Gillian Tett's recent insights into the era of geoeconomics underscore that today's power dynamics rest not merely on economic might but on strategic control points: finance in the United States and manufacturing in China, leaving technological dominance as contested terrain.
I’d like to pull on that last thread a little harder and suggest that today’s technology isn't simply another industry. Artificial Intelligence has introduced a fundamentally new form of power: intelligence itself, purchasable and scalable, making Rich Sutton's prescient "bitter lesson"1 (which one can roughly interpret as positing that "more compute means more intelligence") starkly geopolitical.
AI is not merely another technological advancement; it is an industrial-scale means of producing intelligence. Historically, a nation's capacity for innovation and decision-making was limited by human cognitive resources, rare geniuses, and scarce expertise. Today, however, the performance of AI systems, measured in their intellectual output, scales predictably with the raw computational power available. Intelligence, once scarce, is now increasingly a matter of capital investment in compute power.
Not everyone agrees with this position, of course. There is no shortage loud pundits, claiming that modern AI is actually all just smoke and mirrors. (In the meantime, AI systems continue to become demonstrably more capable, pundit opinion notwithstanding2.)
If you agree with me, geoeconomics looks different. Intelligence as a commodity means the key strategic nodes are shifting. Where 20th-century geopolitics revolved around oil, minerals, and currency reserves, the 21st-century battlefield may centre on GPUs, semiconductor fabrication, data centres, and scaleable/renewable energy.
Mapping New Nodes of Power
Consider semiconductor manufacturing, particularly advanced GPUs. Nvidia's chips, which the U.S. restricts from export to China, are a new kind of strategic resource like oil in the past century. Similarly, ASML's EUV lithography machines, the critical bottleneck in chip manufacturing, constitute another decisive node. Control over such nodes, achieved through alliances, export controls, and sanctions, allows nations unprecedented leverage over rivals' cognitive capabilities.
Energy, too, has shifted. Data centres (the “refineries” of intelligence) require massive amounts of stable, renewable energy. The ability to deliver reliable electricity at scale now translates into geopolitical advantage, putting regions with abundant hydroelectric, nuclear, or renewable potential at the centre of a new global power dynamic.
Integrated cloud computing infrastructure represents yet another possible node. Hyperscale platforms like AWS, Azure, and Google Cloud could be viewed as a kind of “digital central bank”, issuing computational credit globally. Their power to throttle or enable access mirrors financial sanctions but affects cognitive capacity rather than capital liquidity.
Explaining the AI Arms Race
Viewing AI through this lens clarifies several recent policy choices:
Export Controls: Restrictions on chip exports to China, often seen as narrow economic protectionism, instead appear rational when viewed as measures to prevent rivals from acquiring potentially transformative intelligence capabilities.
(Edit: Or not. Shortly after posting this, the New York Times reported that the restrictions on chip sales to China will be lifted!)
Massive State Subsidies: The U.S. CHIPS Act, EU Chips Act, and China's semiconductor funding become strategic investments to ensure national control over intelligence production.
Data Centre Competition: Sudden global rushes to secure GPU stocks or data centre locations are less speculative bubbles and more rational accumulations of critical strategic assets.
Strategic Implications (2025-2030)
One can imagine three distinct blocks emerging:
Dollar-CUDA Bloc: Led by the U.S., encompassing its allies leveraging American technology ecosystems. (Although who the U.S. defines as an ally these days is not always trivial to discern, so it may be that there are additional blocks like Euro-CUDA, too).
China's Sovereign Bloc: Investing heavily in indigenous chip design (RISC-V) and sovereign cloud services to bypass U.S. dominance.
Energy-Rich, Compute-Poor States: Countries like UAE, Qatar, Kenya, and Canada leveraging abundant renewable resources to trade power for intelligence capacity, gaining new geopolitical bargaining power.
“Power-grid-rich” nations now have an opportunity akin to oil-rich states of the 20th century. Canada, Norway, and others with substantial renewable or nuclear capacity could reposition themselves strategically as global intelligence hubs, hosting the data centres that power AI.
Policy Recommendations
To navigate this new geoeconomic landscape, policymakers in all but the greatest powers must:
Treat Compute as Strategic Infrastructure: Regulate computing infrastructure akin to energy grids and financial systems.
Align Energy and Compute Strategies: Site data centres strategically near clean energy sources and secure long-term power agreements proactively.
Strengthen and Broaden Alliances: Given the complexity and scale, no country can afford complete technological autarky. Alliances like those between the U.S., Netherlands, and Japan around lithography equipment will be essential.
Invest in Open AI Ecosystems: Supporting open-source AI models can mitigate risks of monopolization and allow smaller states to participate without building prohibitively expensive infrastructure.
The Era of Compute Diplomacy?
We are entering a fundamentally new era, one in which computational power equals national intelligence. Policymakers must urgently adapt frameworks and strategies to this reality. The geoeconomic lens suggests that the AI arms race is not a transient technological contest; it represents the dawn of an epoch where national power hinges on control over computational resources.
Understanding and adapting to this new geoeconomic landscape will define national security, economic prosperity, and global influence for decades to come.
In 2019, Rich Sutton taught us The Bitter Lesson : simple, general, models with access to greater computational power consistently outperform handcrafted, baroque, and apparently clever complex models. The secret to “more intelligence” is “more compute.”
This is, of course, utterly unsurprising to me. LLMs are Turing-complete and can thus compute any computable function. Full stop. If you believe that Turing-completeness is insufficient to reach “human level intelligence” then you are, like it or not, a kind of mysterian apologist. There’s no shame in that, but be clear with yourself that what you believe is that there is something “magic” about human intelligence — a spark of the divine, a soul, something beyond the physical.
Of course, some computationally universal systems will be much more efficient intelligence machines (like the human brain), and there are probably better architectures out there than the transformer, but these are matters of complexity, not computability.