EP 121
Inside Anthropic’s S-1: A Stock with “Humanity’s Future” as a Risk Factor?
October’s clash of optimism and pessimism, and Anthropic’s IPO 0:00
Chester Roh Today, as we record, is November 1, 2026, a Sunday morning. We’ve had a really tough October. Optimism and pessimism have mixed, sending the stock market swinging, as people ask whether our future will be good or bad. Those views seem to be clashing fiercely in the market. Amid all this, Anthropic has arguably become the bellwether of the AI market. Anthropic, ahead of a stock market listing, filed an S-1. We’ll also look at the numbers in it later.
First, one striking thing about that S-1 filing is that its risk factors section expresses concern about the future of humanity. That’s a remarkable thing to write when going public, unless you’re a nuclear weapons company. It reflects the debate over whether AI will bring a bright future or a dark one. People seem deeply divided on that. Seungjoon Choi will describe this as a split, so let’s take a quick look at that and then talk about the S-1 filing. Seungjoon Choi, are things getting better for us or worse?
Seungjoon Choi I don’t know. We’re at a kind of turning point. The mood these days is strange. This year feels especially strange. We’ve kept talking about mathematics, and it makes me think about the idea of everyone having their own AlphaGo moment.
GPT-6.1 Terra news and the OpenAI video 1:35
Seungjoon Choi So here’s some news: It didn’t get much attention, but GPT-6.1 Terra came out. Terra is good at making 3D things like this. This was made with a generative video model, then used for 3D modeling to show that Terra is capable, too. Gemini’s Carbon came out as well. What caught my eye was a video released yesterday. It was quite intriguing. It’s about a minute long, so let’s take a quick look. It’s called the Museum of What Could Be, a museum of things that might happen in the future, where a character writes something.
Jonghyun Park It was posted on OpenAI’s official channel.
Seungjoon Choi People are divided about how it was made. Was it made in 3D rather than with a video model? Was it just rendered incredibly well? That’s one debate. Another angle is a line that appears around the 36-second mark.
“Optimism achieved internally” and the Bogdanov–Takens bifurcation 2:37
Seungjoon Choi Optimism achieved internally. “Achieved internally” is a phrase Big Tech, especially OpenAI, uses a lot. “AGI achieved internally.” They’ve used it that way many times. So OAI refers to OpenAI while also standing for “Optimism achieved internally.” Now, there’s anticipation because there’s been talk of the next model, Bel, coming out this year. Sam Altman had already said that by year-end, the model after Astra would come out. So that builds anticipation. And in the video itself, another interesting thing is that people disagreed about what this graph shows.
I looked it up and found that it’s called the Bogdanov–Takens bifurcation. Apparently, it’s mathematics about how bifurcations occur. It dates back to the 1970s. Could it be sending a message about a fork in the road? There’s also anticipation that internal models are making breakthroughs in math and other areas. And on October 26, a problem was solved that could signal progress on the Exponential Time Hypothesis or P vs NP. With developments like that, some think the next leap could be a turning point for humanity. People are reading that much into it these days.
Big Tech’s rosy future and critiques of TESCREAL 4:02
Seungjoon Choi So if the rosy future Big Tech is selling actually comes true, There are new drugs, and we grow wealthier, everyone grows wealthier. Alongside that story, there’s also deep pessimism these days. There’s anticipation for Fable 5.5, too. Cereal devoted considerable attention to that pessimism for about an hour. Bernie Sanders appeared, as did Big Tech oligarchs. It likened them to priests and covered the criticism and resistance they face, including people like these in red robes. It explored how this has become a social issue. I felt the overall message was well researched, but I found the latter half more compelling. The earlier part felt a little dated to me.
As for these voices of pessimism, some see this as a revival of eugenics. Critics use the term TESCREAL. It covers transhumanism, the pursuit of limitless progress through technology, singularitarianism, cosmism, Silicon Valley rationalism, effective altruism, which Anthropic is said to be closely tied to, and longtermism. They criticize it from that perspective, and it’s becoming a social issue. They talk about where opinions diverge, but setting that aside, I brought some numbers. Hearing both sides made me curious.
Limits to generative AI adoption and user resistance 5:42
Seungjoon Choi Apparently, there are 8.3 billion people in the world. Of those 8.3 billion, 6 billion reportedly use the internet. You need a computer to use the internet, right? So I don’t know how many computers there are, but 6 billion people use the internet, and they also make up a market. These are rough figures. Apparently, about 5 billion people use smartphones. Would that be around 70% of humanity? They use smartphones. But free generative AI users number around 2 billion, I’m told. That’s also a very rough estimate. Paid generative AI users number 100 million. It may be far fewer, but that’s the divide right now.
This has to do with access and people’s reluctance, and I think that reluctance matters. People who dislike the internet are much rarer now, though there may have been more in the past. But people who dislike smartphones seem a little more common. Think about it the other way around. Smartphones have become essential, yet people still tend to want some distance from them. For a moment, I wondered if that reflects a growing sense of aversion. Or a feeling that you don’t need them.
Chester Roh It could be aversion, but couldn’t we also see it simply as a technology spreading?
Seungjoon Choi It hasn’t spread this far yet, and I think resistance is especially strong around this point. Of course, because of smartphones—social media, really— there have been protests. But have people protested over smartphones themselves? I’m not sure. But with social media, there’s been a lot of talk about its dangers. And with AI, I feel those concerns are coming to the fore these days. I also want to connect this
AI agent investment and the job replacement dilemma 7:32
Seungjoon Choi to last week. I couldn’t join you, but you discussed the economic reasoning behind things that have drawn too much investment and things that still need more. Expectations are enormous right now, aren’t they? But we don’t know what will make it happen. Aren’t many people betting on agents right now? They expect agents to create a new economy.
