Original compilation: DeepChain TechFlow
Guest: Doug O'Loughlin, Analyst at SemiAnalysis (former founder of Fabricated Knowledge)
Host: Dylan Patel, Founder of SemiAnalysis
Podcast source: SemiAnalysis Weekly
Ep. 022 - Market Drawdown, Historic Bubbles, Funding the Buildout, AI Politics (Doug Is Back)
Air date: July 29, 2026
Disclosure: Doug O'Loughlin and Dylan Patel are both employees of SemiAnalysis, a paid research firm in the semiconductor industry whose business model depends on industry conditions. The content below faithfully reflects the original conversation and does not constitute investment advice.
Key Points Summary
Doug O'Loughlin returns to SemiAnalysis Weekly after a long absence, following a sharp correction in the semiconductor sector after its "best first half on record." The Korean KOSPI dropped 40%, retail investors with 2x leverage were liquidated, and SK Hynix missed expectations due to slowing price gains from its shift to LTA. Doug draws parallels between the current situation and the 1980s Taiwan bubble, noting that the behavioral patterns are highly similar, but fundamentals remain sound.
Two individuals engaged in a heated debate over “how large AI demand really is.” Dylan drew from SemiAnalysis’s own experience: after deploying coding agents, the company’s AI spending increased 100-fold, user count grew from 9 to 90, and per-user usage rose tenfold. Doug did not deny the strength of demand but raised a core concern: while supply can be quantified, demand is a “trillion-dollar question” with no clear answers. More critically, scaling laws require chip capacity to double, but physical and institutional bottlenecks—such as electricians, capital, and permits—cannot scale in tandem. This year, hyperscale cloud providers have issued $450 billion in bonds, funded by pensions and annuities, yet the pension pool itself is shrinking.
Summary of Key Insights
Regarding market pullback
By the end of Q2, this was the best performance in semiconductor history. Then we started paying it back. The faster something rises, the stronger gravity pulls it down.
Koreans have a 20-year track record: they always buy at the top. They bought banks in 2007, SaaS in 2021, and this time they’ve YOLO’d themselves.
The KOSPI dropped 40%, wiping out those with 2x leverage. Then came a self-fulfilling spiral: everyone saw their accounts shrink and decided to sell, further accelerating the decline.
Regarding memory cycles
SK Hynix has shifted toward more LTAs, with price increases slowing from threefold to 30–50%. The financial crowd has lost its mind, focusing only on rates of change. As the second derivative drops, they assume the cycle is over.
The script for semiconductors is always the same: during shortages, everyone places double orders; factories, seeing the surge in demand, ramp up production疯狂扩产. Then, when demand sneezes, supply is still ramping up, utilization drops from 100% to 50%, and price cuts become inevitable.
Regarding AI requirements
The demand curve is the trillion-dollar question. The supply curve is relatively easy to understand, but no one knows whether demand is ten times or a hundred times greater.
SemiAnalysis itself is a case study: after deploying the coding agent, the number of users grew from 9 technical users to 90 full-time users, with each user’s token usage increasing tenfold—and the company’s AI spending increased 100-fold.
Regarding supply chain bottlenecks
The U.S. is short 100,000 electricians. Mid-level electricians earn $250,000 annually, and those willing to work overtime can make $400,000 to $500,000. Some companies even use Cessna aircraft to transport electricians to remote job sites.
This year, hyperscale cloud providers issued $450 billion in debt, second only to the borrowing levels of the U.S. government and China. This money comes from pension and annuity funds, but pension pools will not double.
TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If that were to double, Taiwan would need to have more children just to have enough workers.
About AI Politics
AI is less popular than ice cream, less popular than politicians. This has not been priced in. In the midterm elections, AI will become a scapegoat for the cost-of-living issue.
The ROSA bill passed the House 300 to 20 but is stalled in the Senate. Corporate lobbying efforts are blocking legislation that would restrict China’s remote access to GPUs.
