EP 120
Where Does Value Go? The AI Market Correction, CoreWeave's Crisis and the Operation Loop
AI Market Correction Amid Semiconductor Plunge and Excessive Pessimism 0:00
Chester Roh We’re recording on October 24, 2026, a Saturday morning. The market has been extremely volatile over the past two weeks. New models arriving at just the right time has become something we take for granted. And I think the swings in the stock market were probably the biggest shock for people. Major semiconductor companies expected to make huge profits next year and the year after, especially Korean memory chip companies, saw their share prices swing wildly and fall by more than 40% from their peaks at one point. On top of that, people are asking whether this is a bubble, while a16z and others are bringing out all sorts of statistics and flooding the market with different signals. As for how to interpret all this, opinions seem to be divided. Personally, I think the reaction is a bit excessive. The build-out phase isn’t over yet, but there’s a sense that too much is being invested, that demand is inflated and double-counted, and that real demand isn’t keeping up. That seems to be fueling a prevailing sense of anxiety.
Chester Roh Then again, people in finance must be looking at all those numbers, but everyone around us is watching stock prices too. They’ve got Hynix, Samsung Electronics, and US stocks on their screens. So the capital that could flow in is already in the market, and with no more buyers waiting in the wings, the market is running a little low on energy. Add pessimism at a time like this, and it’s an ideal opportunity for short sellers. Sure enough, if you look at companies at the center of it all, like CoreWeave and Nebius, their outstanding short positions are enormous.
The Great AI Transformation Where Every Layer Changes Simultaneously 2:00
Chester Roh I think everyone is asking the same question now. I felt this myself during our last recording, too: the world is changing so quickly that trying to keep up day by day and reacting to every turn is exhausting. It’s time to step back and ask where on earth the world is headed. This feels like a time to explore those questions. Even looking at history, although we talked last time about the railroad and telecom booms, and then the internet and mobile, one difference is that this is happening at a speed those can’t compare with. The second is that it isn’t just one layer at work. It’s not just one kind of infrastructure, such as core computing or core communications infrastructure. I don’t know if this analogy works, but from assembler to Python, as things move constantly up and down the stack, every layer is being transformed. In the past, things would build up step by step, and perhaps a new layer would be added on top of the applications we know. You’d expect one change like that, but now everything in between is changing at once.
Chester Roh With all of that happening together, interpretations are all over the place. People in finance, people in AI, and people unfamiliar with the subject are all taking part in this market, sharing their opinions on YouTube or Twitter and exchanging views with one another. So this is a time when different perspectives are clashing, and I don’t think I can say with certainty exactly what’s going on. But my sense is that the build-out phase isn’t over yet, and demand may continue for a while longer, perhaps much longer. For one thing, I think pessimism arrived too early. Second, in terms of the financial cycle, there are no new buyers left, so short selling dominates and the market goes through a correction. Even so, I personally think there may still be a lot of upside left.
The Age of Exploration Analogy and Demand After Video Modality 4:16
Jonghyun Park Yes, I think much the same. I don’t know much about finance, but ultimately it comes down to whether demand really grows and people keep using it extensively. Will people continue to use AI more and more in the years ahead? That’s what it will come down to, however much the market fluctuates. From what I hear, GPU companies and similar businesses still have a great deal of demand.
Chester Roh Yes, exactly. But hasn’t almost everyone who was going to come in already come in? But even if we start with that conclusion, there’s still so much to do. You can’t just make one thing and expect the revenue to follow. It’s an age of exploration. You board a ship and endure a dangerous voyage, and if you make it through, you may discover a new continent and claim it all. For example, you might develop a new drug and patent it, develop a new material, or even solve aging. When you see math problems and the like being solved, you realize people have become numb to the scale of it. It’s unbelievable, really. We’re talking about the singularity, really: things that would take all of humanity working together for an enormous amount of time are happening all at once.
Chester Roh I don’t think we’ve adjusted yet to that exponential scale. So whenever we discover another new continent, the potential payoff is so large that there’s still plenty of reason to pour in venture capital, and there will be for a long time yet. Of course, from a traditional B2B or B2C software application perspective, you could say demand isn’t keeping up and the value still hasn’t been proven. When I take this a step further and ask Jeongkyu Shin what he thinks of my pessimistic view, he sets everything else aside and simply says, “Video hasn’t arrived yet.”
Chester Roh Right now, text goes in and text comes out, or images go in and a few PowerPoint slides or images come out. That’s the world we’re in. But imagine video going in and video coming out. That would be enormous. In the early days of the internet, people said we didn’t need more bandwidth. But with that extra bandwidth, Netflix and YouTube came along, and now more than 90% of all YouTube traffic is video. The point is, models could very well follow the same path. Our imagination is limited; I think there’s still much more to come.
Chester Roh In the episode from two weeks ago, I asked how on earth we should interpret this, and we talked about it for some 40 to 50 minutes. I said we needed to look at the domain gap and time gap separately, and we discussed things like that. Today, we’re going deeper into that discussion. I needed to understand for myself how the numbers work and how these companies are structured, so I had to research those things. I used Claude and ChatGPT, ran countless bots, researched various questions, pulled out the material I needed, and even wrote an essay with Fable. That’s what I’ve been doing. Let’s talk about it.
Model Prices That Fail to Rise Through the Lens of Domain Gap and Time Gap 7:21
Chester Roh As recently as late last year and early this year, the thinking was that model prices would rise sharply as models improved. There would be $3,000 plans and $10,000 plans. Why? Because even at $10,000, you could get more than your money’s worth. Frontier models would be that capable. That was the thesis. But it hasn’t worked out that way. GPU prices keep rising, while prices per token keep falling. Kimi K3.1 came out this morning, and DeepSeek V4.1 Flash came out this week. The rumor on my timeline is that however much you use these models, your token balance barely goes down. That’s how efficient they are, and the token cost per unit of capability keeps falling.
Margin Structures Through the Lens of Value Chains and Scarcity 8:47
Chester Roh In Jensen’s diagram, energy is at the bottom, compute is above it, AI factories are above that, and models and applications are at the top. Even Jensen, when drawing the diagram, makes the application layer the thickest. And in the compute layer, he draws only Nvidia. History teaches us this again and again: when we buy a car, cosmetics, or any other product, it goes from raw materials to the consumer through many links in a supply chain.
Chester Roh Right? Each business takes a margin along the way, and its margin depends on how scarce it is in that value chain. If there are many suppliers, competition arises and prices fall. As competition intensifies, we call the product a commodity. The quality of being scarce is called scarcity. That’s what scarcity means, and scarce resources always make a lot of money. In the pharmaceutical business, for example, the drug company upstream makes the most money, while distributors downstream make less than the drug company.
Chester Roh But with cosmetics, stationery, or power strips we buy on Coupang, so many people make these products that the businesses putting them in front of customers, the ones that command consumers’ attention, such as Coupang and search engines, make the most money. Influencers make a lot, too. That’s how it works. So if you ask where value goes in a value chain, the answer is always the same: it goes where scarcity is. Value capture goes where scarcity is.
