Kevin Chen has been an adjunct associate professor at NYU, teaching global political risk at the graduate level and fintech to undergraduates. He and Bruno Sergi, who teaches at Harvard, published “Can Singapore Become the Switzerland of the AI Chip Economy?” on the London School of Economics’ Southeast Asia blog this week. They argue that Singapore already is becoming exactly that, not by out-manufacturing anyone, but by making itself the indispensable connector of capital, talent, and supply chains that the chip economy actually runs on.
The argument from “Altitude 1”
Chen and Sergi’s case rests on a shift most people watching the AI chip race have missed. The competitive edge in semiconductors is moving away from who can fabricate the smallest, fastest chip and toward who controls packaging, memory optimization, software, and system integration, the layer that determines whether raw fabrication capacity turns into usable computing power. Nvidia’s constraint right now is not fabrication; it is packaging capacity. Micron has committed seven billion dollars to an advanced packaging facility in Singapore. GlobalFoundries has pledged four billion more for specialty semiconductor expansion there. Singapore already holds more than ten percent of global semiconductor market share, not by building the most advanced fab in the world, but by becoming the place where fabrication, packaging, capital, and talent meet and get coordinated into something usable. As Chen and Sergi put it, the decisive question is no longer who can manufacture the smallest chip, it is who can deploy computing power effectively at scale.
This is one piece of a larger project. Chen has published several more pieces this month, including an argument that export controls will not win the AI race with China. The pieces are consolidating toward a book chapter, currently in editorial review with Palgrave Macmillan. Taken together they describe a bottleneck that has moved. Restricting access to the most advanced chips does not resolve the underlying constraint. It just relocates where that constraint shows up next.
The same shift, one level down
We spend a lot of our time at the organizational level rather than the geopolitical one, but the pattern Chen and Sergi describe is one we keep running into under a different name.
Most organizations assume their AI adoption problem is a capability problem: the model is not good enough yet, or they do not have enough compute. That is rarely where the real constraint sits, a point we made at length in “Everyone’s Impressed by the Model. Almost Nobody Can Reach Their Own Data.” The constraint is almost always the integration layer, the data architecture, governance structure, and workflow design that determines whether a capable model actually reaches the person who needs to use it. This is the core claim behind what we call Tier 3 in our own framework, and it traces back to the argument we first made in “No Facts Inside the Building.” Most organizations are stuck at Tier 1, using AI as an assistive layer bolted onto whatever data infrastructure already exists, broken or not.
AI layered on top of broken systems does not transform them; it accelerates them.
Read side by side, Chen and Sergi’s argument and our own are the same structural claim made from two different altitudes. At the geopolitical level, the bottleneck moved from fabrication to the ecosystem that connects packaging, capital, talent, and data. At the organizational level, it moved from the model to the integration layer that connects raw capability to actual use. Neither shift is really about chips or about AI models specifically. Both are about what happens to a hard constraint once it gets solved: it does not disappear; it relocates to whatever was previously the second-hardest problem, which promptly becomes the new hardest one. Export controls that restrict chip access do not fix an organization’s ability to use the chips it already has, in the same way that a more powerful model does not fix an organization’s inability to reach its own data. The lesson from both altitudes is the same. Solving the visible constraint just reveals the constraint that was hiding behind it.
That is a genuinely useful framework for anyone trying to figure out where their own AI adoption effort is actually stuck, and it is why we would point readers toward Chen and Sergi’s work generally, not just this one piece. There are more articles coming as the book chapter takes shape, and we will be reading them as they land.
If you are working through where your own bottleneck has relocated to, we would welcome the conversation at info@stormkinganalytics.com.