The Geopolitical Chokepoints Shaping the Future of AI
Published on 25.03.2026
The Geopolitical Chokepoints of Artificial Intelligence
TLDR: Frontier AI capability is no longer just about algorithms — it is increasingly gated by access to advanced chips, enormous electricity supplies, and billions in capital. This piece from AI Supremacy, featuring a 4,000-word guest contribution from Oxford researcher Julian Alexander Brown, maps out exactly how geopolitics, semiconductors, and energy are converging to concentrate AI power in a very small number of nations.
Summary:
Alright, let me tell you, this is one of those pieces that makes you sit back and go "oh, right, atoms still matter." We spend so much time talking about transformer architectures and training recipes and benchmark scores, and then reality walks in and says "hey, did you know helium is a critical input for semiconductor fabrication, and a huge chunk of the global supply just got stranded because of the Strait of Hormuz closure?" That is exactly what is happening right now in early 2026. Qatar supplies about 34 percent of global helium, and South Korea — home of Samsung and SK Hynix, who together produce over 60 percent of global memory including the High Bandwidth Memory that goes into every Nvidia AI GPU — sources nearly 65 percent of its helium from Qatar. So a geopolitical crisis in the Middle East is creating a semiconductor crisis in East Asia which is creating an AI infrastructure crisis everywhere. The industry is calling it "RAMageddon," and it is not just a catchy name.
The guest contributor here, Julian Alexander Brown, who is doing his master's at Oxford and has worked at the White House and the State Department, lays out a really compelling framework. His core argument is that for the past decade, AI progress appeared to be driven by things that spread easily — papers, open-source code, algorithmic breakthroughs. And that was true for a while. But the binding constraints have shifted. What matters now are the things that do not diffuse: physical chips, massive amounts of reliable electricity, capital at a scale that only a handful of nations can muster, and the political relationships that determine who gets access to what. Frontier AI data centers draw between 100 and 500 megawatts of power around the clock. A single 300-megawatt facility consumes about as much electricity per year as a mid-size city. And the new projects being announced are even bigger, designed from the start to scale toward gigawatt-level demand.
Brown addresses the DeepSeek moment head-on, and I think he gets it exactly right. Yes, DeepSeek showed that algorithmic efficiency can compress timelines and deliver impressive benchmark performance at lower compute cost. But narrowing the gap is not the same as overtaking the frontier. The leading U.S. labs absorbed similar techniques and applied them to much larger compute budgets. More importantly, the hardware gap is widening, not shrinking. U.S. export controls are designed to prevent China from accessing each successive generation of frontier chips — Blackwell and beyond. As those hardware asymmetries compound generation over generation, the window for software-only leapfrogs gets narrower, not wider. Brown calls this the Jevons paradox of AI: better efficiency does not reduce total compute demand, it increases it, because it becomes economically viable to scale even further.
The piece also maps out the semiconductor chokepoint ecosystem in impressive detail. ASML in the Netherlands makes the extreme ultraviolet lithography machines that are essential for manufacturing chips at five nanometers and below — each one costs over 200 million dollars, and the next-generation High-NA EUV tools approach 400 million. China has never received a commercially viable EUV tool. TSMC in Taiwan fabricates over 95 percent of the world's most advanced semiconductors. Japan controls critical manufacturing equipment through firms like Tokyo Electron and Nikon. The U.S. controls chip design, EDA software through Synopsys and Cadence, and the export control regime that ties it all together. And then there is the capital intensity: announced foreign direct investment in data centers exceeded 270 billion dollars in 2025, with individual campuses requiring 5 to 10 billion in upfront investment. This is not a game you can enter cheaply.
What I find most thought-provoking is the discussion of China's dual-loop strategy. On the digital side, China has gone all-in on open source — Alibaba's Qwen models now have over 100,000 derivatives on Hugging Face and account for an estimated 80 percent of U.S. AI startup usage. On the physical side, China is deploying AI across its massive manufacturing base, generating proprietary industrial data that feeds back into model improvement. U.S. export controls primarily target the digital loop — restricting training compute — but they are not well suited to addressing the physical loop of deployment-driven data creation across China's factories and logistics networks. China also produces nearly half of the world's top-tier AI researchers now, and its institutions put out more AI research publications than the U.S., UK, and EU combined. The piece lists a truly sobering number of additional bottlenecks beyond chips and energy: advanced packaging with 52-week lead times, grid capacity with 4 to 10 year wait times for high-voltage interconnections, 2.5-year lead times for large transformers, critical shortages of liquid cooling components and specialized labor. It is chokepoints all the way down.
Key takeaways:
- The Iran War and Strait of Hormuz closure have created an unexpected helium shortage that directly threatens semiconductor fabrication, particularly HBM production in South Korea
- Frontier AI capability is now gated by physical and institutional chokepoints — chips, energy, capital — not by algorithmic knowledge which diffuses freely
- U.S. export controls target the entire semiconductor ecosystem, not just individual chips, with allied cooperation from the Netherlands (ASML), Japan, South Korea, and Taiwan (TSMC)
- Algorithmic efficiency triggers a Jevons paradox: it makes scaling more economically viable, increasing total compute demand rather than leveling the playing field
- China's dual-loop strategy (open-source digital plus manufacturing physical) may build AI advantages that are independent of access to frontier training hardware
- New data center announcements fell by half in Q4 2025, and only 33 percent of disclosed data center capacity is under active development
- The industry faces "RAMageddon" — a structural shortage of HBM and DRAM driven by reallocation of silicon wafer capacity toward AI
- ARM announced its first physical AI chip (ARM AGI CPU), departing from its 35-year licensing-only business model
Why do I care: If you are building anything that depends on cloud compute, AI APIs, or GPU availability, this piece is a wake-up call that the infrastructure you rely on is subject to geopolitical forces far beyond your control. The semiconductor supply chain is not just a hardware story — it directly determines what models get trained, where inference happens, and ultimately what software is possible to build. As a senior developer or architect, you need to be thinking about multi-cloud strategies, on-premise inference options, and model efficiency not just as engineering choices but as hedges against a world where compute access is increasingly politicized. The ARM AGI CPU announcement is also worth watching closely — if ARM succeeds in selling physical chips optimized for agentic AI workloads, it could reshape the inference hardware landscape in ways that matter for how we architect production systems.