Chester Roh In our terminology, that scarcity is IP, a moat, competitiveness, or a strategic advantage. There are many ways to describe it, but it must not leak away. That’s what it means to have a moat. Having no leaks and having a moat are the same thing. That’s why I said value goes to scarcity that doesn’t leak away. That gives us a lens through which to view this industry. Start with electricity at the very bottom. What’s interesting now is that the discussion starts with electricity and cooling—power and cooling— and goes all the way up to applications. Things are happening at every single layer.
Chester Roh So when we look through this lens, we can see which areas will keep making money, and for how long, and which areas will become commodities. That’s what this lens helps us see. When we look at models through that lens, why can’t frontier labs raise prices?
Near Frontier Labs Catching Up and Frontier Labs Unable to Raise Prices 11:33
Jonghyun Park Because Chinese models have caught up so well, frontier models are no longer scarce.
Chester Roh Even in the US, Muse is catching up, Grok is catching up, and there’s Google—Google, the perennial powerhouse. That’s already four or five players in the US alone. Microsoft and Amazon could also decide to enter that market at any time. So there are already five or six players there. China has more than five or six, too— it has dozens of frontier labs. I saw Kimi K3.1’s benchmark results this morning, and its Terminal-Bench score was close to Opus 5.5’s. So we call OpenAI and Anthropic frontier labs, and the ones behind them near-frontier labs. Some call them Quasi Frontier Labs, and others call them Near Frontier Labs. But the gap between Near Frontier Labs and Frontier Labs has narrowed sharply.
Jonghyun Park I’ve been noticing something else lately, too. Since I follow video closely, I’ve seen Canva start releasing its own models. HeyGen has also started releasing models that rank near the top. And DeepMind’s team has started releasing models like Elorian that generate visuals well. So the companies we already know aren’t the only ones catching up. New model companies are emerging too; people often call them NeoLabs. They’re all entering the field, and models are becoming less and less scarce. That’s the trend I’m seeing.
Vertical Domain Strategies and Leakage-Proof Scarcity 13:11
Chester Roh Because they feel their own services are facing an existential threat. As Frontier Labs buy up data, they gain access to that layer, that domain. If they simply exploit that domain, those companies could lose their businesses. I see this as an effort to protect their own verticals from Frontier Labs. Some are doing that on a smaller scale, within a single domain, while others are doing it across the board. I’ll make today’s material public, so please take it, add it to your models’ context, and use it as source material. There’s quite a lot here, including many numbers, so I won’t explain every point on every slide. I’ll keep the story moving toward the conclusion I want to reach. Ultimately, where does scarcity that doesn’t leak away go? Where does scarcity that doesn’t leak away end up? Broadly speaking, one destination is Frontier Labs. The second question is: what can’t Frontier Labs capture? That includes some of the vertical domains Jonghyun Park mentioned that NeoLabs are pursuing, but also services whose operations I own and that run within my business— services from which a Frontier Lab could obtain any signal only at tremendous cost. I think that’s where we’ll land, so I wanted to say it up front.
Overinvestment in Infrastructure and the Application Era Opening After the Correction 14:49
Chester Roh Andreessen Horowitz has been putting out a lot of data points lately. On a16z’s YouTube channel, they present data-heavy slides or invite big-name guests to share all kinds of figures. Their recent theme is this: They don’t state a firm conclusion, but they say every cycle has had an infrastructure buildout period, and once that buildout is largely complete, a major market always emerges on top of it. Here, they’re talking about roughly 2010 through 2016— the era of mobile, cloud, and the mobile internet. Right? First came Qualcomm and Arm, followed by device makers like Samsung and Apple. Then countless software companies built on top of that, including B2B SaaS companies and Airbnb, and all those mobile internet winners we always talk about. Software and services flourished. That’s the story they’re telling.
Chester Roh I think something similar will happen with AI, but we still don’t have a clear sense of what form it will take. So they’re exploring things like Muse and Grokbot on the application side to see what works. If you simplify the pattern, it always looks like this: Things rise during the infrastructure buildout, as inflated expectations of the future drive infrastructure investment. As I said in the last episode, demand gets overstated. For example, between New York and Chicago, one or two rail lines would actually be enough. But people say, “Think of all the traffic between this industrial region and New York.” So they overinvest and build three lines, which end up competing with one another.
Chester Roh Then people realize they’ve overinvested, and there’s a correction. That correction doesn’t mean a total collapse. The overbuilt assets may change hands. The original investors might go under, and a new generation of investors buys the assets. Those new owners have a different cost basis, so they can operate at a much lower cost. A new structure rises over the ruins of those who failed. On top of that infrastructure buildout, the era of applications and services blossoms and lasts for quite a while, until a new paradigm arrives.
Jonghyun Park And in the AI era, we’re probably in that infrastructure buildout phase now— that’s how many people see it, right?
Chester Roh They’d say we’re about here now. We’re near the very end of the infrastructure buildout. If you look at stock prices, they’re falling sharply. And a great many people who invested in it are losing money now, aren’t they? So this is probably where we are. Does that mean all of it will end in losses? But when there’s overinvestment, investors losing money is practically a given. Investors and creditors who funded the industry can be expected to take some losses. Value-added businesses that build applications on top sometimes capture an even larger share. Or those benefits may simply spread throughout society and benefit everyone.
Anthropic’s Revenue Surge and the Time Lag in Knowledge Worker Adoption 18:31
Chester Roh But I think this caught the market a little off guard. Anthropic had more revenue in the first quarter of 2026 than it did in all of 2025. Its Q1 2026 revenue exceeded its full-year 2025 revenue. Claude Code was, of course, at the center of that. Looking back from late last year to this year, it’s fair to say every coder and near-coder worldwide has come on board by now— probably just about all of them. Among the people around me who code, nobody even opens Visual Studio Code anymore. If you have it open now, people treat you like a fool. So all those people have come on board.
Chester Roh But Andreessen Horowitz says there are about 30 million coders worldwide. All 30 million have come on board. And the revenue those 30 million people generated is Anthropic’s current revenue. Was Anthropic the only one that grew? OpenAI and the rest of the coding market grew tremendously too. Devin grew tremendously as well. But the next wave of people isn’t coming in so readily.
Jonghyun Park Right. People who aren’t developers need to come in too.
Chester Roh The number of knowledge workers who aren’t coders, according to Andreessen Horowitz, is about 1.5B, or 1.5 billion people. Thirty million versus 1.5 billion—that’s a huge difference. Now that 30 million have come in, if this is the result, imagine when 1.5 billion start coming in, The assumption was about how large this market’s TAM would get, but now it’s, wait— they’re not coming in as fast as expected. There seems to be a bit of a lag. That’s the situation the market is now viewing through a biased lens. So if they’d kept coming in, Q3 revenue would have risen above Q2, and growth should have held steady. Instead, it slowed. The Q4 revenue growth rate will probably tell us a lot.