Body
The best first half in semiconductor history, followed by debt repayment.
Dylan: The stock market is pulling back, and all the AI names are falling. Today, we’re either adding fuel to the fire or offering some comfort.
Doug: By the end of Q2 on June 30, this was likely the best performance in semiconductor history. Then we started to unwind. Much of this can be attributed to technical factors: leverage and momentum reversal. But the reality is, the faster something rises, the stronger gravity pulls it back. We’re paying for the previous wild momentum rally.
The situation in Korea is insane. Every day, stocks are hitting their daily trading limits downward. One tweet asked, “How do I do my job?” The HR director lost all his money, and everyone is depressed because all the stocks have crashed. If you look back at the history of Asian financial markets, events like this happen more frequently than you might think.
My favorite book is about the Taiwan bubble. Taiwan experienced a 100-fold bubble on a per capita basis, with bank transactions at 500 times price-to-earnings ratios—everything went crazy.
Dylan: When is this?
Doug: Late 1980s.
Dylan: Do you think South Korea's fundamentals are different now compared to then?
Doug: The fundamentals are strong. But the issue is that things are never as bad as fear makes them seem, nor as good as you imagine. SK Hynix missed expectations today because they shifted toward more LTAs. Ironically, during their ADR roadshow, they were complaining that Micron was securing lower prices with LTAs.
Memory prices rose about threefold last year; they won’t triple again next year—perhaps rising only 30% to 50%. But the financial world has lost its mind, focusing solely on rates of change. Historically, in memory cycles, when the second derivative turns negative, it’s usually the end. Because the rate of change doesn’t stall at 30%—it plunges straight to -50%.
The script for this cycle is always the same: everyone invests in building factories, capacity comes online, and then they realize, “Oh my god, why is demand so low?” Previously, there was double and even triple ordering. Factory utilization drops from 100% to 50%, and the only way to break even is to slash prices. That’s the essence of the semiconductor market.
The KOSPI has now dropped 40%. Those with 2x leverage have been completely wiped out. Then comes a self-fulfilling spiral: everyone sees their account value shrinking and decides to sell, further accelerating the decline.
Chinese memory: May spoil the party, but demand still exceeds supply
Dylan: Recently, Chinese memory companies have entered the ecosystem, with CXMT and YMTC launching major IPOs—what are your thoughts?
Doug: They have the capacity, and even low yields don’t matter. Chinese companies aren’t competing on profit margins or EPS.
CXMT is now clearly the fourth-largest player in the market, yet it’s still profitable despite the supply shortage. Apple has already started using CXMT’s memory because Micron is engaging in “price gouging.” No one’s crying at the casino, Tim Apple—you have to buy at market price.
CXMT might ruin the party, but the reality is that demand still outstrips supply. The real trillion-dollar question is: Where is the demand? The supply curve is relatively easy to understand. We don’t know the demand curve. We know that coding agents and chatbots mean more demand, but we don’t know if it’s 10 times or 100 times more. Supply will ramp up blindly until one day it hits the demand curve.
The coding agent is a turning point: SemiAnalysis's own 100x AI spending
Dylan: I think demand is clearly very strong and will remain so for a long time. Just looking at internal usage within my own company is enough. If you believe demand will flatten or decline in the future, you’d have to believe the models won’t improve anymore. I see no signs of stagnation—only signals pointing in the opposite direction.
Doug: Let me play devil’s advocate. What’s the strongest bearish argument? The pace of technological advancement might outstrip people’s adoption of it. Suppose the killer app for AI is data entry—Kimi K3 is already sufficient. We keep building faster cars and better products, but the actual demand curve is already satisfied by a product we already have.
It was like the dot-com bubble: back then, people said demand doubled every 90 days, but fiber optic technology improved by 2 to 3 times per year. Eventually, the performance of a single fiber became 500,000 times better, and then everyone said, "Wait, we probably don't need that much fiber."