But what agents do is work, right? And since they replace jobs, if today’s investments produce economic returns by doing work, and that work doesn’t leave existing jobs intact but replaces them, how does that investment succeed? Wouldn’t a lot have to be replaced? Looking at the numbers, I wonder if that’s why people are so worried these days.
Jonghyun Park So optimism about AI itself—right. Since replacing labor is likely its biggest effect, it could be bad for society. Is that what you mean?
Seungjoon Choi Right. If the investment succeeds, skipping a few steps, it ultimately leads to replacement. And if people lose their jobs as a result, they won’t be able to spend, right? That would be an own goal, wouldn’t it? So right now, I feel there’s a dilemma here. What do the rest of you think?
How looms changed labor and markets over time 9:00
Chester Roh What Seungjoon Choi said is quite right, but back in the mid-1800s, when most people’s clothes had to be made by hand using looms and weaving machines, there was a certain stability in the market. Only so many clothes could be supplied, and clothes had a certain value. Of course, the population was growing rapidly then too, but demand for clothes couldn’t outgrow the population. So the market must have been in a kind of equilibrium.
But when machines first arrived and clothing production surged, clothing prices fell for a time, and the people employed in making clothes lost their jobs. But over the long term, for humanity as a whole, clothing went from an essential necessity to something people buy by choice. Even we now have dozens of outfits hanging in our closets, don’t we? We wear clothes for more than warmth, and the market has grown much bigger.
I’m bringing this up as an example for knowledge workers. For the people who made clothes back then, over the next 10, 20, or 30 years, the impact on their lives was sharply negative. But for humanity as a whole, after 50 to 100 years, it turned into an enormous gain. Entire new markets emerged on a massive scale. Changes like these took a century back then, so historians could analyze the process and explain what happened. But as we saw earlier, this change is happening within a single year.
Labor demand in the AI era through Jevons Paradox 10:50
Seungjoon Choi Right.
Jonghyun Park I had a similar thought. We discussed this at a research meeting this week. The topic was Jevons paradox, in the context of the first steam engines. When steam engines first appeared, coal use became much more efficient. But that didn’t leave us with surplus coal. In fact, as steam engines multiplied, coal use grew more efficient, yet coal became even scarcer. That was the paradox we discussed.
The same could apply to human labor. People say AI is making work more efficient and might replace us all. But in the future, we might actually end up working more. If AI lets us create more value with the same amount of work, we might have an incentive to work more and create even more value. That could happen. So will that be the future? No, I think people have different views on these issues, ranging from pessimism to optimism. Everyone seems to have their own opinion.
Seungjoon Choi I can’t really picture what that would look like, and I think that’s the problem these days. So new jobs like that could emerge, and we might even develop a new concept of work itself. But agents would do that work too, wouldn’t they?
Jonghyun Park If I think about what I’m doing
Video production, efficiency, and limited demand 12:08
Jonghyun Park and consider some examples and possible futures, take video, for example. Traditionally, making a video involves filming, editing, and countless other steps. But AI has come along, and those steps are being compressed dramatically. As a result, many people who used to find it hard to make videos are making more of them now. So the total amount of content is growing, but the number of people making videos and the time they spend making them haven’t decreased. The process itself has become more efficient, though.
Seungjoon Choi But demand is limited, right? There are only so many people willing to pay for it. As you said last time, even if scarcity leads to more production, that probably won’t translate directly into a livelihood for each person, will it?
Jonghyun Park I’ve been thinking about that a lot too. With video, the total population and the total time people spend watching YouTube aren’t going to increase.
Niche content and the asymmetries technology removes 13:08
Jonghyun Park Those are fixed. But broadcasting used to be highly centralized, and now it’s splintering into niche content. That gives people more ways to find the content they like. I think that’s one way things will change, at least. And if I imagine something a bit more radical, will people be the only ones watching videos? I think AI might watch them too in the future.
Chester Roh But this isn’t unique to AI. It’s what technological progress does. Technological progress always removes some kind of asymmetry. Gutenberg’s printing press did, and so did the steam engine. If you look at newer media, including YouTube, they remove some kind of asymmetry too.
What KTX reveals about concentration and the long tail 13:51
Chester Roh Let me give you an example: the KTX. The KTX made same-day trips possible across the country, right? So when it was first built, people thought many of Seoul’s functions would spread to other regions. But we tend to pursue an ideal. We always picture the ideal outcome. Yet when an asymmetry disappears like this, things always become more concentrated.
The places hurt most by the KTX were regional hub cities. Gwangju, Jeonju, and Cheongju, for example. Those hub cities used to serve the needs of people in the surrounding regions. If someone gets a little sick, they can just take the KTX to Seoul. If they want to buy something, they can go to Seoul. If they want to see a show, they can go to Seoul. So Seoul’s functions became far more concentrated, and it generated much more added value. Things became even more concentrated. The hub cities lost out. But conversely, as Jonghyun Park just said, those smaller places further down, townships and county-level towns, you know? They actually benefited. Tourists who never would have visited before now come.
When we think of an industry as stable, we picture a few stars, a middle tier of people, and people below them supporting it all. We’re used to industries with that stable pyramid structure. When technology removes information asymmetry, what always happens is the peak rises higher, the middle collapses, and those at the edges benefit. It creates a very sharp long-tail pattern. We’ve seen this happen everywhere. That’s what we can generally observe, and knowledge work will probably be no different, will it?
People who know their field, shape public opinion, and use AI to make a hundred videos at a time— people like Jonghyun Park—will become more popular and get more work. And people who have never done it, who couldn’t make anything themselves, can use this to create something of their own and gain recognition from people around them. Meanwhile, the many people in the middle who did this for a living will go under.