Slowing Growth Rates and Expectations Priced into Anthropic’s Valuation 20:10
Chester Roh Perhaps that’s why Anthropic is pressing ahead with an IPO in Q4. We have our own barometers around us, too. Coders have all come in, but companies outside coding haven’t even formed a view yet on how best to use this. So I think there may be a slight lag before they truly enter this investment cycle. That’s why I think we may go through a somewhat flat stretch along the way.
Jonghyun Park But looking at these numbers, growth from Q2 to Q3 was 45%. That’s still enormous. Seeing expectations fall over that makes me realize just how high they were.
Chester Roh Yes, expectations are that high, which is why Anthropic’s valuation is put at 2T and things like that. The company has only existed for five years. Yet compare its market capitalization with Apple’s and Google’s. A startup that’s only five years old can be valued that highly. In terms of revenue, it’s frankly tiny compared with Meta or Google. For the market to accept that valuation means it expects this sector to grow much larger. If that forecast falls apart, that valuation becomes hard to justify. So we’ll have to watch and see. No one knows whether those bold enough to bet or those managing risk will come out ahead. Everyone can just keep quiet and place their own bets.
CoreWeave’s Data Center Financing Structure 22:02
Chester Roh Another thing that happened in the meantime involves CoreWeave, which is central to this story. Its corporate structure is complicated. I once heard CoreWeave’s CEO explain it somewhere: they create a box, put a 15-year data center contract into it, then buy equipment and put that in, and add a customer contract. That lasts three years. Then, to support all that, they raise debt. They issue five-year bonds and keep it all within that one box. Hyperscalers usually buy those boxes. Anthropic, OpenAI, and a third company reportedly account for nearly 70% of the total contracted volume. They buy the boxes, and some bonds even have guarantees from the buyers. A company like Anthropic guarantees it will buy. That reduces the box’s risk and lowers the borrowing rate. After they’ve lowered the rate, they deduct the expenses and financing costs, and what remains is CoreWeave’s profit. The business is structured so those profits and risks accumulate at the parent company, CoreWeave itself. So the CoreWeave parent company doesn’t operate the entire IDC on its own. Below it are separate entities, like subsidiaries, that hold those contracts as distinct businesses. That’s how it’s structured, but the figure we’re looking at is for the CoreWeave parent company.
Chester Roh When we talk about its bonds, we mean bonds that aren’t backed by revenue— unsecured, subordinated bonds. It initially issued these at 9.75%. The five-year bond maturing in 2031 was probably issued early this year. Treasury yields were four-point-something percent then, and they offered a spread—an extra 5% in interest— when they issued it. That’s a very high interest rate. But last week, and then this week, it was nearly 15.78%? At times it was even higher. Then there’s CDS, or credit default swap. That bond could default, right? So someone who bought it, knowing it could default, also buys something called CDS. It’s a kind of insurance. If you pay the premium, even if the bond defaults, the CDS seller covers the loss. But these two numbers make no sense. A yield to maturity of 15.78% means you can buy a 100-won bond for about 80 won right now. So you buy it for 80 won, with a face value of 100 won, and collect 9.75% interest going forward. Since you bought it at a 20-won discount, you also recover that difference when the principal is repaid. That’s how the yield works out to 15.78%.
Risks Revealed by Bond Yields and CDS Spreads 24:01
Chester Roh That part is fine, but to offset the risk, if you try to buy CDS, the premium is 1,100 basis points, or 11%. That’s 11% of 100 won each year. You have to pay 11 won a year in premiums for someone to say, “I’ll guarantee this bond for you.” If you combine that CDS with the bond, you end up holding a risk-free product. The soaring CDS premium means the market sees this business as enormously risky.
Jonghyun Park To put it as simply as I understand it, CoreWeave borrows money to build GPU facilities and uses their earnings to run its business. If it can’t sell that GPU capacity, CoreWeave goes under. People now think that might happen, so the yield rises and the insurance premium shoots up. That’s how I understand it.
Infrastructure Businesses Shaken by Financial Conditions, Not Demand 25:52
Chester Roh Yes, that’s right. But the key point is everyone still sees demand as solid. The problem is the financing needed to meet that demand. Interest rates fell a bit this week, but last week they reached the 5% range. Nobody knows exactly why that happens. It depends on market sentiment at the time: sometimes rates rise, and sometimes they fall. But overall, interest rates are rising, and many infrastructure investment projects use products that create this kind of arbitrage. Financial techniques are applied to projects that must go ahead, and investors get involved in large infrastructure projects this way all the time. What derails those projects is usually the financing environment, rather than demand. Interest rates may suddenly rise, putting the project in the red. Or they may need to issue a new bond to cover the old one, but can’t. That’s usually when a default happens.
Chester Roh And when a default happens—this is the key point— the industry doesn’t stop. It was the same during the railroad boom. Ownership simply changes through restructuring. With CoreWeave, for example, many people say that if the 10-year Treasury yield exceeds 6%, CoreWeave’s business will collapse. That’s because its business was designed when rates were low, at around 4%.
Chester Roh But I’m not crunching those numbers in Excel to determine definitively whether CoreWeave will fail. When the time comes, SemiAnalysis or another research firm will publish an analysis at the speed of light, explaining what’s going on. Financial firms will probably publish the most. Financial firms take analyses like these, circulate them among themselves, and shape market sentiment. The tail ends up wagging the dog quite often. I think this time, too, less as a bubble bursting and more as a sneeze or a hiccup. I think prices will probably recover.
Jonghyun Park Someone must be waiting for a chance to buy it.
Chester Roh They’d have to go under first. But they haven’t gone under yet. It’ll take some time for them all to fail. Though nobody knows if that will happen. As Seungjoon Choi mentioned, Terence Tao looks a bit worn out. Near Frontier and Quasi Frontier Labs like DeepSeek and Kimi are racing to catch up. Among the smaller bits of news, the Tesla Roadster flew. So demand is real and interest rates have risen, but the weakest link’s CDS spread has soared. That makes us wonder why. These are this week’s figures, and things got worse than last week. On the positive side, Treasury yields fell a bit. CoreWeave’s spread peaked at 1,195 bp, and it’s still at 1,115 bp, which is extremely high. Its share price is probably near a 52-week low. It fell to one of its lowest levels. What’s interesting, though, is that demand to short it was the highest. So Nebius and CoreWeave now have enormous short positions against them. About 23.7% of the shares available to trade— nearly 30%—are in short positions. It’s a falling knife with no buyers to catch it, so the price has kept plunging.
CoreWeave’s Decline Driven by Short Selling and Market Sentiment 28:25
Chester Roh Short sellers’ unrealized gains are reportedly about $6 billion. So the fundamentals may not have actually worsened. People on Wall Street have different views. Citadel’s Ken Griffin says unequivocally that it’s a bubble, while others say it isn’t. Goldman Sachs, for example, says it’s aggressively long. Views are all over the place right now. Someone will be right. It’ll either go up or down. So I’ll move on from that.