Dylan: I don’t agree, but it’s worth discussing. My counterpoint is: there are 100 to 1,000 times more people who aren’t using any models at all. Second, AI’s use cases extend far beyond coding—it can also be used for video generation, drug discovery, and materials science. Someone is using AI to develop superconducting components—how much is that worth? It’s worth many GPUs.
Moreover, coding itself is not just about “centering a div.” It represents an entire category of tasks with economic value far exceeding frontend debugging. Sam Altman has been discussing RSI (recursive self-improvement), and Anthropic is preparing to release a new model. The coding agent in Claude 4.5 was a clear inflection point: you cross a line of intelligence, and an entirely new market emerges. What you couldn’t do the day before, you can do the next day.
Doug: You were the prototype user. This time last year, fewer than 10 people on the SemiAnalysis engineering team were using the coding agent, and you told Dylan, “Everyone in the company needs to learn how to use this.” Now we have 90 users.
Dylan: From 9 to 90, a 10x increase. Then, within three to four months, individual usage also increased by roughly 10x, and the company’s AI spending increased 100x. The question now is: Will every company do this? Probably not at our intensity, but many companies have substantial work that can be cut.
The H100 won't become scrap metal, but models are getting larger.
Doug: I think older chips will become worthless. Everyone says, “The H100 is an appreciating asset,” but one day, running a model will require 100 H100s. At that point, you’ll say, “Retire the old girl and get a B300.” The real confirmation signal will be a pricing divergence between the B200 and B300.
Dylan: I completely disagree. The fundamental reason is that no one will take out an H100 and replace it with a B300. Data center designs are entirely different. You can’t simply swap Hopper for Blackwell or Rubin in the same facility—you’d have to tear everything down and rebuild it. So, to justify retiring an entire Hopper data center, you first need to prove that the revenue generated by those chips is already below their operating costs. This isn’t a variable cost—it’s a sunk cost.
Doug: You're right in a frictionless world, but the world we live in is becoming increasingly friction-heavy. The friction in building new computing power includes electricity permits, land acquisition, and approvals.
Dylan: Yes, I agree. The scenario where GPU prices fall corresponds to stagnation in model progress, while the scenario where they rise corresponds to continued model progress. There’s also an X factor: government intervention in frontier labs. If access to the latest and best chips is restricted, demand will be compressed, and prices for older chips will fall as well.
Capital and Electricity: The Physical Limits of Scaling Laws
Doug: What worries me most isn't demand, but the physical bottlenecks on the supply side. First, electricians. The U.S. is short 100,000 electricians. Mid-level electricians earn $250,000 annually; those willing to work 18-hour shifts can make $400,000 to $500,000. A website tracks electrician job postings—on the Wayback Machine, you can see hourly rates rising from $15–$20 to $50, $100, even $200. It takes 18 months to train one electrician. And we’ve never trained anywhere near the number needed if we were to double that figure.
The second is capital. This year, hyperscale cloud providers issued approximately $450 billion in debt—the largest amount in history, second only to the U.S. and Chinese governments. Someone has to buy these bonds. To get them to buy more, higher interest rates are needed. Much of this capital comes from pensions and annuities. Pension funds are structurally shrinking, as retirement savings have shifted heavily into 401(k)s, which do not purchase bonds. So you’re essentially asking everyone to need twice as much insurance—but that doesn’t make sense.
Scaling laws say, "Great, let's make the model twice as large." But not everything can be scaled up two or three times in sync.
Dylan: Wait, you're saying pensions are paying for data center construction?
Doug: Yes. Annuities are primarily purchased before retirement, and as the baby boomer generation retires, this asset pool is large. But can it double? Can it triple? I don’t think so. Life insurance is another source, but you’d have to believe everyone needs twice as much coverage—no one buys twice the amount of life insurance.
Dylan: That's interesting. Pensions are structurally shrinking, but there's still a lot of money there.