Collateral damage and the cost of a transition across generations 16:06
Seungjoon Choi That’s why people are now talking about “collateral damage” and saying, “Don’t call us collateral.” As this shift or progress happens, they don’t want to be the ones who suffer collateral damage. I think that’s why we’re hearing those voices.
Chester Roh But as always, this tends to resolve itself over generations. The people affected almost never find another job or something else to fall back on in their own time. They live on government support through fiscal policy, and in the next generation, new kinds of jobs emerge.
So I think what we should focus on now is this: One approach is to worry about disappearing jobs and regulate to keep them from disappearing. The other, if they’ll disappear anyway, is to move quickly in a different direction. But there’s a radish right in front of us that needs cutting, and people are saying we shouldn’t cut it. Because that’s the easiest thing to say. But policies that forbid it won’t stop it. People will make the efficient choice. They say one thing, demanding in their own field that people use only human-made work, then turn around and buy things on AliExpress. It’s contradictory. So the market will inevitably find its equilibrium. I believe we should prepare quickly.
Employment rules and tax redistribution in the age of automation 17:35
Seungjoon Choi I tried a thought experiment this morning that I haven’t resolved yet. Consider restaurants. Now there are ordering kiosks at every table, right? That must have eased the serving workload somewhat. Taking that idea further, robots cook and robots serve, and restaurants can easily prepare and serve all of it at a high standard. I wondered whether, even then, the government should require restaurants to employ people through regulation.
Chester Roh Companies that gain wealth from this, like OpenAI or Anthropic— we’ll see leading companies like them emerge in Korea too. We’ll see leading AI companies here. Take SK Hynix and Samsung Electronics, for example. I think they’re paying nearly a third to almost half of our country’s tax revenue this year. Employing only about 50,000 to 60,000 people, just two companies pay two-thirds of South Korea’s tax revenue? That’s an enormous concentration. But the government then redistributes that revenue through fiscal policy. That system needs to get moving quickly.
Seungjoon Choi Right. And this isn’t unique to Korea. If this investment succeeds, we could see the government–Big Tech collusion that’s such a common sci-fi cliché. People picture everything becoming concentrated and centralized, with those institutions gaining unchecked power. That’s the picture they have in mind, and I think it worries them. Not right away, though.
Chester Roh At least we have democracy as a tool.
Seungjoon Choi I worry that if this goes too smoothly, it could cause problems. I’m hopeful too, but also very worried. There are signs that 2026 may be a kind of bifurcation, a turning point. That’s the feeling I’m getting. I can’t even imagine 2027.
Chester Roh This was over a year ago, but a close older friend of mine, a former VP at Kakao, Shin Min-gyun—I call him Min-gyun— told me something.
New value in sports and caregiving in an age of abundance 19:44
Chester Roh For a while, he kept saying the future was sports. Take the 100-meter dash, for example. Is a person faster than a car? Of course the car is faster. Machines outperform us in every sport. Yet we make rules to create sports, build stories around them and give them an arena. That’s how the industry grows.
So, as Seungjoon Choi said earlier, if AI and robots do everything for us, the so-called age of abundance that Elon Musk and Sam Altman talk about arrives. Wealth would become extremely concentrated, but taxes would send some of it back to the government, which would use fiscal policy to support everyone. But then I wonder: Do we just live on a government allowance until we die? I don’t think it has to be that way.
I think of entertainment, sports, art, or what we call care work. Caring for people. Traditionally, a grandmother looks after her grandchild. People say that creates no economic value. But what if we started asking, “How many hours a week do you care for your grandchild?” I think the government should pay a substantial salary for that. And an economy could emerge where people naturally expect to pay for things like that. Today, it’s “We’re family, so it’s free,” or “You did the dishes, so it’s free.” But what people provide through their labor or emotional care isn’t valued in money today. In the future, it could be worth a great deal. That kind of era could come. I think it should.
Claude’s refusal policy and the centralization of AI 21:39
Seungjoon Choi But the problem with centralization is clear from Claude’s recent policy. Reverse engineering and cloning games had been a big issue for a while. But apparently, in the past few days, something was added to the system prompt or elsewhere that nudges it to refuse those requests. People pushed back against that too. Some opposed cloning, but also opposed blocking it. “Why do you get to decide?” “A tool should be neutral.” “That’s for me to judge.” So when this becomes centralized, decisions about sensitive matters, sometimes ethical ones, are all made by a tech company.
Chester Roh Maybe it’s because I believe in markets, but won’t another company offer that? Won’t a Chinese company provide it all?
Seungjoon Choi Some people did say, “Just use open weights.” Anyway, that’s what I’ve been thinking about lately. I’ll hand it back to Jonghyun Park to hear about the S-1, then I’d like to return to this.
An overview of the Anthropic S-1 filing and its risk factors 22:42
Chester Roh Anthropic’s S-1 filing is the talk of the town. It’s all over the news. Shall we hear about the filing? Jonghyun, could you give us a quick rundown?
Jonghyun Park I’ll sum it up quickly.
Seungjoon Choi What’s an S-1? It’s something you file to enter the stock market, right?
Chester Roh It’s basically a document saying we’re going public, with everything we can disclose about the company and all that information bundled together in one publicly available book. It must be incredibly thick. I downloaded it but haven’t read it.
Jonghyun Park It’s about 500 pages long.
Chester Roh Shall we look at the main points, Jonghyun?
Jonghyun Park Since it’s a filing for going public, it’s packed with figures, especially financial ones. But unusually, as you just mentioned, it goes beyond centralization to the threats AI models could actually pose. Those risks are listed in it. It even says they could pose a real danger to humanity, including the risk of losing control. It also mentions things like the Hugging Face hacking incident in the filing.
