[BidClub_]
SemiAnalysis · · 50 min

Ep. 022 - Market Drawdown, Historic Bubbles, Funding The Buildout, AI Politics (Doug is Back)

Doug

YouTube
TL;DR
  • The memory trade is unwinding from euphoria, not necessarily from broken fundamentals. Doug says DRAM and NAND prices roughly tripled over the past year but may rise only 30–50% next year; that slowing rate of change, SK hynix’s earnings miss after shifting more volume into LTAs, and widespread leverage helped turn the KOSPI’s 40% fall into forced selling. “What goes up really, really fast often has a little bit of gravity.”

  • A 40% correction need not end the cycle. Doug compares Korea’s behavior—not its magnitude—with Taiwan’s late-1980s bubble, where banks reached 500 times earnings and the market endured two 40% retracements before continuing higher. His qualified call: “I’m not saying it’s over. I don’t think it’s over,” but a boom-bust this violent usually takes time to repair.

  • Chinese memory may compress margins before it eliminates scarcity. CXMT is incentivized to maximize provincial output rather than shareholder returns and could “ruin the party” by accepting roughly 10% gross margins, but Doug still sees more demand than supply. Apple turning to CXMT while accusing Micron of price gouging earns his verdict: “No crying in the casino, Apple.”

  • AI demand is the trillion-dollar unknown. Doug’s bear case is that chips, utilization, inference, and models improve faster than useful workloads expand—the fiber-boom pattern, where one strand eventually became 500,000 times more performant. The host counters that models grow, think longer, attract more users, and unlock new markets: SemiAnalysis went from roughly nine to 90 coding-agent users, then perhaps 10× usage per person, producing 100× AI spend.

  • H100 obsolescence produced the sharpest disagreement. Doug expects sufficiently large models to make old accelerators uneconomic and says a B200/B300 pricing divergence would confirm it; the host argues Hopper facilities cannot simply be swapped into B300, GB300, or Rubin because power, cooling, permits, and layouts differ. “We live in a world with friction,” and that friction can preserve older hardware value.

  • The buildout could outrun its financing even if AI ultimately works. Nate Silver, as labeled in the transcript, estimates roughly $150 billion of ecosystem ARR against $1 trillion of capex already committed; at a hypothetical 50% profit margin, that is only a 7.5% return. Revenue reaching $500 billion may support $2–3 trillion of investment, but $5 trillion against $500 billion of revenue is “a house that you cannot pay for.”

  • Capital, electricians, and politics may bind before chips do. Nate Silver cites a roughly 100,000-person US electrician gap and about $450 billion of hyperscaler debt raised year-to-date; doubling again could demand higher yields and crowd out other borrowers. Nate Silver and the host also discuss AI becoming a cost-of-living scapegoat in the midterms, creating regulatory risk even if voters do not treat AI itself as a top-three issue.

Digest · the substance, structured for research

1. Korea’s memory boom hit leverage before it hit a fundamental wall

  • Doug dates the setup to the end of Q2, which he calls effectively the best semiconductor performance in history. The subsequent unwind may look “technical”—various factors, excess leverage, and forced selling—but his simpler explanation is that “what goes up really, really fast often has a little bit of gravity.”

  • The speculative behavior in Korea echoed earlier Asian bubbles: investors borrowed aggressively and in some cases took second mortgages to buy stocks. Doug’s deliberately harsh history lesson is that Korean investors bought banks in 2007, CMBS in 2007, and SaaS in 2021 before going “ultra mega yolo into itself.”

  • SK hynix missed consensus, partly because it shifted more volume into long-term agreements, making future price increases more conservative than the market’s euphoric expectations. DRAM and NAND may have tripled over the prior year, but another tripling was not expected; even 30–50% price growth looks disappointing when “finance brains are just absolutely broken. It’s all about rate of change.”

  • Doug says the KOSPI was down 40%; anyone above 2× leverage could therefore be wiped out, creating margin calls and reflexive selling. Yet he expects markets to overshoot in both directions: “Things are never as bad as feared, and they’re never as good as you think they will be.”

2. Taiwan’s historic bubble leaves room for another rally—and a warning

  • The comparison is about behavior, not equal magnitude. Doug says late-1980s Taiwan produced roughly a 100× per-capita bubble, including a bank at 500 times earnings; recreating it would require Korea’s market to become perhaps 15–20 times more valuable.

  • That Taiwanese bubble nevertheless contained two 40% retracements before moving higher. Doug therefore refuses to declare the current memory cycle finished, while warning that a leveraged boom-bust “usually takes a little bit of time to retrace out.”

  • The host presses the bullish fundamentals: LTAs remain in place, businesses are sold out for years, and new production takes time. Doug agrees there may be smart money waiting, but argues that lower future price increases also deserve a lower capitalization multiple—even if this cycle proves “bigger and longer and stronger” than prior ones.

3. Chinese memory threatens pricing, while the real demand curve remains invisible

  • Doug’s historical rule is that “everything they touch gets dumped” because Chinese producers can prioritize volume over returns. His framing is provincial competition for GDP: the government is effectively the shareholder, so CXMT can accept around 10% gross margins rather than protect shareholder returns.

  • CXMT is still, in Doug’s view, clearly number four, entering during a shortage and therefore likely to make substantial money first. Apple is reportedly using CXMT after accusing Micron of price gouging, but Doug dismisses the complaint: “The reality is you have to buy it at the market price.”

  • The supply curve is comparatively knowable; the demand curve is not. Coding agents and knowledge work clearly add consumption, but the market cannot yet tell whether the increase is 50%, 100×, or something in between or larger. Supply will keep “ramping blindly into this curve” until it finally crosses demand.

  • Shortages corrupt the signal because a gigawatt-cluster builder may double-order equipment and triple-order memory, knowing excess parts can be resold. Factories interpret those orders as durable demand, add capacity over several years, then face a demand slowdown; utilization can fall from 100% to 50%, forcing price cuts—the semiconductor bullwhip in miniature.

4. AI usage could outrun efficiency—or repeat the fiber glut

  • Doug’s strongest bear case is useful-market saturation: perhaps Kimi K3 becomes good enough for data entry while chips, software, and models keep improving. The internet carried similar claims that demand doubled every 90 days, yet one fiber strand eventually became 500,000 times more performant and capacity could no longer be filled.

