerous geopolitical landscape.

We are likely to remain in this state of flux and uncertainty for the foreseeable future, with no stable anchors in the form of empowered international institutions of governance—and with a concurrent, constant erosion of international law and of the norms that have generally guided inter-state relations. This puts a premium on resilience, both societal and at the level of the state. A brittle polity and a fragmenting society will not be able to meet the challenge that AI has unleashed. Despite its many flaws, a democratic dispensation—one that enjoys popular legitimacy and embraces accountability—may still offer the best hope of navigating these uncharted waters safely.