Seungjoon Choi It’s strange to see that kind of thing in there.
Chester Roh Yes, it’s entirely plausible.
Jonghyun Park Right. As for these risks, everyone seems to be aware of them. Getting back to the main points of the filing, there’s the fact that they’re going public. Then there’s Anthropic’s valuation: it’s incredibly high. It looks set to go public at nearly 2,000 trillion won. That implies AI will be used much more in the future and replace a lot of human labor. Those are the implications. So let’s look at those expectations through the numbers.
Anthropic’s revenue and customers doubling each quarter 24:29
Jonghyun Park First, Anthropic’s revenue is growing incredibly fast. The growth may have slowed slightly, but it’s still growing incredibly fast. It’s doubling quarter over quarter, and if you look at the Q2 figure, it’s over 10 trillion won— close to 15 trillion won. Revenue is growing remarkably fast.
Chester Roh People hope revenue keeps growing at that pace, but growth slowed a bit in Q3. Some say it will probably slow further in Q4. But I also heard about this issue. Anthropic blocked API calls from Chinese companies and others using them for distillation. Revenue would have been higher if they’d allowed those calls, so blocking them may have had an impact. I heard that somewhere.
Jonghyun Park Yes, I heard they blocked a lot of it with Opus 5.5. Looking at it year over year, first-half revenue was 14 times last year’s. So it really is growing fast. And the number of customers spending large amounts has grown a lot. The customer count is rising, and what each customer pays is growing tenfold. So the growth is solid. That’s how I see it.
Chester Roh Customers spending over 100K a year— about 150 million won— numbered just 950 last year, but that’s grown tenfold this year. And 1M is 1.5 billion won. There are more customers paying over 1.5 billion won, too.
Jonghyun Park Yes, as many as 1,900.
Seungjoon Choi Is that from the API side?
Jonghyun Park There’s another revenue figure: customers can pay for subscriptions or pay to use the API,
The API behind 80% of revenue and the subscription math 26:01
Jonghyun Park but subscriptions are usually for individuals. Of Anthropic’s revenue, 80% reportedly comes from API usage.
Seungjoon Choi Then it might not be developers using it.
Jonghyun Park Developers at companies probably use it the most.
Seungjoon Choi So rather than using the API directly, I did briefly wonder if there were services like that for a moment. Developers all use subscriptions.
Chester Roh Yes, under the terms, if you run a service, you can’t use a subscription account. You have to use the API. Say I built a service for legal documents or for finance. For example, if I’m Harvey, then I have to provide that service through the API only.
Seungjoon Choi So there must be more services like that now.
Chester Roh Many popular character chat services use Anthropic too. So that likely accounts for a lot as well. It’s surprising that subscriptions make up less than the API, but it makes sense. Right?
Jonghyun Park Yes. If you look at the pricing, subscriptions are priced absurdly low compared with the API, so that probably has a big effect too.
Chester Roh Yes. But while they’re cheap, if you look closer— I have no objective evidence for this, but the reason Anthropic and OpenAI keep their $200 plans and similar offerings may be that many subscribers use them heavily, but many barely use them at all. Those unused subscriptions are all margin. That calculation is probably built into those plans. I think.
Seungjoon Choi The plans changed a few days ago, right? To Max and Ultra. Only the names changed.
Chester Roh Then it’s the same policy as OpenAI’s.
Jonghyun Park No. Probably not the pricing.
Seungjoon Choi Right, they just renamed the 5X plan Max, and then Ultra— they call the existing $200 plan Ultra. They must be preparing something else.
Chester Roh All right, back to the financials.
Anthropic’s net losses, operating cash flow, and token margins 27:45
Chester Roh The net loss is $80.7 billion, but operations are profitable.
Jonghyun Park Yes, let’s look at the financials once more. First, a quick look at the numbers: revenue is the money they brought in. To earn that, they must spend a lot on compute and other things, and a lot on model development too. So how much are they actually making? Revenue was about $16B, with a loss of around $0.7B, so they lost roughly $1B. That’s after subtracting compute, labor costs, and so on.
Next, operating cash flow. If you subtract direct costs but leave out future expenses like research, and account for the cost of running compute, how much cash is left from revenue? About $4B, or nearly 5 to 6 trillion won in cash. That’s how I read it. Then there’s this net loss. The reported loss is enormous, close to 100 trillion won. Could you explain that?
Chester Roh That probably relates to, before the IPO, convertible bonds that had built up over time or SAFEs they used to raise funding. I don’t know the exact structure myself. They probably recorded a temporary loss on the books to make an adjustment. So it probably only appears before the listing, and that number won’t show up next quarter.
So what we should focus on is that level of revenue. Looking at it, it’s fair to say profit is positive. They made about $1.6B in profit in the first half. People thought Anthropic must be losing huge sums because of compute costs, but the figures show it’s profitable. That’s what has everyone surprised. Wow, the token margin is nearly 70%. NVIDIA has a 70% margin, The company selling the model takes a 70% margin, and then the company selling the service takes a 70% margin too. That means an enormous amount of value is being created.
Frontier Lab margins and the decline of knowledge work 29:54
Chester Roh As Seungjoon Choi mentioned earlier, these margins are coming at a painful cost to some people right now. Jobs are disappearing because of it.
Seungjoon Choi Right. What was it Microsoft announced? A voluntary retirement program recently. I understand about a quarter are taking it.
Chester Roh So, as we keep saying, human knowledge work doesn’t mean creating things or coming up with new insights every day. Much of it is simply using what you know to move a file from one folder to another, or update an Excel sheet, compile it into a PowerPoint presentation and send it somewhere else. That kind of information processing makes up most of the work. It’s like getting an official document at the district office. A person used to be paid to issue it, but now a machine does it. The same thing is happening here. So the key point here is that Anthropic
Jonghyun Park is actually making money.