  • The host’s rebuttal is that four inflationary forces offset deflation: models get bigger, think more, reach more people, and are used more per person. Kimi K3 reportedly tripled in size and fits on a single B300 or MI355X node, but cannot run that way on Hopper.

  • The host also sees 100–1,000 times more potential users plus workloads extending beyond chat: coding, research, video, images, drug discovery, and material science. His best example is that few startups are pursuing weather prediction because everyone is currently distracted by coding—not because the opportunity set ends there.

  • Dylan says the breakthrough in Claude 4.5 changed his mind by crossing a capability threshold: work impossible the day before suddenly became feasible. SemiAnalysis then moved from fewer than 10 coding-agent users to around 90. Doug confirms average usage is higher but is unsure whether tokens rose 10× because his initial usage included intense 14-hour sessions; Dylan argues spend rose 10× per person through sub-agents and related tooling, taking company AI spend up 100× or more.

5. Model growth does not settle what old GPUs are worth

  • Doug speculates that old chips eventually become uneconomic: if inference requires 100 H100s, operators may prefer one newer system and “just let the old girl go.” He would look for a pricing divergence between B200 and B300 as confirmation.

  • The host “completely” disagrees because almost nobody can rip H100s out and replace them directly with B300, GB300, or Rubin. Hopper data centers have different power and physical designs; retiring them requires demand to fall below the operating cost of running them, or such extreme scarcity of permitted power and land that owners demolish functioning facilities.

  • Doug concedes the friction, while the host reduces the broad trade to two cases: stalled model progress pressures GPU prices, while continued progress supports them. Government restriction is the X factor—limiting access to frontier systems could reduce aggregate demand and particularly hurt older GPUs—though Doug declines the scenario as containing “too many what ifs.”

6. AI politics will likely arrive disguised as cost-of-living politics

  • Nate Silver, as labeled in the transcript, guesses AI will be a top-five but not top-three issue for most midterm candidates: widely mentioned, rarely a platform’s centerpiece. When an issue lacks that priority, he expects corporate lobbying to dominate.

  • The host points to the ROSA bill, which he tentatively describes as the Remote Access Security Act. It reportedly passed the House roughly 300–20 but remained stuck in the Senate amid lobbying—an example of how a fourth- or fifth-ranked public concern can lose institutionally.

  • Their shared political concern is scapegoating. Nate Silver expects cost of living to be voters’ number-one issue, with AI and “tech bros” blamed for inflation, housing, employment, healthcare, or climate pressures: not a referendum on whether America should sponsor AI, but AI as a “whipping boy” for existing grievances.

7. Cash-flow timing, debt capacity, and labor narrow the buildout path

  • Nate Silver’s central financing warning is that the future can arrive too late. A technology boom may ultimately justify its vision, yet investors can spend $1 trillion to reach $100 billion of revenue and wait five more years for that figure to become $1 trillion: “You built a house that you cannot pay for.”

  • His rough ledger is $150 billion of ecosystem ARR against $1 trillion of capex. At 50% profit, the implied return is around 7.5%; growing ARR to $500 billion could probably carry $2–3 trillion of investment, but a later doubling toward $5 trillion makes the path much narrower and demands adoption “like yesterday.”

  • The host sees enterprise, banks, telcos, defense, intelligence, and government agencies as more important than getting every grandmother to subscribe. Nate Silver’s concern is operational: chips alone do not create deployments, and the US already has a roughly 100,000-electrician gap; training takes perhaps 18 months, while journeyman-level electricians can make $250,000 and potentially $400,000–$500,000 on extreme hours.

  • Capital has its own supply curve. Nate Silver puts hyperscaler debt issuance near $450 billion year-to-date; more borrowing may require 7–8% yields and push up mortgage rates compared with hypothetical mortgage borrowing near 5%. Life insurers and annuities are important funding sources, but pensions are secular decliners and “everyone needs twice as much insurance” is not a scalable financing thesis.

8. Data centers spread the boom locally, but permitting can reverse it

  • Taiwan illustrates the physical limit: Dylan says TSMC and its supporting ecosystem may represent roughly 20% of the economy, while Taiwan’s GDP rose about 25% as chip output surged. The host also notes that TSMC directly employs fewer than 100,000 people in a population above 20 million and can expand abroad; Nate Silver answers, “No electricians in Arizona, bro.”

  • Their oil-boom analogy cuts both ways. Data centers can create construction and service economies in remote places because energy availability and permits—not natural deposits—determine location; a permitting or political shift can therefore strand a community as quickly as a commodity bust.

  • Nate Silver cites a poll suggesting that people who dislike data centers often do not live near one, while communities—especially younger people—can become more favorable after jobs arrive. Doug adds that construction work is accessible and honest employment, and uses a tentative estimate of 10,000 jobs per gigawatt: 70 gigawatts could mean roughly 700,000 jobs. Unlike an app built in San Francisco, a $10 billion information factory in a rural area creates “broad participation” and more sustainable investment.

Dylan Patel

I'm joined this week by Doug O'Loughlin, who, for some reason, hasn't been on the podcast for a little while. Doug, what's going on, man?

Doug O'Loughlin

I thought I wasn't allowed to come on here, but apparently that was never the case. I don't know. I've just been doing a lot of things here at SemiAnalysis. It's been a crazy season.

Dylan Patel

Doug has been busy, but he's back. Today, we're going to talk a little bit about the drawdown that's going on in stocks right now: memory, optics, and lots of AI names are down. We're going to either fan the flames or provide people with some soothing positivity.

Doug O'Loughlin

Positivity.

Dylan Patel

TBD on that one. We're also going to juxtapose that drawdown with some of what's going on in the labs, like Sam Altman warning staff that their new pre-trained Doug is going to be RSI in 6 months. I'm just continuing to build a ton of models. Anthropic as well—there are rumors of new versions of their models coming out. We saw Opus 5 get released, GPT-3, and we'll get Doug's reaction to that. There have also been some moves around the industry, with people leaving.

Okay, let's start with the lay of the land, man. Stocks are in a drawdown. What's going on?

Doug O'Loughlin

So, let's just talk about it. I've been doing this spiel a little bit. Pretty much up until very recently—let's say by June 30, at the end of Q2—it was pretty much the best performance in semiconductor history. We've had a little bit of an unwind since then.

You can argue that a lot of it is “technical,” meaning there are reasons related to various factors, and people were maybe overlevered and things like that. But the reality is that what goes up really, really fast often has a little bit of gravity, and we're just paying for the massive, almost crazy momentum rally we've had.