Chester Roh Some numbers from Anthropic’s S-1 filing also came out. There was also talk that Kimi K3.1’s numbers were very good. So I ran a thought experiment, and I want to say a bit more about the conclusion I mentioned earlier: going where you own the operation and have an operation loop. I think that’s the only way forward. Let me explain. Where should we invest, and what kind of business should we build? That’s always the most important question. Honestly, over the past month, after returning from a business trip to Europe, I stepped back from the trees to look at the forest and spent time thinking about where we should go. I’ve ended up sounding like a Zen master. I’ve been thinking that, rather than the telecom bubble, the cycle of the railroad boom may be repeating itself in a similar way. That’s been on my mind.
Businesses That Own the Operation Loop and the Time Gap Between Labs 30:30
Chester Roh There’s a time gap and a domain gap. The time gap keeps shrinking. The gap between Frontier Labs and Near Frontier Labs is definitely narrowing. But for those of us who aren’t labs, the time gap for selling Frontier Lab capabilities to people further behind who don’t know about them yet is still quite large, I think. As for the domain gap, models can do everything these days. If Opus 5.5 is good at video editing, Kimi is good at video editing too, and so are other models. When I think about why, I imagine someone sells an RL dataset for training that skill to Anthropic first, then takes it and sells it to another provider. When a new model comes out, they say, “This model is really good at this.” It’s like announcing a new car: “This one has heated seats. This one has cruise control. This one has automatic climate control.” They add features one by one and make a big deal of each one. Similarly, domains seem to be built into models as features, and those features become marketing material. So, as you know, it starts with areas where reward signals are strong in digital settings, like coding and math, then moves toward areas where reward signals are harder to get digitally.
Jonghyun Park Yes, I feel exactly the same way. I think frontier labs are conquering domains one by one, in any case. It would be great if we could see what’s next. We should watch what data sellers are collecting now and who they’re selling it to. That gives us some sense of which domain might be next.
Where Value Moves When Models Become Commodities 33:12
Chester Roh Yes, exactly. We hear a lot about that. We have a rough idea of what data packages sellers are bringing to the so-called near-frontier labs. Right now, scarcity seems to be concentrated in about two places. One is the physical bottleneck. When model prices become commoditized, the value captured there moves somewhere else. If it moves down, infrastructure providers capture it. But infrastructure providers usually make, at most, maybe a 10% margin? Of course, AWS has a 35% margin, but they generally maintain low margins. Or the value moves up. Most often, businesses higher up capture more value.
Chester Roh That’s because, as we know, the closer you get to the paying customer, the more value you capture. But what if everything higher up becomes commoditized too? Then no one can capture that value. The whole industry becomes commoditized, and consumers get all the benefits through lower prices. And that’s where it ends.
Jonghyun Park Life gets easier for everyone in the world.
Chester Roh Right, right. And many industries are like that. Railroads were, and electricity is too. Prices keep coming down, don’t they? Thanks to Coupang and Amazon, more often we can buy things that used to cost us more at the lowest price. I should briefly talk about frontier labs. Anthropic and OpenAI are frontier labs, and in today’s market, when those two companies act, the world reacts. So what they do really matters. When they do something, a market forms around it. It’s fascinating. In any case, scaling laws still hold, and we see the performance gains the labs are achieving and say, “Wow, it’s the singularity. Look, math is solved. Look, other fields are solved too.” But I don’t think that’s RSI yet.
Scaling Laws and the Ongoing Compute Race 35:07
Chester Roh For it to be RSI, a model would have to develop models on its own and produce gains beyond what comes from adding more compute. But I don’t think we’re there yet. It still seems to be a combination of compute and the model efficiency gains DeepSeek and Kimi excel at: algorithms, KV caches, and changes like those. I think the gains come from combining those things. Most of the impressive performance we’ve seen comes from putting in more compute.
Chester Roh By today’s standards, a quarter of pre-training at around 1 GW yields a new generation, followed by about as much post-training. That’s how the model comes out. In post-training, it enters the Hyperbolic Time Chamber and comes out having solved every problem. I think that’s how we get to this point. Loss keeps falling, but even with 100 times more compute, the reducible loss falls by only about half. When the Chinchilla paper came out, there was a claim that, however far loss fell, it couldn’t go below 1.68. We’re probably somewhere above that now. The portion of loss you can reduce by adding 10 times more compute is called reducible loss. In theory, you can cut about 30% of it each time. Labs don’t publish those loss figures, so we don’t know, but they’re probably still falling. And the relationship between falling loss and performance gains? We don’t know that either. But to improve the performance we actually experience faster, you need a lot of compute. Doubling compute doesn’t feel like it doubles performance. It feels more like you need 10 times the compute to double performance. So while this competition continues, while the race for performance goes on, the compute required will keep growing.
Jonghyun Park I feel much the same way. Just because loss goes down doesn’t mean the tasks an ordinary person can handle suddenly double in number or anything like that. And with model release cycles, having followed them for years, they do seem to be getting a little faster. But it doesn’t feel like they’ve suddenly shot through the roof or that models are changing dramatically with every release. It feels like the same trend is continuing.
Chester Roh But I should say this upfront: compute is never going to be abundant. It will remain scarce. Because until intelligence becomes superintelligence and makes us look like dogs, this competition probably won’t end. If you view this through the lens of past build-outs, such as the internet or mobile, or even railroads, the slope won’t match what we’re seeing now. The timeframe and the slope are completely different now. I’ve said countless times that we need a new lens. Let me get back to the point. I judge frontier labs by three things: compute, data, and talent. But data and talent tend to spread. Talent spreads when people change jobs. In San Francisco, as Jonghyun Park knows from visiting, it also spreads when friends meet for a meal. People tell each other everything. They don’t share exact details, but they reveal the general direction and which mountain to climb to get results. That’s how talent spreads. With data, a vendor sells it to one company and then keeps selling it to others. It spreads everywhere. You’ve seen how that company improved after buying this, right? You should buy it too. That’s how they keep selling it. So even among those labs, what creates a lead isn’t really who has smarter researchers or one genius on the team. Ilya Sutskever made the wrong bet too, in my view. It’s no longer about talent and philosophy; it’s become entirely a compute competition. They have to maintain an advantage in compute for everything they’re building now to make sense. So Anthropic and OpenAI and these frontier labs can’t stop this endless compute race until one of them is completely wiped out. That’s what I think, and they’ll keep going. The market may swing, but people at those labs come to buy whenever a new data center comes online. They come to buy, regardless of price.