Chester Roh It’s earning a profit. A frontier lab isn’t a money pit.
Jonghyun Park And we’re talking about over 5 trillion won in the first half, just looking at the cash changing hands.
How AI answers threaten Google’s and Reddit’s ad models 31:01
Chester Roh This also ties into Google’s stock falling sharply. Google’s search quality and other things improved, but its ad profit per click fell. Reddit’s stock also fell sharply the other day. The ways people used to consume traditional media are all breaking down. If you had a question, you’d search Google, then visit a related Reddit thread or Quora and ask there. The spaces where you saw ads along the way are disappearing. Now ChatGPT sums everything up and gives you the answer in one go.
Seungjoon Choi It looks at dozens of sites in 10 seconds and gives you an answer. The quality is on another level.
Chester Roh Here too, the middle will probably disappear. Companies in tiny niches will profit, blockbusters will do even better, and the uncertain middle will disappear once again. Let’s move on.
Jonghyun Park Anyway, they’re already making money, but their valuation is still enormous. So will demand and revenue keep growing this quickly? I brought in some figures from another report to explore that.
AI adoption by occupation and the shock of GDPval 32:10
Jonghyun Park It shows how widely AI is used across different occupations. Take engineering, especially coding: people started using it about a year ago, and now almost everyone does. Other fields, like law and finance, started using it this year, and adoption is rising even faster. So you could say most occupations have already adopted AI. The next factor is how much each person is spending. So you could also argue that there isn’t much room for more people to start using it.
Seungjoon Choi While researching the numbers earlier, something occurred to me, so I looked up where GDPval stands. OpenAI released it last year. It’s not fully saturated, but it’s in the 80% range.
Jonghyun Park That covers tasks across various occupations, right?
Chester Roh Then, out of 100 people, theoretically, if you think about it rationally, 80 should end up unemployed.
Seungjoon Choi If this plan succeeds, there will clearly be collateral damage. I can’t help but feel that collateral damage is inevitable.
Chester Roh Yes, it will happen. It will.
Seungjoon Choi But Chester, you kept saying the entire middle tier would collapse.
Chester Roh I have no desire to spread doomerism about things collapsing. The three of us hosting this show are in that middle tier too. We’re not Big Tech. We’re just people in the middle, talking with you here in our little corner of the world. We wouldn’t be needed. If a podcast like All-In came out fully dubbed in Korean every week, you’d watch that. Why would you watch us? So we’d take the hit.
Seungjoon Choi We’d take the hit and have to move. We need to think about leaving or changing course.
Chester Roh My message is: let’s move. Think seriously about where to go and take action. I think the direction we need to move in is clear.
Seungjoon Choi What is it?
Chester Roh Toward what’s human, where scarcity still exists.
Seungjoon Choi It’s time to think hard about that.
Chester Roh There’s a lot to do. Shall we look at this graph?
The plunging price of intelligence and faster industry change 34:23
Jonghyun Park Moving on, this graph looks at AI—at intelligence— and what it costs. This ties back to what we said about replacing labor. Intelligence keeps getting cheaper, and it replaces people because it costs less. How does that compare with other goods we’ve had? How much cheaper did they get after they appeared? Take electricity: over 80 years, it became about 10 to 30 times cheaper. Compute, over about 60 years, fell to nearly one hundred-billionth of its original cost. But AI is getting cheaper even faster. Earlier, you said change unfolds across generations. But with AI, change may come faster than a generation.
Chester Roh We’ll see the results within our generation. I mean, look how fast things have moved in just one year. How many fields have frontier labs wiped out in a single year?
Jonghyun Park You could say many businesses are already starting to be reshaped.
Chester Roh Our country has relatively stable employment, so it happens a bit later. But the pressure builds and eventually bursts.
Jonghyun Park That reminds me: you mentioned character chat. Character.AI was a leader in that space. It rose, then went all the way to bankruptcy— it’s already been there and back. That shows how quickly things are changing. It’s another example of that.
Chester Roh In a niche the frontier labs briefly overlooked, you can use an agent like Muse and set up your own persona to get a Character.AI-style service. Even character services doing well now may struggle to go mainstream. The world changes fast.
AI-native companies and the San Francisco gold rush 36:16
Jonghyun Park Let’s look at one more thing: Are AI companies actually making a lot of money? If these companies make more money, they’ll also spend more with model providers like Anthropic and OpenAI. Then demand will keep growing. So we classified them as AI-native companies. These are companies where AI itself is the product, like character chat, for example. Those companies reportedly make far more money than companies that added a little AI to existing software— the AI-enabled companies. So businesses actively using AI are creating value in the market right now. That suggests demand may grow even faster. At least, that’s one possibility.
Chester Roh It’s a gold rush right now. Everyone’s starting an AI company, and San Francisco is still buzzing, apparently. You were there until last week, after all.
Jonghyun Park That’s why I’m trying to do it too. I think it’s time for me to make a move.
Chester Roh So you’ll be moving to the US soon.
Jonghyun Park If I can. And among AI companies,
Model competition in the benchmarks and Kimi’s decline 37:22
Jonghyun Park if we focus on Anthropic and OpenAI, how smart the model is is ultimately one of a frontier lab’s competitive strengths. AAII, the metric we look at most, changed this week. So let me pull up the scores. Even under the revised scoring, Claude and GPT both remain at the very top. For context, after the metric changed, Gemini 4 Carbon appeared, and Carbon is now in first place. But Carbon isn’t public yet, so we can’t use it. It’ll probably be available soon.