It's gotten a little sloppy, and the tape has been very interesting recently. South Korea is pretty crazy. Every single day, the stock markets essentially hit their limit. There’s a tweet right now that says, “How do I even do my job?” The head of HR lost everything, everyone is super depressed, and their stocks are all down. It's just kind of crazy.

If you look back through the history of Asian financial markets, this ironically seems to happen a lot more often than you think. One of my favorite books of all time is The Great Taiwan Bubble. I think Taiwan, on a per-capita basis, had something like a 100x bubble. Just incredible numbers. Everything went to something like 1,000 times. You were talking about a bank trading at 500 times earnings. It was crazy stuff everywhere.

Dylan Patel

When was this?

Doug O'Loughlin

It was in the late ’80s.

Dylan Patel

So, are you comparing the current situation in Korea to Taiwan in the late ’80s, or are the vibes similar?

Doug O'Loughlin

The vibes are similar. The bubble in Taiwan is truly historic. I just don't think anything—I don't know. Just put it this way: If we were to hit that bubble in Korea today, I think Korea's stock market would be worth 15 to 20 times more. It was such a crazy bubble. I don't think we should ever be able to recreate that. If they can go off, I guess.

But I do think some of the lessons learned and some of the behaviors that happened during that bubble are often indicative of this one.

Dylan Patel

Yeah.

Doug O'Loughlin

I don't know what to call it—a bubble or whatever—but maybe this crazy run-off. What happens is that you have these blow-off tops that get absolutely crushed. At the absolute top, everyone's buying as much as they can. They go superlevered. People are literally getting second mortgages on their houses to buy more stocks.

This is actually what happened in Korea. Koreans, particularly, for some reason, have this crazy hit rate of always buying the top of every cycle. They buy the top of every cycle. They were buying banks in 2007. I'm not joking. They were buying CMBS in 2007. They were buying SaaS companies in 2021. They have this 20-year track record of always buying the top.

Korea just went ultra-mega-YOLO into itself, and I think it's a—

Dylan Patel

Can you clarify a little bit about what's going on here? Do you think the fundamentals are different in this case?

Doug O'Loughlin

On the fundamental side, I think they're good. But the problem is that things are never as bad as feared, and they're never as good as you think they will be.

SK hynix missed earnings today and missed consensus. The reason why is that they shifted more to LTAs. It's ironic because they were in the United States, and when they were doing the ADR, they were dumping on Micron for doing LTAs and effectively getting a lower price.

One of the things that's happening in this memory cycle is that memory prices—DRAM and even NAND—have probably tripled over the last year. Next year, they're not going to triple again. They're going to go up 30% to 50% or something like that.

But finance brains are absolutely broken. It's all about the rate of change. Historically, in the memory cycle, when that rate of change goes down—when the second derivative goes down—it's usually the end, because it doesn't just go up to a 30% price increase; it usually goes to -50% price increases.

What happens is that they invest all this capital, all the bits come online, and then, once upon a time, they're saying, “These factories—we can't even run them hot enough because we're running out of chips. There's so much demand.” Then they build another factory and say, “Why was there so much demand then? There's so little demand now?”

People were double- or triple-ordering in order to get their orders in. This is a typical bullwhip in a cyclical thing. But we do think that this cycle is bigger, longer, and stronger than past cycles.

I think what's happening is that people are looking out, and the price increases that the memory companies are showing in terms of future price increases because of the LTAs are lower and more conservative than what the market had hoped for. The market was euphoric. People were way over their skis at the top. They're levered up, and the stocks are moving against them. They're down 2x or 3x now.

If you're 2x levered and your stocks go down 50%, you don't have any more stocks, right? Many people were at that level. The entire KOSPI is down 40% right now, so if you're over 2x levered, you're completely wiped out. We're seeing crazy margin calls, and that becomes a self-fulfilling thing. Everyone who has a lot of money on the table looks at it and says, “Damn, this pile of money is shrinking. Maybe I should just sell.” Then that exacerbates it.

Part of this, too, is that the run-up was truly historic. I don't know. Being cautious in any way during that period of time, everyone pretty much laughed at you. Now people are probably going to freak out and overshoot on the downside. This is what markets are all about.

Dylan Patel

Yeah. Do you think there's some smart money on the sidelines right now that's still looking at this and saying, “Demand is strong. These LTAs are going to continue. These are really healthy businesses that are completely sold out for the next number of years, and it's going to take a little while to bring more production online anyway”?

Doug O'Loughlin

I think so, but the reality is that now there's going to be this overhang: price increases aren't going to go up as much as they did, and the multiple they capitalize it at is going to be lower. I'm not saying it's over. I don't think it's over. But when you have that big boom-bust, it usually takes a little bit of time to retrace.

Dylan Patel

That, or it's going straight back.

Doug O'Loughlin

Okay, sorry. Let's bring this all the way back to the Taiwanese stock bubble. I was up late last night, literally looking at stock charts of the great Taiwanese stock bubble. During the bubble, you had two 40% retracements. It's not unheard of to have this giant retracement and then continue to go higher, but it's pretty dramatic, and there's a lot of pain right now.

Dylan Patel

Well, okay, let's talk about the people who are calling a top, selling, and trying to short right now. They have different approaches to calling bullshit on all of the demand. We can go through a few of them.

One that was hot in the news recently was the introduction of Chinese memory to the ecosystem: CXMT and YMTC—obviously, big IPO. What's your take on Chinese memory coming into the memory-makers' industry right now?

Doug O'Loughlin

I think on the Chinese side, historically, everything they touch gets dumped.

And that's why people are scared: China has the ability to ramp a lot of capacity, and even if it's at a worse yield, it's not like China is trying to make a 10% gross margin or a 50% gross margin. Historically, they usually make 10% gross margins, and they're more than happy to dump to win market share.

The correct framing for how Chinese companies compete against each other is really provinces competing against each other for GDP output. They're not actually competing against each other for profit margins, EPS, or shareholders, right? The shareholder is the government. The government incentivizes production, and each province competes against the others. CXMT is incentivized to produce.

I do think the CXMT LTAs are a little bit—I don't know. I think they're clearly number 4 in the market, and it's a shortage, so they're going to make a lot of money in the meantime. Apple, for example, is down to using CXMT memory because they're accusing Micron of price gouging, right? No crying in the casino, Tim Apple. Well, I guess he's retired. No crying in the casino, Apple. The reality is, you have to buy it at the market price.