Scarce Compute and Leaking Data and Talent 38:30
Chester Roh So what does 1 GW amount to? These days, people don’t talk about GPU counts. Just as meat is sold by weight, capacity is sold by power, isn’t it? For Blackwell GPUs, roughly 400,000 to 500,000 of them. In dollar terms, about $50 billion to $80 billion. Compared with two or three years ago, memory prices are up tenfold, and chip prices have risen too. As this build-out race intensifies, the people selling jeans and pickaxes are making enormous amounts of money. If demand for their products vanished, they’d all go under, but demand may work differently this time.
Chester Roh What’s different from the past, I think, is that demand can keep emerging in new places. With telecom, someone actually had to use the bandwidth. Jonghyun mentioned this earlier too. Everyone in San Francisco is building their own model. Right? They have an incentive to protect themselves and create new businesses that way, even pulling capital forward to put compute to work to find and defend those opportunities. Every company has that incentive now. Every startup does too. If I had to make a comparison, we’re like Europe before the discovery of India and the Americas, just as seafaring had been invented.
From AI That Replaces Demand to AI That Creates Demand 41:20
Jonghyun Park I think the Age of Exploration analogy fits too, because so far, these companies have taken work people do and used AI to replace it. But once all the work people do has been made more efficient, demand runs out. I think we’re at that point, but take coding, for example. As people become able to do much more coding, it doesn’t just end with AI replacing developers. The total amount of code produced soars. The internet is overflowing with slop, and everyone seems to be targeting the layer above it. It feels like discovering a new continent, and if this happens in other fields too, demand won’t stop at replacing what’s already there. It could multiply and break through the ceiling. That’s what I imagine.
Chester Roh Everyone’s getting a ship ready and setting off in every direction. Some have the bad luck to reach the North Pole, some reach a land of gold, and some die along the way. That sort of thing is happening all the time now, and when someone discovers a new land, the rewards are enormous. It may sound silly coming from me, but I said this means everyone becomes a talent agency, remember? The payoff from becoming Lee Hyori is so huge that all the expected value is concentrated in one star. Because that payoff is so large, there are tens of thousands of trainees behind that star. Yet everyone, from trainees to Lee Hyori, puts in an enormous amount of practice, and that’s what we’re seeing with compute. So you might ask, is there no demand for this? No. The new perspective is that demand is being created.
Labs’ Attention Order and Licensed Professions Automated by Buyable Data 43:16
Chester Roh Speaking of data, data is the domain here, so to speak. There’s a set order in which these data domains get used. Naturally, labs will start with large, verifiable, digital ones and begin adding them as capabilities of their models. I’ve used the phrase “lab attention” to describe this. Where that attention goes matters to us too. Because if a frontier lab goes after a domain, we can’t pursue it. So market size divided by the difficulty of buying data probably determines how much attention it gets. They’ll start with domains where data is easy to buy and the market is large and take those domains over completely.
Chester Roh But funny as it sounds, these are fields filled with traditional licensed professions. Fields with even a hint of those professional titles will probably follow that order. Starting with accountants, lawyers, and doctors at the top, down to insurance agents at the bottom. In fields with these professions, once a lab buys the data, it’s basically over. So we need to think carefully about that data. There’s data you can buy and data you can’t buy.
Chester Roh Dataset companies will create all the buyable data and sell it to labs. Take accountants, for example. Accounting, too, is a highly non-verifiable profession. When judging a particular situation, there may be more than one valid answer. It’s unlike math or coding, where you can say yes if it passes a test and no if it fails. You need a way to measure whether a judgment is good or bad. But if you bring in accountants, have them work through practice problems for every possible case, and label the answers, the problem becomes verifiable. It’s a matter of coverage.
Chester Roh That’s what dataset companies are doing now: entering those fields, hiring experts by the hour, and turning all that coverage into datasets. Creating them is simple, too. You just need to write the questions and answers. Our model handles the steps in between, discarding wrong answers and rewarding correct ones. That’s how the system works. If this happens on an enormous scale, the assumption that all knowledge industries will disappear becomes reality. Unbuyable data.
Unbuyable Data and Frontier Labs’ Attack Paths 45:50
Jonghyun Park I’ve seen that same signal a lot: job postings from data vendors hiring experts. There were a huge number of those postings, and there were many rumors that OpenAI was hiring experts directly.
Chester Roh All the labs are doing it. Across many fields, especially physical AI. For a specific task, they put a device on someone’s head and have them do it, you know? And that dataset sells for a high price. So ultimately, if people can generate datasets in some digital form to cover a field, it will all become a feature of a frontier lab’s model. It’s just a matter of time. So where is that absolutely impossible? In small, niche areas. Little niches. Maybe there are regulations that don’t attract a lab’s attention, or the market is too small for labs to enter, or something is protected by a patent. It’s gotten harder, in fact. Labs have three paths of attack. First, buying data.
Chester Roh Then, if buying data isn’t feasible, in biology or pharmaceuticals, for example, the expected payoff is so large. They just build a lab themselves. They build a wet lab to get a reward signal: “We’ve set up a drug development lab in San Francisco.” Then they combine it with their own model. The third, in my view, is a less convincing argument: Midjourney or video generation. Lots of companies are now entering those areas to build their own vertical products, right? The kinds of things Jonghyun Park mentioned earlier. But won’t the labs enter those areas too? With Goliath driving a train toward them, those companies will probably disappear too.
Chester Roh We already talked about talent earlier. One beer and I’m out. So the so-called frontier recipe is no longer a scarce resource. It’s not something only Noam Brown can do; if he can do it, plenty of people elsewhere can too. As for the so-called Near Frontier Labs, Kimi and DeepSeek may be just one or two iterations away from being counted among the Frontier Labs, I think. So let’s say China and the United States move into a different relationship.
The Spread of Frontier Recipes and Regulatory Moats 47:46
Chester Roh For the others, the so-called Near Frontier or Quasi Frontier Labs, the moat is regulation. Regulation or some political reason. That’s why the word “sovereign” comes up. Unless they can drop the “Near,” think of companies that rely on public procurement. In Korea, Tmax and Hancom are prominent examples. Quite a few may end up with revenues around that level.
Chester Roh And open and closed models seem to advance at somewhat different speeds. That’s because Kimi and DeepSeek are also working hard to grow their own services. Really, that whole loop needs to run in one place. Data from customers using the service, new intents that emerge from it, or knowledge about a domain— all of that needs to be extracted and fed into the next generation of pre-training or post-training as training problems. With open-weight models, that loop is broken. When it’s broken, I think the loop runs a bit more slowly.
The Gap Between Closed Operation Loops and General User Demand 48:52
Chester Roh a16z also published figures on token consumption. The people who use it use enormous amounts. And 1% of companies use 90% of all that compute, funded by venture capital or by really large companies. They’re also saying that real demand hasn’t arrived yet. Even if you look at ordinary consumer use, it’s all coding, productivity, or video generation; people aren’t using it for the kinds of applications we usually think of. That’s why people are asking whether there’s a bit of a gap.