But here’s why I brought this up. From Anthropic’s perspective, Chinese labs are among its biggest threats. Kimi, in particular, had been making rapid gains. Before the revamp, Kimi K3.1 Pro came out last week—about a week ago now. It made a big leap as soon as it came out. Its scores suggested it might be at this level and genuinely threatening the frontier labs. But when the benchmark changed completely, Kimi dropped down a bit. So it seems the frontier labs are still holding strong. That’s one way to look at it.
Seungjoon Choi How did the benchmark change?
Jonghyun Park I think the private test sets probably carry a bit more weight now. And benchmarks keep changing anyway. Once the questions become known, they get solved, so benchmarks move toward harder, previously unseen questions. Those changes seem to have shifted the scores quite a lot. So models whose rankings suddenly shifted, like Muse and Kimi, may have optimized for the old benchmark. That’s one way to interpret it.
Seungjoon Choi So they may have gotten good at specific benchmarks without necessarily being more intelligent overall. That’s another way to look at it.
Jonghyun Park Scores could also differ by field. And there may have been a gap between benchmark scores and how capable the models feel in practice. As an aside, with Kimi, I met someone from the company this week, so I had a chance to ask them directly. I asked them a few things, and they said they’re putting a lot of effort into coding performance. Since I’m interested in video, I asked, “Are you planning to work on video?” They said they had no plans to for the time being. That was their view. But when it comes to making video with code— like the 3D work you showed earlier— this Kimi model is also really good at Blender and similar tools.
Chester Roh As I said about models in my last session with Jonghyun Park, a model might be good at video or something else. It’s like adding features to a car. We could think of each model as a finished car model. So with a model from a particular company, we might say, “The OpenAI 7 series has automatic climate control?” I think that’s where we’re headed.
Jonghyun Park But those features are arriving so fast. If I don’t make a move quickly too, I’m reaching a point where I may not survive.
Chester Roh Either become a major player or build a strong niche brand with 2,000 customers you can count on, or go mainstream. If your role is that of a midsize city caught in between, I’d say that’s a tough position. All right, let’s move on.
Anthropic’s massive compute contracts and the overinvestment debate 40:34
Jonghyun Park Next, the compute this lab is spending on— I’d like to take a look at that. As for how much they’re spending on compute, the figures are here, so let’s take a look. For now, about $600 billion in spending is committed. Of course, that figure could keep changing. And it’s not all spending for next year. Some contracts run through 2033, and the longest ones cover nearly 10 years of compute, so the spending is spread out over time. $600 billion is nearly 1,000 trillion won. That’s the scale of spending we’re talking about.
As you mentioned last week, some are long-term lease contracts, while others are short-term contracts. With the SpaceX contract, the termination clause allows either party to end it. With 90 days’ notice, Anthropic can terminate it, or the supplier, SpaceX, can. I guess that means the price can keep changing. Anyway, their compute spending is on the order of 1,000 trillion won.
But if we think about that, based on the revenue trend we saw earlier, by year-end they’ll probably reach roughly $100 billion in annualized revenue. So assuming they keep growing, the amount currently committed doesn’t look all that large. They’d spend about 1,000 trillion won and bring in even more revenue than that. I think that’s one way to look at it.
Chester Roh Compute is the core asset, so everything comes down to compute. So compute is basically oil now. Oil.
Jonghyun Park Anyway, the compute they’ve contracted for— it’s the same with oil. If demand for what you make with that oil keeps growing, it’ll remain in short supply. On the other hand, AI is seeing enormous spending, so if demand doesn’t keep up, there’ll probably be a surplus. That would be the phase of overinvestment you mentioned last time, I suppose. At this scale, I don’t think it’s overinvestment.
Chester Roh That’s how it feels to me.
Anthropic and OpenAI valuations and the gap with Chinese Labs 42:50
Jonghyun Park Let’s think a little more about the competitors. Anthropic raised funding in May at a valuation close to $1 trillion. About 1,400 trillion won? It was recently valued at around that level, and OpenAI is similar. But Chinese labs like Kimi are valued at about one-twentieth of that? Since they’re raising funds at that scale, there seems to be a valuation gap.
Chester Roh OpenAI and Anthropic seem to be valued on the idea that they could be the next Google or Meta. Even before an IPO, Anthropic has over $120 billion in cash. For compute, as Jonghyun Park mentioned earlier, as well as talent and market expansion, customer acquisition and marketing, they seem to have plenty of room to spend.
Meanwhile, OpenAI is reportedly preparing to raise about $30 billion at a valuation of around $1.25 trillion, but I hear those meetings aren’t coming together. Things have a way of turning around, and Anthropic seems to have the upper hand right now, but I think this dynamic will change soon. Even the people putting up the capital and others like them— I think very few really understand exactly how the technology works. When I hear people in finance talk, they seem to piece together widely known news and speak about it superficially. So I think OpenAI’s current $1.25 trillion valuation, compared with Anthropic’s, is far too low.
Jonghyun Park In OpenAI’s case, I hear its spending is enormous right now. As for valuation, everyone seems to expect Anthropic to go public at around $1.5 trillion to $2 trillion. So I looked at companies of that size among the Big Tech firms already making money. Meta is worth about that much. Apple is about two or three times bigger. And how much are those companies earning? Some $10 billion to $30 billion per quarter. So if Anthropic earns as much as Meta—roughly ten times what it earns now—this valuation makes sense. I think it might even earn ten times what it does now. So is this valuation too high? Maybe it isn’t.
Anthropic and OpenAI aim to become the next Google 45:27
Chester Roh OpenAI and Anthropic are pursuing both B2B and B2C. B2B was traditionally the domain of Office and B2B SaaS companies, while B2C included content and social platforms like Meta, or search services funded by ads. They’re entering both areas. OpenAI and Anthropic are going after what Microsoft and Google do at the same time.