Dylan Patel

Yeah.

Doug O'Loughlin

And so I think CXMT might ruin the party, but the reality is there's still just more demand than supply. And the real question here is how do we know the demand is going to be strong? Because, okay, now I'm just going to go on a rant. Demand and supply are always yearning toward each other for the crossover of price. But in a new market like this, they're blindly searching. We do know the supply curve. The supply curve is relatively easy to understand. The demand curve we actually don't know. We know that coding agents and stuff like that means that there's a lot more demand. We know that chatbots and the value of knowledge work mean that there's more demand. But we don't know if it's 10x more demand—maybe DoorDash is 100x more demand, or maybe it's 50% more demand. Also, supply is getting better all the time, so supply is just going to ramp blindly into this curve until one day it meets demand. When something is super in demand like it is right now, you're sitting there and saying, "Okay, I want to make a data center today. I want a gigawatt cluster. If I make a gigawatt cluster, I'm going to land Anthropic tomorrow. I need my [__] to come on time. What am I going to do? I'm going to double-order my equipment. I'm going to triple-order my memory because, hey, if I get memory and it's a little bit late, I could just sell it to someone else. It's a shortage everywhere." During semiconductor cycles on the upside, everyone's double-ordering, so the factory looks at it and says, "Oh my god, there's so much demand." They almost always overbuild for a level of demand that doesn't come. Historically, there's usually some kind of economic wobble, like a financial crisis or the Fed freaking out. Then demand weakens for a quarter, while supply ramps because you can't stop a factory from coming online. Your factory goes from 100% utilization to 50% utilization, and the only way to make money back on your factory is to cut price. That is essentially the semiconductor market in a nutshell. The real question is, where is demand? Supply can ramp 2x over 2 years, and SemiAnalysis exclusively focuses on that. The real question is where demand is—the trillion-dollar question. And we think it's wrong. So.

Dylan Patel

Yeah, I mean, I have my opinions on that. In my view, I think demand is pretty obviously very strong for a very long period of time here. I just look at my own internal usage. I look at how others are using it. If you're calling a future where demand stays flat, does not increase that much, or starts to decrease, you really have to believe that the models are not going to be getting better in the future. I see zero signs of this. I only see signs of the opposite. It seems hard to imagine a future where next year or the year after demand is suddenly caught up to by the supply curve as you're describing. Maybe I'm—

Doug O'Loughlin

Wait, so I'm going to play devil's advocate here because I do agree with you. As a guy who now spends a crazy amount of tokens a week and is probably hooked on tokens, the 2 things that people will say are that the technology gets better at a faster rate than people can use it.

For example, let's say the real killer application for AI happens to be data entry, and Kimi K3 is good enough. Effectively, we make faster and faster cars and better and better products, but the real demand curve that matters gets saturated by a product that we already mastered. You can argue this was the internet, right?

The reason why the internet bubble—"Oh my god, demand is doubling every 90 days" was one of the common frames—it wasn't. The technology behind making the links faster literally got 2x or 3x better every single year. Then, all of a sudden, after everything happened, one strand of fiber became 500,000 times more performant. They were like, "Wait, I don't think we actually need—we can't actually fill this fiber demand."

That would be the pushback: the models get 100 times smarter and 10 times cheaper, and that level of intelligence saturates the big market. I think that's probably the biggest bear case I could think of, and I'm wondering if that's probably true for some parts of the economy.

You don't really need to say, "Center my div." Does Kimmy need to do that, or does Fable need to do that, or can Sonnet do that?

Dylan Patel

My burrito and stuff. No, I think absolutely not, but to me there's still 100 to 1,000 times more people who are currently not using any of the currently good-enough models who will learn to use them over time. That's 1. Number 2, I see a whole variety of other use cases.

The internet was connecting people to the internet, but AI is coding. It's also chat and research. It's also video generation, image generation, drug discovery, or material science. Somebody could start making superconducting elements with AI or doing all sorts of other material science work.

Doug O'Loughlin

What's that worth?

Dylan Patel

Well, I think it's worth a lot of spend on GPUs, for example. In many cases, the fact that we don't have 10 startups all working on weather prediction right now for farming, leisure, or whatever is almost purely because everybody's distracted by coding.

Meanwhile, coding in and of itself—the claim that centering a div is going to run out because everybody's going to have a website for themselves, fair enough. But coding also represents a whole bunch of other tasks that are so much more economically valuable intrinsically than centering a div, and you can pursue those in this. I mean, there are rumors of Sam Altman talking about RSI, right?

Doug O'Loughlin

Yeah.

Dylan Patel

Yeah.

Doug O'Loughlin

Yeah.

Dylan Patel

I'm going to push back because I don't believe this. I'm more on your side of the camp. The question is, will that still fulfill and satiate all of this? Will it push and do all this economic good work while we build all the data centers?

I think the problem is that the chips get faster every year. We also get better at using the chips, and then the models get faster and the models also get better. So you have 4 rounds of deflation to just make really good products.

All of a sudden, you can look up after 3 years and you're like, "Wow, every person in the world has an 8x B200 to do whatever, every single token, every single day." And then we're like, "Ah, we've run out."

Doug O'Loughlin

You walked through those 4 that you say are deflationary. Let me tell you the inflationary ones, right? The models get bigger every year. The models think more every year. More people use the models, and more people use the models per day. Each individual user uses the model more.

You can look at this as it's happened. Kimi just tripled in size, and now it can only fit on a B300 or an MI355X single node. You can't run Kimmy K3 on a Hopper, right?

Dylan Patel

Yeah. Yeah, I was going to actually say that. I was actually going to talk about that. Yeah, I was actually—do you want to talk about that because that's a spicy take? We talked about the—oh, sorry, two—one last thing because I know we're bantering on the demand thing. The thing that changed my mind more than anything else truly, though, was last year in coding agents, in Claude 4.5. I don't know about you, but there's never been a clearer moment where you hit some level of intelligence on the curve and an entire new market showed up. You know what I'm talking about? I could not do this before, and then the day after it came out, you could do it, and I—

Doug O'Loughlin

You were the prototype, right? There was a time last year when the technical staff at SemiAnalysis—call it under 10 people—were using coding agents. Then, sometime in November or December, you and Dylan said, "Every single person at the company needs to learn how to use this thing." Now we have around 90 users of it.

