A reordering is under way, and it does not look like the revolutions earlier generations were trained to recognize: it looks like a leaderboard. In April 2026, the artificial intelligence (AI) model that briefly led the open-weight rankings on the Artificial Analysis Intelligence Index was not ChatGPT, nor Claude, nor Gemini — it was Kimi K2.6, a Chinese model from a Beijing start-up called Moonshot AI. Kimi K2.6 is free to download, fine-tunable in any language and unfamiliar to almost any Western policy maker who will now have to reckon with it.
That is the shape of the geopolitical contest over AI, but it is only the visible edge. The deeper story is that access to frontier AI models, and to the compute, chips, energy and data that produce them, has become a lever of statecraft. In today’s economy, a nation’s sovereignty over its cognitive layer has become as critical as its sovereignty over energy had been in the twentieth century. The competition to protect that sovereignty is now running between America and China; through them, between Washington and its allies; and into the boardrooms of every major Western firm whose strategy depends on infrastructure it does not own. Control over frontier AI is the struggle for the commanding heights of this century’s economy: not the digital economy, the economy.
In the coming decade, the choice that corporate boards, government ministries and investors face is not whether to participate in the AI economy — that decision has already been made for them — but on whose terms, and what to do when the terms change. Externalities are over — geopolitics is now the operating environment.
Two Countries, Two Systems
The American AI model is hyperscaled and centralized. A handful of firms — OpenAI, Anthropic, Google, Microsoft — build frontier systems as a centralized core and rent capability outward. This means that weights stay in the lab and safety is controlled at the centre. As a result, commercial rents and geopolitical leverage concentrate accordingly. According to Stanford’s 2025 AI Index, American private investment in AI reached roughly $109 billion in 2024 against $9 billion in China; by 2025, the American figure had risen to about $286 billion. The five largest American hyperscalers are projected to spend $600–700 billion on capital expenditure in 2026 alone. Stargate, a single OpenAI-SoftBank-Oracle consortium, has committed $500 billion over four years. The 2025–2028 build-out is on pace to exceed the cost of the Apollo program, the Manhattan Project, the Marshall Plan, US rural electrification and the interstate highway system combined — roughly $1.2 trillion in today’s dollars.
By contrast, the Chinese model is decentralized, open-weight and built around a different theory of advantage. AI models such as DeepSeek, Qwen, Yi and Moonshot’s Kimi K2.6 can be downloaded, fine-tuned in local languages and run on hardware that does not need a hyperscale data centre behind it. AfriqueQwen-14B, a Qwen variant, has been continuously pretrained on 20 African languages by researchers at McGill-NLP, a Montreal-based natural language processing group led by a Nigerian computer scientist. Chinese laboratories have quietly become the world’s dominant publishers of open-weight AI. Foreign Affairs identified their strategic logic in July 2025: the technology is cheap, capable, open, locally adaptable. Put another way, China’s strategy is soft power as a download.
The questions around centralized versus decentralized or open-weight versus proprietary systems are not neutral technical choices: they are statecraft itself. Decentralization is not a feature of the Chinese stack; it is the argument. It says to the rest of the world do not trust your most important systems to American firms whose terms of access can be changed by an American president. Build your own, on our weights, in your own jurisdiction. As SecDev wrote last November, this may be the bet that ages best. Hyperscale is impressive, but it also has a head office, a court of jurisdiction and a single political climate around it. Smaller, sovereign and “good enough” technology may outlast bigger, smarter and conditional models.
Technology now sits at the heart of geopolitics and will shape its outcomes. Yanis Varoufakis, in Technofeudalism (2023), calls the new asset class “cloud capital,” meaning an asset whose owners extract rents, like feudal landlords rather than industrial capitalists. The Council on Foreign Relations describes an AI sovereignty paradox: the United States hosts roughly 75 percent of global AI supercomputer performance, China another 15 percent, and everyone else competes for the rest. Recent peer-reviewed work frames sovereignty as a continuum, with states differentiated by their steering capacity over compute, talent and standards. The boundary between technology policy and foreign policy has dissolved.
The Evidence: Crypto
If anyone doubts that the break from American technological dominance is real, then follow the money. The cleanest evidence that an alternative system has arrived is not in AI rankings; it is in cryptocurrency adoption across the Global South. Over the past three years, the use of stablecoins and bitcoin in South Asia, Sub-Saharan Africa and Southeast Asia has grown sharply. The drivers are not speculative but structural: the weaponization of dollar sanctions, the de-risking of entire jurisdictions by Western banks, and the simple fact that for the unbanked majority of the world’s adult population, a TronLink wallet on a $40 phone is more useful and accessible than a traditional banking relationship.
The cumulative effect is the construction of a parallel financial infrastructure. Stablecoins have largely solved the off-ramp from cryptocurrency to dollars. Where the balance of trade increasingly runs to China, transactions need not settle in dollars at all. In Burma, Cambodia and Laos, industrial-scale fraud operations tied to crypto rails are now estimated at roughly 40 percent of those countries’ combined formal GDP. Crypto has become an alternative system capable of operating regardless of Western controls or approval.