They aren’t simply remaining suppliers of models built on compute—in other words, suppliers of intelligence. As I said before, the value created by the model is falling, and value is moving above and below it. Anthropic and OpenAI are also building the services above the model themselves. If they complete the whole stack, from those services down to compute, they should be seen as the next Google or the next Tesla. Then shouldn’t they be worth at least as much as Google and Meta are now, or even more? It’s Microsoft and Google combined. We’ll have to see how the competition between them plays out, but I think that’s the reasoning behind the market’s current valuations.
Shrinking search ads and the shift in value 46:46
Chester Roh One thing is clear: if someone profits from those figures, a gain somewhere means a loss somewhere else. Because, as Seungjoon Choi said earlier, there are only so many customers and only so many ads they can click. So as Anthropic and OpenAI keep rising, Google’s stock has also fallen a lot, because its search ad revenue is declining. What’s been predicted is happening. People are increasingly getting information through AI and doing everything there. You could say they’re getting used to it. Perhaps by early next year, these numbers will show it even more clearly. I don’t click Google ads anymore, and I can’t remember the last time I visited Naver.
Jonghyun Park Same here. With Reddit, I think they blocked AI traffic from coming in—or was that what they blocked? I think something like that happened recently. Then traffic dropped by 90%, was it? I heard something along those lines. My friends and I jokingly call it the dead internet. But a lot of the accounts posting comments online may not be people. I think AI is already doing much of that. But AI doesn’t look at ads and click on them or consuming something—but that’s not the case now, is it? Isn’t all of that value shifting? That’s how I see it.
Would we buy Anthropic stock when it goes public? 48:06
Jonghyun Park So, to wrap this up, will you buy when they go public? I’m asking everyone watching the same question.
Chester Roh Yes, I’ll buy. As for Anthropic and OpenAI stock, for the record, I don’t own shares in either company yet. I’ve been buying shares in other AI companies when I can. When those two go public, well, you know how the market is these days. After an IPO, the price dips a little below the offering price, then climbs back above it. It goes back and forth like that. But with Anthropic and OpenAI, as I said in my last session, I’d buy as insurance. Even if my business fails—and if it fails, that means those companies are doing really well— I’d have their shares as a hedge.
And the center of services is shifting. What’s clear is that it’s moving away from the familiar Google, Meta, and Naver toward AI model services. We’re already seeing that trend play out. So I think these companies could grow further— at least two to threefold. I think they have room to do that. Those are my two reasons for buying. Conversely, companies on the other side of that shift might be worth shorting.
Jonghyun Park Do you mean their competitors?
Chester Roh That’s right. The familiar gatekeeper companies in Korea are at risk, in my view.
Jonghyun Park I think I’d buy too. Ultimately, as Seungjoon Choi worried earlier, they could become centralized. If that happens, I think they could wield more than ten times as much influence. Would that be a happy future? Maybe not, but still.
Chester Roh Everyone worries about their job and their company, but when it comes to their personal finances, they’ll all buy Anthropic stock. Everyone bought SK Hynix stock, too. I don’t know anyone around me who didn’t. And it’s still very much in motion. So let’s leave our discussion of Anthropic’s S-1 filing there. My timeline is full of talk about the S-1 filing right now, so let’s wrap up this topic. Seungjoon Choi, what would you like to say before we finish?
Play and joy after AI surpasses us 50:24
Seungjoon Choi Okay, I’ll go through this quickly. I gave several presentations last week and this week. This one was at an early childhood curriculum society. Coexisting with AI was a major topic there, too, even in early childhood education. I took part as a discussant, and in one of the slides I made, as Chester Roh mentioned earlier, I brought up sports and pure play. Lately I’ve been asking myself, “What will remain?” What do I mean by that? As math and Go have already shown, when even the people at the forefront of a field are surpassed, what still remains? What would still be worth doing myself? That’s what I’ve been thinking about.
It made me wonder: What about math? Other disciplines? What about art? I thought it best not to put the probability at zero. That seemed like a good hedge. When it comes to making software and writing code, AI surpassed me long ago. But I don’t mean implementation. If it becomes better than me at coming up with the artistic ideas themselves, what will I still have? I find myself imagining things like that. But even if AI surpasses me there, wouldn’t the fun of it still be mine? Or something children growing up would still have? That’s the nuance I was going for.
It’s about looking for optimism amid uncertainty— optimism, but not the kind Big Tech talks about. Everything is so uncertain right now. But in early childhood education, Loris Malaguzzi once said something about embracing uncertainty as an important attitude. I find comfort in that. I don’t know what will change or how, but I’ve been wondering how to find my own sense of optimism within it.
So, the math world getting hit by this TNT— I tried to visualize it. Go had Move 37, and math had its own moment. Yet people still play Go. Mathematicians may head in new directions, but I think they’ll still do math. The joy and curiosity it brings belong to each of us, so they’re hard to take away.
Staying curious and trying things yourself— the joy of learning through experience. Children ultimately go through the same process humanity went through: exploring, trying and failing, experiencing. That’s hard to take away. Because the joy of discovering even little things and understanding math for myself will always be there. Even if I’m not at the frontier. That’s always been true. There have always been people more talented than me. Those things will continue, so I’d like us to focus on them. Still, so much uncertainty can shake me.
Uneven Future and your own AlphaGo moment 53:29
Seungjoon Choi But as we move through this and face friction, I’ve been thinking about what we can do together. I’ve worked with artist friends of mine a few times under the theme “Uneven Future.” It means more than a bumpy, uneven future. When I rephrased and translated it, I found myself facing an uneven future I hadn’t chosen, and feeling shaken these days. So I was capturing this moment in 2026.
As for imagination, what I presented yesterday went a step further. I talked about how, after 30 years of coding, I recently stopped coding. For programmers, SICP even talks about spells. Programming was tied to imagination— words working like magic.