Dylan Patel

So that's 10× right there, from 9 to 90.

Doug O'Loughlin

Yeah, and I think the interesting thing is that some of the work gets better. But anyway, smarter models and new capabilities—that's the first part. The second part: do you use more tokens?

Dylan Patel

Sorry, from the time you started using it in November to today, do you use more tokens than you used then, per user?

Doug O'Loughlin

On average, yes. On average, yes.

Dylan Patel

I think it's like 10 times more tokens than initially, right?

Doug O'Loughlin

I'm unsure, because that first week I literally did 14-hour shifts for a little while. I can't—my initial coding psychosis was pretty hard. That was sub-agents and bigger models. Maybe do it by spend, not—

Dylan Patel

Actually, yeah, by spend. But 100%. If we do it by spend, not even close, right? Because of sub-agents. Do you remember that conversation in Claude Code? When it first came out, everyone remembers that one random-ass paper called Gastown, which was like, "I'm going to create this self-healing whatever agent harness that would do this crap." And everyone was like, "This is crack cocaine, but I kind of vibe with it." I would argue we were kind of there, but you're seeing sub-agents, tool usage, agents, and multi-agent systems—the fact that you fan out tools like this. You're already seeing the realized version of this, but I think back then it was impossible to market.

Thinking mode, right? They're doing more. So, in our experience, that was 10 times more users almost overnight, and then over a span of 3–4 months, people probably 10×ed their individual usage. That led to our company going 100× on AI spend, if not more.

Doug O'Loughlin

Yes. Yes.

Dylan Patel

Now, the question is: does every—

Doug O'Loughlin

Every company do this? Probably not at our level, but I do think a lot of companies have a lot of work to chop.

Dylan Patel

Well, your argument is that this will happen over time, but it will happen on the cheaper models because companies will have budget constraints or something like that.

Doug O'Loughlin

Yeah, something like that. Actually, two things. Let's talk about the H100 thing, because I think that's really interesting. We talk about these new models, and it's very clear to me that the old chips will be worthless. Everyone's like, "Oh, the H100 is an appreciating asset." But at some point, it's going to take 100 H100s to run inference on one of these models, and you're just like, "Dude, just let the old girl go. Just get a B300 or, you know, a VR200 instead."

That's my speculation, and we'll see. We'll obviously be writing notes about it, but the true confirmation of this trend is if there's a pricing divergence between the B200 and B300.

Dylan Patel

Okay, I completely disagree with this, but I'm interested in you fleshing it out. Maybe the most fundamental reason I disagree with this is that I don't know anybody who's ripping out H100s to replace them with B300, GB300, or Rubin, specifically because the data centers are so completely differently designed.

So, you would need to justify retiring a Hopper data center because there's no demand for the chips above the price to run them—the input-cost OPEX on energy and people to maintain the data center—to justify it. For the bulk of the market, you can't replace a Hopper data center in the same footprint with Blackwell or Rubin without completely ripping things out or just knocking it down and building new, right?

Doug O'Loughlin

Yeah, so it's a variable cost.

Dylan Patel

That is a potential scenario if there's so much demand for the new stuff that people literally don't have power permits or land, and so they just knock down an old data center.

Doug O'Loughlin

In a frictionless world, that is true, but we live in a world with friction, and the friction is getting worse when it comes to new compute. Is that fair? So, I agree with you.

Dylan Patel

That's the first thing. Now, the second thing is the camp for GPU prices going down is that model progress stalls, roughly, and GPU prices go up if model progress continues, meaning demand continues. That's the rough thing.

Now, there's 1 X factor that I don't think is being considered, and I want your take on it: if there is government intervention on the frontier labs. To me, if that happens, GPU prices go down, especially for the old stuff.

Doug O'Loughlin

Explain why you think that.

Dylan Patel

Because they're going to restrict who can have access to the latest and greatest, and that's going to restrict how much demand there can be. Therefore, there will be more demand for alternative stuff. That's going to cause people to use the—

Doug O'Loughlin

There are too many what-ifs in that one. I agree, sort of, but that's such a top-down argument. In my heart of hearts, I'm a bottoms-up guy, meaning that the world changes one mind at a time until it becomes mass consensus—people choosing things from the bottom up, like product-led growth, right? I don't believe in CTOs. I do believe in CTOs. But see—

Dylan Patel

You don't believe in the great man theory of history, Doug?

Doug O'Loughlin

We've had this conversation. I'm not going to have this conversation.

Dylan Patel

Come on, now.

Doug O'Loughlin

I've had this one too many times, with too many what-ifs and maybes. I'm not going to have this conversation.

Dylan Patel

You're in a Jensen Huang shirt, not believing in the great man theory of history.

Doug O'Loughlin

He was just atoms pushed along by the greater Molochian will.

Dylan Patel

Right place, right time, eh?

Doug O'Loughlin

Yeah, a little bit. I mean, I'm going to be pragmatic. I feel like there are many variables. It's hard to say that the only variable was him, right? I'm sure there are some great guys who are like—In my experience, really great horses make the rider look better.

I will admit, Jensen's a hell of a rider, and he had a hell of a horse. So, it's very hard—anyways.

Dylan Patel

This also doesn't make sense to me, because you're implying that government intervention is a single person as opposed to a groundswell of minds, whereas I pretty fundamentally believe right now that a lot of people in the U.S. really don't like AI. And I think that is not being priced in.

Nate Silver

Not priced in. I agree. How this will be priced in, I think unfortunately, is the midterms. This is the vibe: we get a little bit of wind testing in about 2 months from now, and I think my guess is that it's not a top-5 priority. No, no, it's not a top-3 priority, but it is a top-5. Is that fair? Healthcare—

Dylan Patel

For most of the candidates?

Nate Silver

For most of the candidates, yeah.

Dylan Patel

Yeah.

Nate Silver

Everyone's going to talk about it, but no one is going to platform on it. Does that make sense? And I think that's where—okay, so usually what happens is, if no one platforms on it, that's where corporate interests win.

Dylan Patel

Yeah, and corporate interests are currently stopping, for example, the ROSA bill in the Senate, the Remote Access Security Act or something like that, which passed the House like 300 to 20 and yet is stuck in the Senate while people lobby against it.

Nate Silver

Yeah.

Dylan Patel

If you're saying that it's a fourth- or fifth-priority issue, what would move it to number 3, where it gets some attention?

