SPEAKER_1: Alright, so last lecture we established that DeepSeek doesn't just compete on architecture — it actually repriced the market. Competitors had to cut their own API costs after R1 dropped. I want to stay on that thread today and really understand the economic mechanism. How does a single lab's pricing decision restructure an entire industry? SPEAKER_2: That's exactly the right question to pull on. And the short answer is: DeepSeek didn't just offer a discount. It exposed that the prevailing price of AI inference was structurally inflated — not by necessity, but by the absence of real competition. SPEAKER_1: So what's the actual magnitude of the price drop we're talking about? SPEAKER_2: R1 came in at 96% cheaper than OpenAI's o1 for comparable reasoning tasks. Inference costs overall — across DeepSeek's model lineup — ran 20 to 50 times cheaper than any frontier lab. That's not a promotional rate. That's what the architecture actually costs to run, because fewer active parameters per token means less compute per query. SPEAKER_1: And that connects directly back to the MoE efficiency we covered — sparse activation keeping the per-token compute low. SPEAKER_2: Exactly. The economics follow from the architecture. When you're only firing 37 billion parameters instead of the full 671 billion on every token, your marginal cost per query drops dramatically. DeepSeek didn't choose to be cheap. It built something that is structurally cheap to operate. SPEAKER_1: So here's what I think our listener might find counterintuitive — how does dropping prices this aggressively actually increase market share? SPEAKER_2: This is where it gets interesting. In AI, the demand curve is highly elastic at the high end. When inference costs are high, entire categories of users — startups, independent researchers, developers in emerging markets — are priced out. They don't use less AI. They use none. Drop the price by 96% and you don't just take customers from competitors. You create a new market that didn't exist before. SPEAKER_1: So it's less about stealing OpenAI's customers and more about unlocking a completely different tier of users. SPEAKER_2: Right. And that's the primary economic mechanism here — what economists call demand expansion rather than demand substitution. DeepSeek's pricing didn't just redistribute existing AI spending. It activated latent demand from builders who had real use cases but couldn't justify the cost at previous price points. SPEAKER_1: How did OpenAI and Google actually respond? Because they can't just ignore a 96% price gap. SPEAKER_2: They cut prices. OpenAI reduced reasoning model costs after R1 launched. Google followed. Anthropic adjusted its positioning. The competitive pressure was immediate and public. What's significant is that these cuts happened within weeks — which tells you the previous pricing wasn't cost-floor pricing. There was margin being extracted that the market had simply accepted because there was no alternative. SPEAKER_1: That's a striking admission embedded in those price cuts — that the old prices were... optional. SPEAKER_2: Essentially, yes. And for our listener tracking the long-term picture, that has a compounding effect. Once the market anchors to a lower price level, it's very hard to raise it again. The token economy — the per-query pricing model that the entire AI industry runs on — has been permanently reset downward. SPEAKER_1: What are the longer-term economic consequences of that reset? Because cheaper AI sounds good, but there have to be second-order effects. SPEAKER_2: Several. First, the viable business model for AI labs shifts. If inference margins compress to near zero, revenue has to come from volume, from enterprise contracts, from adjacent services — not from charging premium rates per token. Second, the barrier to building AI-native products drops significantly, which accelerates the number of startups that can actually reach product-market fit without burning through capital on API costs. SPEAKER_1: And on the flip side — does this create pressure that could actually harm the ecosystem? Like, can smaller labs survive if the price floor keeps dropping? SPEAKER_2: That's the real tension. If inference becomes a commodity priced near marginal cost, only labs with structural cost advantages — like DeepSeek's architectural efficiency — can sustain it. Labs that rely on brute-force compute without equivalent efficiency innovations get squeezed. It's deflationary for the industry in a way that rewards architectural discipline and punishes capital-heavy approaches. SPEAKER_1: So the factors that let DeepSeek offer these prices — it's not just one thing, right? What's the full picture of why they can do this when others can't? SPEAKER_2: It's a stack of decisions, not a single trick. Sparse MoE activation keeps per-token compute low. MLA compression reduces memory bandwidth costs. A lean team of roughly 100 people keeps operational overhead minimal. No venture capital means no pressure to inflate margins for investor returns. And a training cost of around $6 million for V3 means the capital recovery threshold is orders of magnitude lower than competitors who spent hundreds of millions. SPEAKER_1: So for Yunying and everyone following this course — what's the one thing they should carry forward from this economic picture? SPEAKER_2: DeepSeek's pricing strategy has initiated a structural reset in AI pricing. Advanced AI is now accessible to startups and researchers who were previously priced out entirely. The labs that survive this shift will be the ones that built efficiency into their architecture from the start, not the ones that assumed high margins were permanent. The cost of intelligence just collapsed, and the entire industry had to respond.