Chester Roh At first, I wondered whether there really was a gap. Because it’s absolutely certain that ordinary businesses and ordinary consumers can’t keep up right now. They haven’t even formed a view yet of how to use it. Only the people who’ve decided to keep up with all this are caught up in the bustle; step just outside that group, and people have no idea. Seungjoon Choi, you once told me privately that you felt lonely. Because if you step just outside this field, there’s no one you can talk to about it. So that’s where we still are.
GPU Demand Through Spot and Contract Prices 50:40
Chester Roh And this is another important indicator: “Spot” and “contract” are terms that come up when looking at CoreWeave. For example, xAI’s latest results are good. SpaceX, I mean. Elon built a huge number of data centers. Apparently, he’s built about 3 GW of capacity. They use some themselves, but sell capacity to others too. And now that it’s a viable business, they’re building more to send into space and installing more in Texas and Virginia. And they sell the compute they build on the spot market rather than through long-term contracts.
Chester Roh But companies like Nebius and CoreWeave lock it in under three-year contracts, guaranteeing it for three years. So they sell it at a somewhat lower price. That’s a contract, with a fixed price, while the spot price is set by the market at the time— literally the spot price. The upside is a higher selling price; the downside is that without demand, it just sits idle. On AWS, Spot Instances— It’s almost impossible to use Spot Instances these days. Because compute is in short supply everywhere. Here’s one way to think about it: Right now, spot is much more expensive than contract rates.
Chester Roh Yes, when I try to rent capacity, there’s no spot capacity available at all. I can’t rent GPUs that way.
Chester Roh But will that ever reverse? Whether it will or won’t is one of the key turning points, I think. Let me briefly touch on finance, too. Interest rates are rising. And this is how CoreWeave is structured. There’s an SPV for each contract, which I believe stands for Special Purpose Vehicle. They set up these special-purpose companies, and each one has different terms. The amount raised through bonds, for example, or the hyperscaler backing it— all these things differ. The terms vary accordingly, and any resulting profits or losses flow up to the parent company and determine the value of CoreWeave. That’s how it’s structured.
CoreWeave’s Interest Rate Sensitivity and Maturity Mismatch 52:07
Chester Roh I wouldn’t call it highly sophisticated. It’s a very typical structure. But when interest rates rise—as they did during the railway boom— businesses that had assumed borrowing rates would stay at a certain level when structuring these deals all went under back then. Many people now say CoreWeave’s structure would collapse if the 10-year government bond yield exceeds 6%. That’s what they’re saying. That may be why CoreWeave’s stock price is low and its CDS spread is extremely high.
Chester Roh But I’m not in a position to make a financial judgment about that. I’m just saying this structure exists, and it carries this risk exposure. I think we’ll have to wait and see how it plays out. And this isn’t something I found myself; Fable found it. What people are questioning right now isn’t whether there will be a funding shortfall, but whether the timing lines up. Customer contracts last three years, while data center leases last 15, and the debt matures in five years. There’s a mismatch in that structure.
Chester Roh Because of this mismatch, they seem to think the next round of financing—the next bond issuance— is likely to fail. But that high likelihood of a failed bond issuance is my personal assessment. I don’t think it would be due to shrinking compute demand. I think they could stumble simply because financial conditions change.
Chester Roh Then, of course, someone would step in and pick up the pieces. So this chart shows companies building data centers. On the far left, Google, Amazon, and Microsoft build their own cloud infrastructure. But they’ve been building it with their operating profits— the money they earned. At least, that’s how it worked over the past few years. But this year, Google made news by issuing bonds for the first time. That’s because they’re now making investments beyond what their own cash flow can support.
A Data Center Risk Map from Hyperscalers to Neoclouds 54:34
Chester Roh Even so, these companies are safe. Next in terms of risk is Meta. Meta has plenty of money too, but for the Hyperion data center it’s building, it’s putting up only 20% of the money. The other 80% comes from Blue Owl, I think— a large private equity firm that invests in infrastructure. They’re putting up the money. They put up the money, and Meta, much like CoreWeave earlier, They guarantee they’ll lease it for several years.
Chester Roh The contract even says they’ll guarantee the payments if that arrangement falls through. This way, Meta sharply reduces the share of debt recorded on its balance sheet. But the potential payment obligations don’t appear on the balance sheet; they’re buried in the fine print of the financial statement notes. Those are major risks.
Chester Roh Next in terms of risk are companies like Oracle. They’re funding it directly with bank loans. When they borrow from banks, they usually raise enormous sums, tens of trillions of won. They raised it from JPMorgan or somewhere like that, but because the amount is so large, the bank splits it up and sells portions to others. It’s a syndicated arrangement.
Chester Roh But I’ve heard those portions aren’t selling well. No one wants to take on that debt. So it sits on the bank’s books as a large potential loss. The riskiest companies are ones like CoreWeave and Nebius. Their bonds are junk bonds. “Junk” doesn’t necessarily mean the bonds are bad; people call them junk when the interest rate tops 10%. So that’s how the risk exposure lines up if we look only at data centers.
Jonghyun Park xAI is over on that side too. It must have raised a huge amount through bonds.
Chester Roh Right. They borrowed a lot at high interest rates. But then there’s Elon’s magic. Taking SpaceX and other companies public, sending a Roadster into space and launching rockets, then using the resulting rise in company value to convert the debt into equity, or something like that, makes the debt suddenly disappear.
Chester Roh I think we need to learn from that. So earlier, when Jonghyun Park was talking about Silicon Valley, you said something like this: people there dream big. Instead of working on half-baked ideas, they set ambitious goals and aim to do big things. That attracts better talent and more money, which helps turn those dreams into reality. I remember you also hinting that, in Korea, building a small tool might not be enough. I think this is somewhat similar.
Jonghyun Park Optimism seems to prevail. People believe it’ll work. They take on this risk because they think it will all work out. And Elon Musk seems to be the leading example. In xAI’s case, I didn’t know it had raised capital like that until you shared these materials. In fact, SpaceX already makes more money from renting out GPUs than from the space industry. That may be possible because it raised capital while taking those risks and moved quickly.
Four Scenarios Divided Between Toyota and Ferrari 58:19
Chester Roh Someone said this: These days, you either become Hyundai or Toyota, or Porsche or Ferrari. Only those two types seem to be left. Everything in between has died. That’s what they were saying. It seems only two types can survive: a large mass-market business or a truly distinctive niche business. In AI, it’s exactly what we discussed earlier. Either you become a frontier lab or own a domain-specific operation loop. There are no other options. So if interest rates rise, the companies on the right become vulnerable first, in the order I mentioned. Still, I don’t think the trouble will spread all the way to the left. I’ve spent the past two weeks asking myself whether this bubble is really about to burst. I don’t think it is. That’s my conclusion.