Nate Silver

How the fuck do I know, man? I mean, you tell me, man. I'm pretty interested because, speaking of which, I think we literally just started commissioning today. We're going to do survey work specifically focused on this problem, because I think it's interesting. Personally, I hope you get a nice little survey in the mail and fill it out.

Dylan Patel

So, here's my view.

Nate Silver

Yeah, I mean, I haven't really thought this through, but to me, let's say that the concept of AI is not at the forefront of people's minds, but it's used as a scapegoat to excuse other things that people care about, like you mentioned healthcare, people's finances, or housing. A lot of people who have—

Dylan Patel

Climate change.

Nate Silver

Literally all of the stuff that people actually do care about. I think there's a big chance that AI and tech in general—or tech people, who a lot of people have backed Trump financially—will be scapegoated for other issues that people care about, like inflation and the economy.

Dylan Patel

100%. It is a whipping boy for what you actually care about. It's important, so you have to attach it.

Nate Silver

Yeah, actually, I think I disagree with you that cost of living is probably the number-one issue in a lot of people's minds, and that's what they're going to vote for.

Dylan Patel

And AI will be like a sub-sub-agent—

Nate Silver

A component, a sub-agent of the cost-of-living debate.

Dylan Patel

Yeah.

Nate Silver

It's not going to be like, "We should decide whether we sponsor AI and push it forward." It's going to be like, "I care about the economy, and you should blame tech bros and AI."

Dylan Patel

It’s probably going to be a big part of the midterms, I think.

Nate Silver

We’ll see.

Dylan Patel

Tech bros and AI.

Nate Silver

Okay, I agree, but I’m curious to see what is really going to resonate and stick. There’s also the horseshoe theory. You go super far on the right, you come to conservatism, right? It’s like, “Oh, my ranch is being affected by the EMF, and the noise from the data center is killing my calves on the farm next door.”

So I’m pretty curious about this one. I’m going to be honest with you: I don’t have a strong enough view to be like, “Yeah, this is definitely what will happen.” Usually, I’m like, “I’ll wait and see,” kind of guy, but we’re doing a lot more work on this.

The last part of what would actually kill it is a slowdown, the midterms, whatever, stuff becomes illegal, and we regress a little bit. But also, the thing I keep thinking about here is that the future for all these technology booms always comes true, but the timing of the cash flows is the issue, right?

You spend $1 trillion to get $100 billion, and it actually does become $1 trillion one day. But it comes 5 years later, and at that point you’re like, “Dude, I don’t have enough money to keep this thing going.” Does that make sense? I think that’s going to be the real issue: “Okay, you spent $2 trillion, and all of a sudden AI rips. AI’s at $300 billion, $400 billion.”

We’ll say we’re spending $5 trillion, and AI’s at $500 billion in revenue. And you’ll be like—

Dylan Patel

OpenAI and Anthropic believe the final pre-train is coming soon because they’re going to IPO and have the final funding round.

Nate Silver

Yes. And then, when that happens—the final, final pre-train happens—it is really good. It’s the best it’s ever been. It makes a ton of revenue. It grows very quickly, but it doesn’t grow at a rate that’s enough to pay the bills.

So you built a house that you cannot pay for. You’re paying for $5 trillion of investment on a $500 billion thing, and you’re like, “Wait, wait, that’s 10 years of spending.”

Dylan Patel

You’re saying this in spite of your pretty deep understanding of these companies’ financials and how profitable serving the existing models is right now.

Nate Silver

I don’t think we’re there yet. This is a chicken-and-egg thing, right? The thing that matters is how much. We don’t have enough. It isn’t enough. We’re still on the narrow path, I think, where you can kind of blink and see where the revenue comes from, and it probably is good enough to fund it.

You also have these hyperscalers who are super good for the money because they have these other businesses that gush cash and are among the most profitable in the history of time. But it just keeps hitting—you know, we talk about blindly scaling the supply-demand wall, right?

I think, legitimately, we’ll say $150 billion of ARR for the entire ecosystem, and $1 trillion of CAPEX so far. Maybe it’s not all in the ground, but $1 trillion has left the door and there’s a lot of it working. So, that’s a 15% return on revenue, but not on profit.

If we say it’s 50% profit, that’s, you know, 7.5%. That’s not the end of the world. Maybe that’s better than it costs people to do, but that’s not super profitable. So you have to believe that the $150 billion becomes $500 billion, which is doable.

That can probably get you to whatever, $2–3 trillion out the door. But then what happens is there’s going to be a doubling where it’s just really hard for it to double that quickly. That’s where I think the gap happens, where you’re like, “Okay, we just spent $5 trillion.” Yeah.

Dylan Patel

What do you think stops it from doubling? I think I’ve made the case that there would be enough demand from all the users and all of the different model types around the world, but there’s also—

Nate Silver

You have to get your grandma to be vibe coding.

Dylan Patel

I’ll give it a go. I think she’s up for it.

Nate Silver

It’s like, “Grandma, I need you to make 12 agents.”

Dylan Patel

I get a really incredible newsletter every month, man. She writes it all on her own, you know. I can totally imagine the research being some AI stuff in there, helping with the design.

Nate Silver

So I think that’s probably the biggest mismatch. I don’t have a strong view right now, and I think we’re nowhere near the point where it’s actually okay, because the path narrows as you have more revenue and more CAPEX. It becomes a tighter path to walk.

It is not a narrow path. There’s a lot of cushion right now. There are a lot of companies that make a crap ton of money. Meta makes a crap ton of money. Their free cash flow goes negative, but if they wanted to, they could stop CAPEX tomorrow and then the profit would print, right? That’s okay, right?

As you do more and more and more, you commit more and more and more. The stakes become higher and higher, and then the path becomes narrower. In that narrow path, you have to essentially demand people to use it. People have to be using it like yesterday, right?

I think the problem is that the decision-makers and the people who actually adopt it are two very different worlds. Zuck is sitting here like, “Of course you’re going to have your Meta sunglasses, and you’re going to have your Meta wearables, and you’re going to be in the whatchamacallit universe—in the metaverse. You’re going to be using a quadrillion tokens every single day.”

Then there’s a grandma in Nebraska who’s like, “Honey, I don’t really know how to get my new iPhone to work.” You have to win every single consumer tomorrow. I think the adoption curve just takes time.