From that perspective, I thought about what I said earlier about imagination. My artistic ideas—it’s time to imagine my own AlphaGo moment. That’s the message I wanted to share yesterday. Just as it hit the math world, unless it hits me, it’s hard to really feel it. It’s only a simulation, but in anything involving intelligence, there’s a nonzero chance of being overwhelmed. And it could go beyond implementing things— to ideas and insights we think of as uniquely human. It could far surpass me in those things, and not just me— even in my field, or Chester Roh’s business, for example. I saw you’d recorded something about that last time. The risks facing professions ending in “-sa.” Doctors (uisa), lawyers (byeonhosa).
Chester Roh That’s right.
Seungjoon Choi Doesn’t “sajang” (CEO) have “sa” in it too? Though each “sa” uses a different Chinese character. These days, I think we need to imagine our own AlphaGo moment. In my own career or business, I expect plenty of ups and downs. But we still have to live, so I’m looking for ways to stay mentally whole and live a full life through it all— to find those ways in a healthy way. Because certain things are bound to happen, I think we need to imagine what that mindset would be like. That’s what I’ve been thinking. 2026 felt like a year that called for it.
Chester Roh Exactly.
Beyond the AI frontier to how humans will live 55:49
Chester Roh We originally started this podcast to follow the AI frontier and study the cutting edge, but lately, our episodes are practically a humanities show. What should we do as humans? Everyone is asking questions like that now, and developments at the frontier have clearly gone beyond what we can simply analyze and predict. The future we imagined early this year— wondering, “What if this happens? What if that happens?”— had already arrived by the middle of the year. Our talks about AGI, ASI, what counts as AGI— we don’t have those conversations anymore.
Seungjoon Choi And earlier, you said there were very few users, right? Put another way, that number can only grow.
Chess and the coexistence of humans and AI 56:33
Jonghyun Park You also mentioned Go earlier, and I hear chess is hugely popular these days. Apparently, it’s incredibly popular. And when I think about chess—I play on Chess.com, the best-known site— the AI there helps you a lot. If you’re nervous at first about playing against another person, you can practice against AI at any skill level. Then, whether you play a person or AI, it shows you which moves were poor and what you could have done better. It gives you feedback with clear explanations. I think that helps people enjoy chess more. When you think about it, could this be one of the best examples of humans and AI coexisting?
Chester Roh Right.
Seungjoon Choi Chester, do you code these days?
Chester Roh I don’t code. Well, I do a lot of coding work, but I don’t write code myself.
Seungjoon Choi You do that kind of work, but you don’t write code by hand, right?
Chester Roh I don’t even open the code.
Seungjoon Choi I used to love coding. But surprisingly, some people don’t miss it. I’m one of them. Others miss it a lot. Some people realize that it wasn’t the code they loved, but the work of imagining something and building it.
Chester Roh I don’t miss carefully writing code to avoid pointer errors and all that. I worked hard to learn coding in college, and it was thrilling to become fluent with pointers and good at managing memory. But I don’t miss it. I don’t want to be someone who’s good at it, either.
Seungjoon Choi What worries me is that I can feel the skills I don’t use getting weaker.
Chester Roh That’s why I’ve started reading long books again. It helps me sustain a train of thought. Even when I read something I find on Twitter, I don’t read more than two or three pages. I feel like my thinking muscles are steadily wasting away.
An age of abundance where taste and relationships become capital 58:33
Chester Roh But isn’t that the kind of era we’re heading into? If we enter an age of abundance and everyone has plenty of money, remember what we said when we talked with CEO Seojun Kim? In the United Arab Emirates, where people don’t have to worry about making a living, they judge someone by how refined their hobbies are. I think our idea of capital will change in a similar way. Instead of money as we know it, it’ll be things like, “What kind of taste do they have? What’s their brand? How many subscribers do they have? How many people show up when they host a gathering?” Young people compete over Instagram follower counts and numbers like that, and everyone wants to become a YouTuber. Maybe that’s the right choice.
Jonghyun Park I think so too. In the end, the desire for human connection won’t disappear. Even if it plays a smaller role. And because it becomes even rarer, that fits the idea of moving toward what’s scarce.
Chester Roh There’s definitely scarcity. Handsome and beautiful people still do well. Because they’re rare.
Seungjoon Choi If we’re still doing this podcast a year from now, in January 2027, we’ll have to see how our thinking has changed.
Chester Roh We need to act, beyond just analyzing all this. Take a different path. Seungjoon and I could even open a café where smartphones and AI aren’t allowed. A place where people come just to talk to each other.
The imagination to recognize the next Google 60:05
Seungjoon Choi And as Chester said before, the entrepreneurs or artists of this new era have probably already been born.
Chester Roh And they could be right next to us, perhaps in San Francisco, or, conversely, in Seoul. We should watch them closely and see where value is flowing. Right now, we’re only watching two frontier labs, but as Jonghyun said earlier, AI-native companies are using them more and building them into services. Perhaps OpenAI and Anthropic could turn out to be Yahoo and AltaVista. The next Google and Amazon might already have been founded somewhere. So we need to make sound assumptions about that and find assets that aren’t overvalued and haven’t found their true worth, then invest wisely in those undervalued assets— I think that’s what we should do now.
Seungjoon Choi That might be a form of imagination, too. The ability to imagine what might happen.
Chester Roh Of course. It’s a very human thing. When I first met Jonghyun, I thought, “He seems promising.” And with certain other people, you can see futures you want to be part of and invest in. Being seen as one of those people, and figuring out how to get there— I think that matters, too. Alright, let’s wrap up today’s episode here.
Seungjoon Choi Thanks for your hard work.
Chester Roh Thank you.