Chester Roh This is the main point today. So what will happen in the future? Ultimately, which probabilities should we bet on? I divided this into an X-axis and a Y-axis. I always enjoy putting everything in two dimensions. Whether the frontier gap closes or widens. That is, whether a company like Anthropic pulls ahead. In that case, models could become scarce resources. If the gap narrows, with Kimi doing it too, DeepSeek doing it too, and the Korean government’s planned Frontier Lab doing it, the gap closes. So what does the model become? The answer is simple: a commodity. No matter how smart a model is, its share of the value approaches zero. That’s what I put on the X-axis. The Y-axis is the measure I showed you earlier for when the industry reaches its turning point. Right now, we’re clearly below it because costs are much higher. Depreciation and interest far exceed revenue. But revenue will eventually exceed depreciation and interest. Right?
Jonghyun Park Hasn’t that happened yet?
Jonghyun Park No, it hasn’t.
Jonghyun Park I was only thinking about investment, but that’s true of depreciation too.
Chester Roh For example, you can say Anthropic makes a certain margin relative to its inference costs, but that’s more like a contribution margin. It only accounts for variable costs. Once you include the money needed for pre-training, staying competitive in the next generation, labor, other expenses, data centers, and upfront commitments elsewhere, they’d probably still be deep in the red.
Jonghyun Park In short, they’re investing more than they’re earning right now. They need to earn more than they invest for the happy scenario, where they don’t go under and things go well, to play out.
Chester Roh Right. I’ve been wondering how this will turn out, estimating the probabilities bit by bit and updating them every day. The quadrant OpenAI and Anthropic most want right now is S3. They maintain the frontier gap and take the whole market, wipe out everyone else, and become monopoly infrastructure providers. As infrastructure providers for intelligence, they’d make enormous amounts of money. Then both their stock prices would soar forever. There’s some chance of that. But I think the gap is more likely to close. I think the number of frontier labs will grow enormously. The gap Anthropic and OpenAI maintain today will close as DeepSeek, Kimi, and others catch up, leaving quite a few frontier labs. So I think the chance of ending up in S1 is the highest right now. Then a few companies that reach what’s called best practice may get there first, or have raised a great deal of investment, or moved early enough to secure some market share and have enough cash flow to keep going. The established players— Google and Meta, and perhaps Anthropic and OpenAI as well—would survive. Labs that don’t reach that best practice—whose revenue falls short of depreciation and interest, whose output is less than what they put in— are what I call marginal builders. They’d all go under. They’d fail and be absorbed by someone else.
Model Commoditization and the Shift of Value to the Operation Loop 62:16
Chester Roh But competition among them would intensify, so frontier models themselves would become a utility. And if they become a utility, a commodity, value won’t accumulate at the frontier labs; it will flow above and below them. The infrastructure below—Nvidia, or power plants and cooling towers—will take sizable margins. They’ll still be scarce resources. Chips do seem likely to stop being scarce, though. Smart people all over the world are trying to replace Nvidia chips for inference, and they’re racing to do it. I think Nvidia’s margins may shift to others too. So where does most of the value go? It’ll move up. Up. So it moves up, but where does it go? To those who own the operation loop. But what if the frontier labs also exploit the entire operation loop by buying datasets or building their own labs, and compete with each other there too? Even that would become a commodity, so all the benefits would go to consumers.
Chester Roh We could enter an era of free superintelligence. But financially, the lower quadrants are interesting, whether the gap widens or closes. Wait, is this a bubble? No profits, just massive investment? That’s S2.
Jonghyun Park A lot of people do see it that way. They see all that investment.
Chester Roh Right. Many do. Then some businesses and infrastructure fail, leaving assets behind to be absorbed by those in the S1 quadrant that have established best practices. S1 and S2 are two sides of the same coin, but the market hasn’t decided yet. The frontier labs are working hard to make money, but the smaller businesses down here, including the many NeoLabs Jonghyun mentioned earlier, could also find a place here. Profitable opportunities will emerge sooner in specific segments, business units, and enterprises where there is demand.
Chester Roh Would you say the dot-com bubble was S2?
Chester Roh Right. It was in S2, but more precisely, looking through the lens we used earlier, the industry may look like S2 now, but as it settles and market uncertainty fades, it always moves up into the S1 quadrant. So together, S2 and S1 have a probability of almost 80% in my view. I think the gap is very likely to close, which is why their combined probability is so high. The turning point for revenue, depreciation, and interest is a matter of timing: when will it arrive? That’s the most important variable. I think S4 is extremely unlikely. So does a business that owns an operation loop necessarily need to do model work? Not necessarily. Owning an operation loop has no necessary connection to owning your own model. I’m already proving that through several companies.
Weight Ownership Depreciates, Signal Ownership Compounds 65:48
Chester Roh Whether it’s Jev or something else, you can take a near-frontier open-weight model and plug it in. Often, just adding compute solves the problem. In fact, the loop you own is much more likely to be in a control plane outside the intelligence layer. And those control planes don’t hold that much data. So even if you gather that data and feed it to a model, the domain often can’t provide enough for the model to learn and improve. Competition is headed in a direction where increasingly intelligent models will inevitably be released with open weights. That’s where things are headed.
Chester Roh When a frontier lab moves ahead, those following it need attention and publicity, so they release their weights to get noticed. That’s why I put this below. I didn’t write it; a model did: owning weights depreciates; owning signal compounds. So the value of model work will keep falling. Many people think the value of model work today lies in data and talent, but I see both as depreciating assets. If compute can replace both and solve the problem, you could conclude that you don’t need to do model work. At least, that follows logically. If you intend to compete with a frontier lab, Of course you have to do model work. But if you choose an area of operations that frontier models can’t reach and build a loop there, you don’t need to do model work.
Jonghyun Park What do you mean by “signal” here?
Operation Signals as Compressed Wisdom 68:01
Chester Roh I’m still working out a precise definition myself. And by “signal,” I don’t mean the size of a dataset or its coverage for the model work we’ve discussed. Every company has signals like that. At our company, for example, revenue might fall in a certain way. Then we identify the problem, fix it, and take the necessary steps. As those cases accumulate, they become a dataset. It’s question and answer, over and over. As that continues, eventually it gets compressed into a few rules. Sometimes we call those rules wisdom; sometimes we call them a rule base. They’re not substantial enough to embed in a model, but they’re compressed forms of wisdom. To draw out those rules, we have signals arising in the business every day. Under certain conditions, this happens, then that happens; customers dislike certain actions, and so on. That’s what I mean by an operations loop. In other words, every existing business has one. Say several companies are in the same industry, including established leaders. In a mature industry, they might have market shares of 70%, 20%, and 10%. The leader will have a lot of operational signals because it already has so many transactions happening within it. If those companies combine their loops with AI and build them well, even frontier labs can’t challenge them. They’ll gain the efficiency improvements AI brings and evolve into the next generation of businesses. If they don’t, a startup from outside that sees the opportunity through this lens will challenge them.
