Placing AI in the lineage of railways, oil and the internet, and reading it through Susan Strange's structural power, makes a useful point: AI cuts across security, production, finance and knowledge at once, so dependence on it is harder to unwind than dependence on any single sector. That seems to change what middle powers can realistically bargain for. Do you see governance leverage coming mainly from controlling compute, or from shaping the standards others build on?
An important analytical contribution. This article was selected among today’s Conflict Engineering Top 5 following comparative review under our published methodology. Its main strength lies in treating artificial intelligence not merely as a technology, but as an emerging infrastructure of power, governance, and international order.
The shift from capability to infrastructure is the right analytical move, and it holds up better if pushed one layer further down. Infrastructure is not only ownership and governance. It is also physical, and its geography is considerably narrower than the language of cloud and stack suggests.
Below the model layer sits a set of very few nodes: EUV lithography from a single Dutch supplier, leading-edge logic concentrated in Taiwan, high-bandwidth memory in South Korea, advanced packaging capacity in a handful of plants, and, increasingly, grid interconnection as the binding constraint on new compute. Sovereign AI programmes of the kind pursued by Singapore, the UAE and Saudi Arabia buy accelerators and build national data centres, which changes who holds the asset. It does not change who can make it. The vulnerability sits upstream, at fabrication, tooling and power, and buying more of the finished good does not reach it.
That matters because the redundancy in this system is more correlated than it looks. A distributed cloud footprint reads as resilience on a map, but if the fabs, the packaging lines and the memory supply fall inside the same contingency, geographic dispersion is bookkeeping rather than insurance.
The European case makes the same point from the other side. The AI Act is real leverage, but it operates at the demand layer, and demand-layer instruments bind only while the regulated firm wants access to the market. Autonomy is a full-stack property. Holding rules without holding compute produces influence over conduct, not control over dependence.
Strange's knowledge structure is the right frame. Its physical floor, though, behaves less like a network and more like a chokepoint system, and the analytical vocabulary developed for maritime straits transfers with surprisingly little loss. A chokepoint is not simply a narrow place. It is a point where volume, substitutability and transit time collapse simultaneously, so that the cost of a closure is set by how long the rest of the system can run without it. Hormuz, Malacca and Bab el-Mandeb qualify on those terms. So do EUV tooling, sub-5nm logic and advanced packaging, on shorter timelines and with even thinner alternatives. Reading AI infrastructure this way changes the governance question from who writes the rules to who can hold a passage closed, and for how long before the dependent parties are forced to concede.
AI militarization, especially with ML is risky because it can trigger actions without human judgment, here we go again the problem with North South divide as well
Nice overview of AI technology, it's infrastructure and global significance. This article also coveres problems in AI adoption and how it's crucial to develop own infrastructure to be independent in the new world
Placing AI in the lineage of railways, oil and the internet, and reading it through Susan Strange's structural power, makes a useful point: AI cuts across security, production, finance and knowledge at once, so dependence on it is harder to unwind than dependence on any single sector. That seems to change what middle powers can realistically bargain for. Do you see governance leverage coming mainly from controlling compute, or from shaping the standards others build on?
An important analytical contribution. This article was selected among today’s Conflict Engineering Top 5 following comparative review under our published methodology. Its main strength lies in treating artificial intelligence not merely as a technology, but as an emerging infrastructure of power, governance, and international order.
Methodology:
https://conflictengineeringhq.substack.com/p/conflict-engineering-top-5-methodology
The shift from capability to infrastructure is the right analytical move, and it holds up better if pushed one layer further down. Infrastructure is not only ownership and governance. It is also physical, and its geography is considerably narrower than the language of cloud and stack suggests.
Below the model layer sits a set of very few nodes: EUV lithography from a single Dutch supplier, leading-edge logic concentrated in Taiwan, high-bandwidth memory in South Korea, advanced packaging capacity in a handful of plants, and, increasingly, grid interconnection as the binding constraint on new compute. Sovereign AI programmes of the kind pursued by Singapore, the UAE and Saudi Arabia buy accelerators and build national data centres, which changes who holds the asset. It does not change who can make it. The vulnerability sits upstream, at fabrication, tooling and power, and buying more of the finished good does not reach it.
That matters because the redundancy in this system is more correlated than it looks. A distributed cloud footprint reads as resilience on a map, but if the fabs, the packaging lines and the memory supply fall inside the same contingency, geographic dispersion is bookkeeping rather than insurance.
The European case makes the same point from the other side. The AI Act is real leverage, but it operates at the demand layer, and demand-layer instruments bind only while the regulated firm wants access to the market. Autonomy is a full-stack property. Holding rules without holding compute produces influence over conduct, not control over dependence.
Strange's knowledge structure is the right frame. Its physical floor, though, behaves less like a network and more like a chokepoint system, and the analytical vocabulary developed for maritime straits transfers with surprisingly little loss. A chokepoint is not simply a narrow place. It is a point where volume, substitutability and transit time collapse simultaneously, so that the cost of a closure is set by how long the rest of the system can run without it. Hormuz, Malacca and Bab el-Mandeb qualify on those terms. So do EUV tooling, sub-5nm logic and advanced packaging, on shorter timelines and with even thinner alternatives. Reading AI infrastructure this way changes the governance question from who writes the rules to who can hold a passage closed, and for how long before the dependent parties are forced to concede.
AI militarization, especially with ML is risky because it can trigger actions without human judgment, here we go again the problem with North South divide as well
Nice overview of AI technology, it's infrastructure and global significance. This article also coveres problems in AI adoption and how it's crucial to develop own infrastructure to be independent in the new world