The benefit of the adoption curve so far is that the internet is a really scalable distribution platform that fits AI super well. Most young people are pretty big adopters. Most working-age people are big adopters. I’m talking to people at funds who are in their 40s and 50s who use it every single day because they have to. It’s a really good technology.

But the question is: will every single person use it? And are they going to all be token-maxing? Because I think you have to believe that.

Dylan Patel

Yeah, and I—

Nate Silver

I do, and I think we both made the case on opposing sides here for a little bit. Maybe the more interesting question is on the supply side. The concept is being able to bring on enough GPUs or hire enough people in order to go and sell, because to me revenue probably doesn’t necessarily—I mean, consumer is such a small part of the revenue of these coding-agent companies in terms of ARR right now that I actually don’t think it depends on the grandmas.

I think it depends a lot on enterprise businesses, defense agencies, intelligence agencies, and all sorts of other federal government agencies around the world actually adopting this stuff at scale. I see a pathway to getting everybody and every bank and every telco and every retail company actually using this in their day job at work, as opposed to having every consumer have a subscription to this sort of stuff.

Maybe the more interesting thing is that you’re going to run out of gas on the ability to bring GPUs online, hire enough people in your B2C go-to-market sales motion, or have them install enough GPUs in their private data centers to actually run this stuff. You need to get to the point where you can actually grow revenue to that point where you’re saying, what, $100 billion ARR goes to $500 billion ARR—

Dylan Patel

Yeah.

Nate Silver

I mean, we’re hitting some physical constraints that are not just making the chips out of the factory. They’re relatively creative. We’re running out of electricians in the United States. We have a 100,000-person gap.

Every electrician who’s a journeyman—let’s say, I forgot the name of it. It’s essentially like a workman or whatever, a mid-level electrician—is able to make $250,000 a year pretty easily. If they really wanted to and worked 18-hour days, I’m sure they could make $400,000 or $500,000 a year.

These are people—it’s a trade. These guys are in demand, dog. There’s actually a really weird website that I go to. I’m not going to leak all of our alpha on this podcast, but there’s a really good website that shows open jobs for electricians. It’s crazy because you can use the Wayback Machine and see the hourly rate go from $15 or $20 an hour to $50, $100, or $200 an hour.

The reality is that it takes time. You don’t just wake up one day and become an electrician. It’s, let’s say, 18 months of training. Another doubling requires an entire— we’ve never trained that many electricians, right?

Another example is capital, which I think is probably the most top-of-mind one for the finance community: debt, right? Let’s say, so far year-to-date, the hyperscalers have raised roughly $450 billion of debt.

That’s effectively the biggest ever. It’s like the third-largest issuer after the United States government and China, with its hyperscalers in aggregate raising capital. The problem is that there’s a limited supply of money. Someone is buying the debt, right? When you issue a bond, someone buys it on the other side.

So, in order for people to buy more bonds, they have to give them a higher rate. You’re hitting a supply-and-demand curve there. On the demand side, I really don’t know what drives it. Well, I have a vibe as to what drives it. It’s life insurance, okay? Retiree assets.

There’s a weird poetic justice that the peak buying of annuities is right before retirement. All the boomers are in retirement, so the actual asset class is larger. But can it double? Can it triple? I don’t think so. That’s going to be one of the problems we start to see: in order for the hyperscalers to raise more debt, they have to offer higher interest rates.

When you’re a lender and you’re saying, “Hey, I can give someone money. I can either buy these mortgage-backed securities for an American, backed by the United States government, at, let’s say, a 5% rate for your mortgage”—I think that’s actually too low. Whatever, for your mortgage. Or I could lend to this hyperscaler, who’s a better lender than the United States government ’cause they make a ton of money and they pay me 7% or 8%.

So, if they keep doing that and all the money goes this way, essentially mortgage prices will start to increase. You have all these ways that the system just can’t handle another doubling. The scaling laws are scaling, and they’re like, “Great, let’s make a 2-times-bigger model.” But I don’t think everything can scale at a 2- or 3-times-bigger rate.

So electricians and capital are 2 of the ones that I think are creative and weird. But I think if you give it time, it will, though. They’re going to issue a ton of money. I’m sure they can issue $1 trillion next year. Truly.

Dylan Patel

I’m actually fascinated by that. I really had not thought it through, but it totally is pension plans that are funding this buildout right now.

Nate Silver

Yeah, but pensions, dude, pensions are a secular decliner. Pensions as a concept have actually gotten down over time. It’s not like people are working. People are working more in local governments than they used to be. Pensions are—also, most pensions are underwater.

Historically, the concept of a pension has shifted over to an equity-participation plan, like a 401(k), right? Pensions have lost share at the expense of 401(k)s. They are buying stocks in this stuff, but they’re not really issuing equity.

The really big dollars that get issued every single year are through debt. Life insurance is a great source, and life insurance is from annuities. Annuities get purchased when someone is about to go into retirement, so that’s a source. General insurance, actually—health insurance, stuff like that.

But you have to literally believe everyone just needs twice as much insurance. You’re like, “Dude, no one’s going to—” You know what I mean? It just doesn’t make sense to me.

Dylan Patel

It’s an interesting balance because there’s a lot of conversation in advanced economies that there’s an inversion in the age pyramid, as many people live longer, retire, and build up this wealth for retirement. But there’s a lot of money there.

Nate Silver

Fewer people—

Dylan Patel

Spend it on data centers. They can spend it on data centers, too. And then, importantly, the boomers who left such a bad legacy—

Nate Silver

Okay, boomers, right? You can downsize. Sell the house, right? Take that cash. You want to buy some life insurance policies, and those are going to fund the data center buildout. So we’re trading houses for data centers, right?

Dylan Patel

And even better, it takes away your job at the end, too. So, yeah, you can’t even own a house. Now you’re poor and penniless, and you’re in the permanent underclass because your boomer parents invested in life insurance and took away—

Nate Silver

But sorry, no—the life insurance is also funding the data centers, which are also building the AIs and the robots that are making sure you don’t have a job, either. So now you don’t have a house, you don’t have a job. Is it going to come for retail food service soon as well? Now you can’t eat.

Dylan Patel

Yeah, dude, that’s when the robots come, bro. No, I mean, that’s being a little trite, but, yeah, I don’t know. It’s pretty interesting. You just can’t say, like—okay, a good example. Another good example of this, actually—my favorite one that makes a lot of sense—is Taiwan.