Read this evidence carefully. What is happening in crypto is the same “transfer of power” dynamic that is happening in AI. Crypto’s decentralized, open architecture is being adopted by the parts of the world that the centralized Western architecture priced out, sanctioned out or simply never reached. The parallel financial system and the parallel AI stack are not coincidences; they are the same structural shift observed in two different domains. In both cases, the decentralized models gain ground when Western institutions weaponize access to the centralized version. Every time Washington tightens chip export controls, it strengthens the case for DeepSeek. Every time it adds a country to a sanctions list, it strengthens the case for stablecoin settlement. In other words, the American policy reflex is producing the very fragmentation it is trying to prevent.
For any institution that depends on the dollar-clearing system, Western payment rails or consumer access to Global South markets, this is not an abstraction: it is a vendor risk. The assumption of a single global financial architecture, taken for granted in every long-range plan written since the end of the Cold War and the beginning of the AI age, is no longer safe.
Access as a Weapon
It is tempting to read all these patterns as only a story about the Global South: about emerging economies adopting Chinese tools because Western ones are out of reach. But that reading misses the wider, and perhaps more consequential, ripple effect. These geopolitical dynamics are also playing out at the core of the Western alliance system and, indeed, inside the boardrooms of every firm that runs its strategy on infrastructure it does not own.
Geopolitics is not just what happens between adversaries anymore. Rather, it can be observed among allies, when the technology underpinning their respective economies sits inside one of their jurisdictions and not the others. European, Canadian, Japanese, Korean, Australian and Gulf institutions all increasingly depend on American frontier models for the cognitive layer of their operations. But that dependence is not symmetric or consensual. The terms of access are set in California, governed by American export-control rules, subject to American sanctions enforcement, and immediately revisable by an American president for reasons that are no longer predictable. Every allied government is increasingly aware that its meaningful technological capabilities now run through centralized systems that it does not control and could be summarily blocked by a tweet.
Allies that built their digital strategies on the assumption of an open American technology base are discovering, late in the cycle, that the assumption was always conditional.
These concerns are not hypothetical. The US CLOUD Act already gives Washington legal reach into data held by American firms regardless of where it sits. Export controls on advanced semiconductors, designed to constrain Beijing, have produced collateral effects across the supply chains of every American ally — notably, a quest for digital sovereignty and away from Silicon Valley’s dominance. The 2025 reorganization of American AI policy under a more transactional administration in Washington has put non-American buyers on notice that frontier capability is now a chip on a negotiating table, not a protected service. Allies that built their digital strategies on the assumption of an open American technology base are discovering, late in the cycle, that the assumption was always conditional.
This is what Alex Karp, the chief executive of Palantir, was getting at in his April 2026 manifesto drawn from his book The Technological Republic. Karp argued that statecraft in this century is no longer organized around the atom but around AI, and that technology firms must start thinking in terms of national interest rather than profit alone. The framing is uncomfortable for Western liberals because it pulls the corporate sector explicitly inside the perimeter of national power. But Karp is largely correct about the direction of travel. Access to AI, like access to nuclear technology before it, is becoming a differentiator between great powers and everyone else, and it can be weaponized, including against partners. That is the deeper meaning of the AI sovereignty argument. It is not only that the Global South needs its own models but rather that the United States’ own allies need their own models, and the tech firms inside those allies need a credible answer to the question of what happens to their core operations when the terms of access change.
This is why the contest is not really about the digital economy. The digital economy is the visible part, but it is the cognitive layer that matters. Underneath the digital overlay, AI now enters every meaningful sector: defence procurement, energy-grid management, agricultural optimization, industrial design, drug discovery, financial settlement, logistics, intelligence and public administration. Whoever controls the cognitive layer is taking a position on the operating system of every other layer beneath it. This is the struggle for the commanding heights of this century’s economy in the original Leninist sense of the phrase. Energy was the commanding height of the twentieth century; cognition is the commanding height of the twenty-first. And, unlike with energy, the terms of access to the cognitive layer can be changed remotely, with software, between one fiscal quarter and the next.
The countries that have understood this dynamic are already moving. Switzerland released Apertus in September 2025, a fully open 70-billion-parameter model trained on its national supercomputer in more than 1,000 languages. France committed €109 billion at its February 2025 AI Action Summit. India has the IndiaAI Mission and BharatGen; Saudi Arabia has HUMAIN; the United Arab Emirates aims to be the world’s first fully AI-native government by 2027. None of this is protectionism. It is the recognition that capability without ownership is not capability at the scale that matters. Hosting is not owning. Renting is not building. Allies who do not build will be tenants, and the rent is going to rise, or the landlord may decide to end the lease altogether.