Doug: Here’s another example—Taiwan. TSMC directly and indirectly accounts for 20% of Taiwan’s GDP. If TSMC were to double or triple again, Taiwan would need to have many more children just to produce enough workers. Taiwan is playing only one game. Taiwan’s GDP grew by 25% this year solely because TSMC is manufacturing chips. But if it doubles again, there simply won’t be enough people.
AI Politicization: The Scapegoat of the Midterm Elections
Dylan: Many people dislike AI, and this hasn't been priced in. How could it be priced in? I think it's the midterm elections.
Doug: AI is probably fifth on the priority list, not in the top three. Healthcare and cost of living come first. No one would run for office on an AI platform.
Dylan: But AI will become a sub-issue of the cost-of-living debate—not “whether we support AI,” but “care about the economy, blame tech bros and AI.” The ROSA Bill passed the House 300 to 20 but is stuck in the Senate. Corporate lobbying is blocking it.
Doug: If it's not a top-three priority, lobbying power will override public opinion.
Dylan: But AI has already become a scapegoat for people on other issues—climate change, housing, inflation—AI and tech bros always get dragged into it.
Doug: There’s an interesting poll: people who dislike data centers usually don’t live near them. Those who do live nearby, especially younger people, tend to have a positive attitude because of the jobs. I visited a data center near Buffalo, and the locals were overwhelmingly supportive. Building data centers in remote areas is actually beneficial—it broadens economic participation. A one-gigawatt data center requires about 10,000 people. Seventy gigawatts means 700,000 jobs. This is starting to influence votes.
Final outcome: $5 trillion in investment, $50 billion in revenue
Doug: The future of technological boom will always materialize—it’s just a question of cash flow timing. You spend a trillion, get back a hundred billion, and yes, it will eventually become a trillion again. But it might take five years, and by then you’ll say, “Dude, I’m out of money.”
Assuming the current ARR of the entire AI ecosystem is $150 billion, with cumulative CAPEX of $1 trillion, a 15% revenue return translates to a 7.5% profit margin at a 50% profit rate. That’s not bad, but it’s not exceptionally profitable. You need to believe that $150 billion can grow to $500 billion—which is achievable. Then, $500 billion could support $2 to $3 trillion in CAPEX. But doubling again beyond that would be very difficult.
OpenAI and Anthropic believe the final pretraining is imminent because they are going public. The models produced by final pretraining are indeed excellent, and revenue is growing rapidly—but not fast enough to cover expenses. You’ve built a house you can’t afford. You’ve invested five trillion, earning five billion in revenue—that’s ten years’ worth of income.
Dylan: You said this is revenue, not profit. And when you said that, you knew how high the profit margins are for these companies right now.
Doug: Yes, we haven’t reached that point yet. We’re still on a narrow path, where you can see how revenues align. Hyperscale cloud providers have other businesses generating massive cash flows—if they wanted to cut CAPEX, their profits would instantly skyrocket. But as you invest more, the stakes rise and the path narrows further. At some point, you actually need everyone to be using it. The problem is that decision-makers and actual users inhabit entirely different worlds. Zuckerberg thinks everyone will be wearing Meta glasses and burning trillions of tokens in the metaverse every day, but a grandmother in Nebraska can’t even use a new iPhone.
Dylan: Income doesn't come from grandmothers—it comes from enterprises, banks, telecommunications companies, retailers, defense, and intelligence agencies. I see every bank, every telecom company, and every retailer using this in their daily operations. The more interesting constraints are on the supply side: Can you install enough GPUs? Can you hire enough people to sell them?
Doug: Yes, the supply-side issues are more interesting and more difficult. Electricians, capital, permits—these things can’t be doubled according to scaling laws. But given time, they will come. They may indeed issue a trillion dollars in bonds next year. The real problem is that the path narrows, the stakes rise, and you require everyone to adopt. This adoption curve takes time.