We can double the output of Taiwan over and over and over. There are economies of scale, but Taiwan’s GDP is up like 25% this year just because TSMC’s cooking chips, right? But if TSMC were to double again, and let’s say there’s some ratio of workers that needs to happen, you’re going to run out of people in Taiwan to make the chips. I’m not joking.

It employs almost directly and indirectly, I think, 20% of the economy. The rest of it is healthcare, retail, and the government. There’s really only one game in town. So, let’s say they double—

Supporting TSMC, though, indirectly, right?

Nate Silver

Yeah, it’s all for supporting TSMC indirectly, yeah. So, let’s say TSMC doubles—triples, triples, okay? Triples the need for workers. What are we going to do? Taiwan needs to literally have more babies to be able to pay for that future. So, yeah, it’s kind of crazy if you think about it.

Dylan Patel

Yeah, I did just look it up. To be clear, though, I think you’re saying “indirectly”—I think that’s doing a lot of work. TSMC has fewer than 100,000 employees, and Taiwan has over 20 million people. So they can double it one more time and find some people to employ directly.

Nate Silver

They can do that. I’m also knocking on the factories that support the factories that support the factories—the material services, all the stuff like that.

Dylan Patel

Yeah, I mean, at the percentage of GDP, they can build a new one in Arizona, right? And expand other things globally. They’re—

Nate Silver

No electricians in Arizona, bro.

Running out of labor.

Dylan Patel

I think it’s the people that operate the machines—lithography machines, anyway.

Nate Silver

There’s actually a crazy thing. To make a new fab, you need an electrician. People are flying electricians via Cessna to little backwater places in order to make it to the job sites and stuff. There are private flights just filled with 16 electricians, like, “Ah, you’re here for your second shift.”

Dylan Patel

So, okay, this really reminds me of the oil fields in Canada. I have some friends who have worked on different oil projects in Fort Mac and stuff like that.

Nate Silver

Yeah.

Dylan Patel

The boom-bust cycle of oil in Calgary and northern Alberta and stuff is really fascinating, as some of these communities get destroyed when oil mining, whatever else, is shut down. It’s going to be fascinating to see how that happens when it’s not even a natural resource that’s drawing people to these areas. It’s literally just a permitting—

Nate Silver

An information factory.

Dylan Patel

Yeah. No, but the reason why they have these data centers here is because of the ability to produce energy, which—obviously, being close to the energy sources is good—but generally, they put up data centers where they can get a permit most easily. That’s where they put the—

Nate Silver

West Texas.

Dylan Patel

Yeah, exactly.

It’s not even like there’s something fundamental where you can just run out of lithium to mine, or gold, and then all of a sudden everybody has to turn around and go home. It’s literally just like, if a permitting regime changes, if the political will of the people changes a little bit, then all of a sudden everybody in there is out of a job and things change a lot.

I think we saw exactly this when we were putting out stuff on New Mexico. We had Ali on a couple weeks ago to talk about that with Jeremy. And the whole—

Nate Silver

But wait, the thing is, okay, the invisible hand of capitalism does work a little bit. There’s also this interesting aspect: if they do this over and over and over, the logical conclusion is that all these new jobs just get disappeared, and then some guy is going to be like, “Whoa, they took our data centers. They took our jobs.” So there is a self-correcting force in this as well.

This is an interesting poll that I saw the other day: people who hate data centers often don’t live near one, while after having a data center in your community, especially for young people, it’s a net favorable thing because of jobs.

Dylan Patel

Yeah, I mean, I went to visit one in the Buffalo area, and just like everybody we met—I mean, of course they like having a job—but everybody was just super pro the whole site and how many people it employs.

Doug O’Laughlin

’Cause they’re like, “F yeah, dude. I got a job, dog.” A construction job is good. It doesn’t take super high-skilled labor. You don’t have to have a PhD. You know what I mean? It’s just like, “No, dude. It’s an honest living, okay?” People are making more money than they have been because it’s clearly in demand. I mean, it’s just supply and demand: when there’s a lot of demand, supply’s got to reach it.

And so I think it’s a net good thing if it’s broad. I think, ironically, because these data centers are being built in the middle of nowhere, it’s actually very good for broad participation, right? You can argue the entire finance-industrial complex and the technology-industrial complex are extremely narrow.

It’s like, “Yeah, dude, San Francisco tech is making me an app so I can deliver something in a rural state.” That is not broad participation in the economy. But building a $10 billion building in the middle of nowhere that has extremely high electricity needs, high levels of service, and is extremely important—people are willing to pay for overages because it’s extremely important for the strategy of these companies.

That’s broad participation, ’cause a thousand people—I don’t know what the math is. I feel like Jeremy knows. I want to say it’s like 10,000 people per gigawatt or something like that. And so it’s like, “Hey, if you bring 70 gigawatts, it’s a lot of jobs. It’s 700,000 jobs.” And that starts to move the needle, dude.

Dylan Patel

In terms of revenue, right? There are many other jobs programs where you can spend less money and create a lot more jobs, but a lot of them are not sustainable in the same way, whereas this is obviously going to run for a very long time. People really like these data centers. And I think that, generally speaking, a lot of people have problems with jobs programs when compared to investment from companies that happens to be directed to these communities and happens to result in jobs and investment.

Doug O’Laughlin

It’s more sustainable, so.

Dylan Patel

Yeah, it’s good stuff. Okay, we’ve got to move to a wrap here.

Doug O’Laughlin

Okay, yeah. I was going to say, I have a call I’m late for. It’s the reason why I’m not on this podcast, bro.

Dylan Patel

Good to have Doug on. Good to debate a little bit. Yeah, good to get some hot takes.

Doug O’Laughlin

Well, some modest takes, I think.

Dylan Patel

Oh, we should get Joey on. I really want to do a Finance Week earnings recap.

Doug O’Laughlin

Okay. We’ll have to get Joey on.

Dylan Patel

We can shoot the breeze. Or we’ll have to get Joey a competitively dynamic shirt like your Jensen Huang lightning leather jacket shirt.

We have a Morris Chang shirt, so.

Doug O’Laughlin

Okay, Joey can get a Morris Chang for that.

Dylan Patel

We can get a Lisa Su in here, too. That’d be sick. We actually have new company uniforms, so you’re looking at them. Okay, sweet. Thanks for coming on, man. Good job.

Doug O’Laughlin

Yeah.

Ep. 022 - Market Drawdown, Historic Bubbles, Funding The Buildout, AI Politics (Doug is Back) | BidClub