Canada is the cautionary case. Despite hosting roughly 10 percent of the world’s top AI research talent and three Turing-laureate-led national institutes in Mila, Vector and Amii, Canada attracts under two percent of global AI venture capital and ranks roughly twenty-third globally for AI infrastructure. Its 2024 budget committed $2 billion over five years to sovereign compute, about two percent of what France committed in a single weekend. The nation's recent AI strategy highlights an important first step, but talent without an industrial policy is not sovereignty: it is an export pipeline.
Additionally, if the AI contest is the present chapter of the two-systems story, quantum computing is the accelerant. Recent research suggests that breaking the cryptography securing bitcoin, Ethereum and much of internet traffic may require far less compute than once assumed, pulling “Q-Day” closer — reportedly with help from AI itself. The dynamic runs both ways: AI is shortening the path to practical quantum, and practical quantum will, in turn, accelerate AI. Whichever bloc arrives first does not gain an edge; it gains a regime change. That is what compresses the window for deliberate choice from a decade to something shorter
When Risk and Strategy Fuse
The pattern across all three frontiers — AI, finance and cryptography — is the same. The marginal cost of a critical capability is collapsing; the institutions designed to govern it are moving slower than the rate of change; the contest for who sets its terms is being run by two states with different theories of how the next century should work, and the terms can be revoked. Further disrupting these three seismic forces will be a steady cadence of cascading shocks: contests over Taiwan, the Arctic, the South China Sea, Ukraine and the Middle East producing sanctions surprises, export-control snapbacks, cyber operations against operational technology, sabotage of undersea cables and energy infrastructure, the next pandemic, the next financial accident, the next AI-enabled disinformation crisis on the eve of an election somewhere consequential. Each shock will arrive faster than the institutional cycle designed to respond to it. The defining problem of strategic stability in this period is that the response time of the international system is now slower than the rate of disturbance.
There is a temptation, when faced with conditions this complex, to retreat into either denial or fatalism. Neither out is available to anyone with fiduciary or constitutional responsibility. The harder discipline, and the one this moment demands, is sense-making: integrating signals from across domains that previously sat in separate silos, identifying which combinations matter, and acting on partial information at speeds the environment will tolerate.
This is, paradoxically, where AI may turn out to be most useful. The same technologies driving the disorder are also the most plausible candidates for navigating it. Frontier systems can ingest news, sanctions filings, scientific literature, satellite imagery, financial data and primary documentation at scales no human team can match, and surface patterns across domains that no individual specialist would notice. They can run scenarios, stress-test assumptions, and translate between the vocabularies of geopolitics, technology, finance and operations. Used well, they extend the cognitive bandwidth of the people responsible for steering institutions through turbulence.
But this is also the strongest argument for achieving sovereignty over the models themselves. An institution whose sense-making capacity runs on infrastructure it does not control is not making sense; it is being told what to make sense of. A government other than the United States that uses an American model to assess American sanctions exposure has, in a serious way, ceded a piece of its own statecraft. A non-Chinese board that uses a Chinese model to assess its supply-chain risk is doing the same in the opposite direction. The integrity of the cognitive layer matters as much as its capability, and that integrity rests, in turn, on provenance: on whether the data the model was trained on is reliably attributable to humans rather than to other AIs. As Ilia Shumailov and colleagues demonstrated in Nature in July 2024, models collapse when trained on recursively generated data, losing the tails of the distribution first — the rare cases that make a model capable of surprising its operator. That is precisely the property an institution most needs in a high-uncertainty environment.
Used carelessly, AI will accelerate the disorder. Used carefully, on infrastructure the user actually controls, with cryptographic and epistemic hygiene as a discipline rather than an afterthought, it is the most powerful sense-making tool any institution has ever had access to. The promise is real. It is also conditional, and the conditions are not the ones the technology vendors prefer to discuss.
A Choice for Decision Makers
The reordering will not wait for institutional readiness. Two states are setting the terms of the cognitive economy; everyone else, including America’s own allies and the firms inside them, is choosing between rented capability and the much harder work of building their own. The window in which that choice can be made deliberately, rather than under shock, is the next decade. Quantum is closing it. What follows are the questions that anyone with constitutional or fiduciary responsibility for an institution caught in this contest has to start asking, and answering, now.
For governments: Is your industrial policy designed to rent capability or to own it, and can your procurement rules attract the long-horizon capital that built earlier strategic industries? Do your security, economic and foreign-policy machineries decide in a shared forum or in separate silos on separate cycles? For corporate boards: Which of your critical processes now depend on a foreign-owned frontier model whose terms of access can be revised unilaterally? Where do your data, customers and suppliers sit, and what will change if any of those jurisdictions realign? And is there a single function with the authority to integrate technology, geopolitical and operational risk, or is that brief still split between the chief information officer, corporate affairs and audit?
Geopolitical risk, technological risk and demographic change are no longer externalities to be quarantined from core decisions: they are the core decisions. The discipline this period demands is to act as though every domain matters at once, because increasingly every domain does.