Chester Roh I think this is what will happen after the competition in frontier models, coding, and model work that we’re familiar with now. That’s why, among my three Out of Model Work candidates, I put B2B first: building loops within operations. As we talk about enterprise AX, I’ll keep discussing this with Jongok Ko going forward. Companies still don’t know what to build or how to build it. Right now, most efforts use AI to replace tasks that existing knowledge workers perform. That makes the work faster. It’s faster, but it only makes those workers’ lives easier. It hasn’t raised productivity across the company yet. So now is the time to redesign workflows completely around AI. That’s the lens we need. Businesses like that will emerge. That’s why companies like Palantir, EightyNinety, OpenAI, and Anthropic have all built their own B2B FDE teams.
B2B and B2C Workflow Redesign and the AI Agent Competition 70:06
Chester Roh In B2C, I also talk about this a lot with Simon Kim, CEO of Hashed. In fact, Muse Spark, Gemini Spark, and Grokbot are all examples. Kakao and Naver will build them, too. So will SK Telecom. Everyone will. Ultimately, the battle in B2C is over who gets in front of consumers. Once you have that position, the businesses you can capture with that reach are enormous. So every company will head in this direction. Then which companies will win? This comes from my experience as a product manager at Google. You can put coverage and comprehensiveness on the X and Y axes. In search, broad coverage is the biggest factor in user satisfaction. So, for example, if you give this agent a job, it has to handle everything from restaurant reservations to small things like keeping me company, and even answering questions.
Jonghyun Park How many different things it can do. Right.
Chester Roh Huh? It can book a flight to Italy, but it can’t buy me Samdasoo water? That’s a gap in coverage. So coverage has to be broad, and for each task, how exceptionally well it can perform is a kind of comprehensiveness. The result is the product of the two. And there’s one more lens to consider: when to enter the market. That timing shifts so dynamically, though. And number three is similar to number one. Ever since I started the podcast, the ideal I’ve had in mind has been Tesla. If I wasn’t building a frontier lab, I’d do it like Tesla: bring in the technology I needed from outside, integrate everything internally from the bottom all the way to the top, and create a signal loop within the company itself. But at some point Elon thought, “I need a frontier lab after all,” so he created xAI and ended up with one of his own. I envy him. So I see two things. You should own frontier lab stocks. It’s insurance. We don’t know when the flipping point for that exposure will come. I don’t think there’s a high chance of one or two frontier labs forming an oligopoly. But as frontier labs emerge— OpenAI, Anthropic, SpaceX, and maybe DeepSeek will go public too; I hope it does— when they go public, you should keep holding their stocks as insurance. Because, fundamentally, they’ll become the infrastructure for all this intelligence, and their shares will be assets that protect you against inflation and everything else. So I think you need to own them.
Vertical Integration and Frontier Lab Investment as Insurance 73:05
Chester Roh And your business should go where they aren’t operating. That’s obvious, of course. Find the biggest opportunity in those areas. Or find companies that already have an operation loop and quickly build something for them. That’s why I’m thinking of starting a B2B enterprise AI company. I’ve been thinking about it since I created Runaways’ Alliance. I kept mulling it over, but now I feel the timing is right. I track an indicator, though it isn’t quantitative. By following a few media sources and organizing what I find, I can derive that indicator. I check it about every week or two and set my own criteria for measuring it. I want to gauge the market’s timing. So I watch how the spread between spot and contract prices narrows and widens. Then number two—we could leave this one out— is whether you use a GPU for three years and retire it, or use it for eight years. That changes everything. So how long does it take to recover the hardware investment through revenue? What’s the payback period? Because if the payback period is short, everything after that becomes margin. For companies like CoreWeave and Nebius, if the hardware is still running after it’s fully depreciated, it generates net profit from then on. So today’s bad numbers could all turn good if demand for compute continues. That’s why I watch this too.
Seven Indicators for Reading AI Market Timing 74:24
Jonghyun Park It seems that’s actually happening now. GPUs we thought had been retired are still generating a lot of revenue. A100s, for example. Because compute is so scarce. That seems to be what people are saying.
Chester Roh Then there’s the gap between open-weight models and the frontier, which is effectively the gap between China’s near-frontier labs and America’s frontier labs. Right now, the two are almost equivalent. That gap was four months, but it now seems to be shrinking to less than two months. You might conclude that China’s labs are becoming frontier labs themselves.
Chester Roh Then, most of the agents we have now are for coding. So I’m watching how many new domains emerge outside coding. I’m looking at areas like legal and finance. Those are obvious ones. Another thing to watch is which major domains come next.
Chester Roh Number six might be an unfamiliar concept, but I see it as one indicator of where value capture is going. I don’t have a definitive measure for it yet. I’m just going by people’s sense of it. Say a consulting firm that runs FDEs goes in and does a project. I look at whether the token cost is far greater than the firm’s margin, or whether its margin exceeds the token cost. It’s the ratio between the two. That gives you a rough idea of where the value is being captured.
RSI Variables and How to Watch Frontier Labs 77:16
Chester Roh And number seven is when RSI is first observed. Once that starts happening, it might be fair to think we’re doomed. Can a model really change the slope of the scaling-law graph as we know it? If something emerges that can improve itself, we could be entering a rather frightening era. I’ve talked about numbers one through seven, but perhaps the biggest risk in watching these indicators is that all of this could become moot if superintelligence simply arrives before the year is out, and that’s it.
Jonghyun Park You’ve spoken for nearly two hours, assuming all of this follows the current trajectory, and I think you’ve explained it all very well. But as you mentioned earlier, with RSI and all, if that’s what Ilya Sutskever is betting on, new talent could bring new research breakthroughs, like continuous learning. If things like that emerge, this whole trajectory could change completely. In any case, I think those are things we need to keep monitoring. I feel the same way.
Chester Roh Yes, exactly. So much has happened over the past two weeks, and we’re interpreting those events, thinking them through and organizing them from our perspective. There is one overarching theme: watching the frontier labs closely is most important. It’s important to see where they fall short. Then watch how the timing shifts in between. I think that’s about it.
Chester Roh I don’t do this every day, so I try it once or twice a year at pivotal moments like this to form my own perspective. If you see things differently, your view is right. Jonghyun, you’ve just returned from the U.S., so take some time to get over the jet lag. Next week, let’s come back with more interesting stories and have some fun talking about practical matters.
Jonghyun Park I really enjoyed listening today. I think people who, like me, have been listening to every episode of AI Frontier Korea probably share a similar perspective. People who’ve bet on AI continuing to grow and transform our lives are hard to find just outside our circle, as Seungjoon said. We often say we’re trapped in that bubble. We tend to put it that way a lot. Everyone will use this to plan their next moves and act on them.
Education for the AI Era and Closing 79:48
Chester Roh Exactly. Right. One last quick plug: Seungjoon Choi couldn’t join us today because he’s busy. He runs Hanmi Kindergarten in the Ilsan area. It’s a rare AI kindergarten in South Korea, isn’t it? So if you’re in the area and, in this AI era, you’re wondering how to educate your young children, then please consider applying to Hanmi Kindergarten. That wraps up today’s recording.
Jonghyun Park Thanks for your hard work.