SPEAKER_1: Alright, so last lecture we established that DeepSeek-R1 proved reinforcement learning alone — no supervised examples — can produce a model that genuinely reasons. That was the technical story. Today I want to zoom out to the geopolitical one, because the implications feel enormous. SPEAKER_2: They are. And the thread connecting both lectures is the same: constraint forced innovation. DeepSeek built world-class AI under strict U.S. GPU export bans. That fact alone rewrote the assumption that raw compute access was the only path to frontier AI. SPEAKER_1: So let's name that assumption directly. The idea was that whoever controls the chips controls the AI race. DeepSeek just... broke that? SPEAKER_2: It cracked it, at minimum. The export controls were designed to preserve what analysts call a 'compute moat' — the idea that restricting access to high-end GPUs would keep frontier AI development concentrated in the U.S. DeepSeek demonstrated that architectural efficiency can partially substitute for raw hardware. You can't export-control a software insight. SPEAKER_1: That's a striking phrase. So what's the actual geopolitical mechanism here? How does one lab's technical achievement translate into a shift in global power dynamics? SPEAKER_2: It works through demonstration effects. Before DeepSeek, the implicit message to every government and enterprise outside Silicon Valley was: you need our infrastructure, our chips, our platforms. After DeepSeek, that message lost credibility. Countries watching this realized that dependency on external AI ecosystems isn't just a technical inconvenience — it's an operational, legal, and reputational risk. That's what's driving the sovereignty conversation. SPEAKER_1: And sovereignty here means something specific — it's not just 'build your own model.' SPEAKER_2: Right, it has four distinct components. Territorial — where data and compute physically reside. Operational — who manages and secures the systems. Technological — who owns the intellectual property. And legal — which jurisdiction governs the AI. A country that outsources any of those four is exposed. DeepSeek's success made that exposure visible in a way it hadn't been before. SPEAKER_1: So how many countries are actually responding to this? Because it's one thing to say the conversation shifted — it's another to see governments move. SPEAKER_2: The EU is the clearest example. Their Apply AI Strategy commits €1 billion specifically toward AI sovereignty — open-source models, homegrown frontier systems, and shifting critical infrastructure away from U.S.-dominated platforms. The target sectors are healthcare, defense, manufacturing, and public services. They're also using procurement reform in public administration to create demand for European AI solutions — essentially building a domestic market by policy. SPEAKER_1: That's a real commitment. But why is it counterintuitive that a lean Chinese lab ends up leading the open-source AI race? Because most people would have predicted the opposite. SPEAKER_2: The intuition was that open-source leadership required massive resources — the kind only well-funded Western labs could sustain. DeepSeek inverted that. A team of roughly 100 people, no venture capital, around 10,000 GPUs, produced models that are now the reference point for open-weights AI globally. The constraints — limited chips, limited capital — pushed them toward architectural decisions that emphasized efficiency and independence. This approach aligns with the broader geopolitical trend of AI sovereignty. SPEAKER_1: So the very restrictions meant to slow them down made their output more globally useful. SPEAKER_2: Exactly. And that has a compounding effect. The shift towards AI sovereignty is not just a policy concern but a strategic geopolitical move. Countries are now layering their strategies: use global frontier models where possible, build sovereign domain-specific or language-specific models where needed. Countries are now layering their strategies: use global frontier models where possible, build sovereign domain-specific or language-specific models where needed. SPEAKER_1: How does that layered approach actually work in practice? Because building a sovereign model sounds expensive. SPEAKER_2: It doesn't have to be. The distillation insight from R1 — that reasoning patterns from large models can be transferred into smaller ones — means a country doesn't need to train a 671-billion-parameter model from scratch. They can fine-tune a smaller sovereign model on domain-specific data, in their language, under their legal jurisdiction. DeepSeek's open weights made that pathway real. SPEAKER_1: So the open-weights release wasn't just a technical gift — it was a geopolitical one. SPEAKER_2: That's a precise way to put it. It gave every government and enterprise a foundation to build on without dependency. The AI agency framework — which distinguishes between access, control, choice, and leverage — captures this well. Sovereignty isn't binary. Countries can identify where to build independently, where to partner, and where to hedge. DeepSeek gave them a credible option in the 'build' column. SPEAKER_1: And what does this mean for global tech policy going forward? Because the EU is one data point — what's the broader trajectory? SPEAKER_2: The trajectory is from a unipolar AI world — centered on Silicon Valley — toward a multipolar one. Sovereign AI has shifted from a policy debate to an economic imperative. Governments are investing in local infrastructure, talent pipelines, and innovation ecosystems because they now believe it's achievable. DeepSeek is the proof of concept. The question every policymaker is now asking is: if a constrained Chinese lab can do this, what can we do with deliberate investment? SPEAKER_1: So for Yunying and everyone following this course — what's the one thing they should carry forward from this lecture? SPEAKER_2: DeepSeek's success demonstrated that AI sovereignty is achievable, reshaping the geopolitical AI landscape. The compute moat was real — until it wasn't. The compute moat was real — until it wasn't. And now every nation, every enterprise asking 'do we control our own AI?' has both a reason to ask and a model for how to answer it.