Organization & Compilation: Deep Tide TechFlow

Guest: Doug O'Loughlin, SemiAnalysis Analyst (Former Founder of Fabricated Knowledge)

Host: Dylan Patel, Founder of SemiAnalysis

Podcast Source: SemiAnalysis Weekly

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

Release Date: July 29, 2026

Disclosure: Both Doug O'Loughlin and Dylan Patel are employees of SemiAnalysis, which is a paid research institution for the semiconductor industry, and its business model relies on industry prosperity. The following content faithfully presents the original dialogue and does not constitute investment advice.

Summary of Key Points

Doug O'Loughlin makes a long-awaited return to SemiAnalysis Weekly, coinciding with a sharp drawdown in the semiconductor sector following its "best first half ever." The Korean KOSPI fell 40%, retail investors lost their 2x leverage, and SK Hynix missed expectations due to a shift towards LTA resulting in slowed price increases. Doug compares the current situation to the 1980s Taiwan bubble, suggesting that the behavior patterns of the bubbles are highly similar, yet the fundamentals remain healthy.

The two engage in a heated debate over "how big is the AI demand?" Dylan draws from SemiAnalysis's own experience: since the coding assistant launched, the company’s AI spending skyrocketed by 100 times, users grew from 9 to 90, and individual usage also increased 10 times. Doug acknowledges the strong demand but raises a core concern: supply can be easily calculated, but the demand side is a "trillion-dollar question" without clear answers. More critically, scaling laws require chips to double, but physical and institutional bottlenecks such as electricians, capital, and licensing cannot double in sync. This year, hyper-scale cloud vendors have issued $450 billion in bonds, with funding coming from pensions and annuities, even as the pension pool itself is shrinking.

Highlights of Key Perspectives

On Market Drawdown

"By the end of Q2, this is the best performance in semiconductor history. Then we started to pay off debts. The faster something rises, the greater the gravitational pull."

"Koreans have a 20-year record: they always buy at the top. They bought banks in 2007, SaaS in 2021, and this time they YOLO'd themselves."

"The KOSPI fell 40%, and those with 2x leverage went completely to zero. It becomes a self-fulfilling spiral: everyone sees their accounts shrink, decides to sell, which intensifies the decline."

On Memory Cycles

"SK Hynix shifted towards more LTA, causing price increases to slow from 3x to 30 to 50%. The finance sector has all gone haywire, only looking at the rate of change. When the second derivative drops, they think the cycle is over."

"The semiconductor script is always the same: when there's a shortage, everyone doubles orders, factories see demand and frantically expand production. Then demand sneezes while supply ramps, utilization drops from 100% to 50%, and all they can do is cut prices."

On AI Demand

"The demand curve is the trillion-dollar question. The supply curve is relatively easy to understand, but whether demand is 10 times or 100 times, no one knows."

"SemiAnalysis itself is a case in point: after launching the coding assistant, the number of technical users grew from 9 to 90, with each person's token usage also increasing 10 times. Company AI spending increased 100 times."

On Supply Chain Bottlenecks

"The U.S. lacks 100,000 electricians. Mid-level electricians earn $250,000 a year, and those willing to work overtime can make $400,000 or $500,000. Some are using Cessna planes to transport electricians to remote sites."

"Hyper-scale cloud vendors this year issued $450 billion in debt, second only to U.S. and Chinese borrowing. This money comes from pensions and annuities, but the pension pool will not double."

"TSMC directly and indirectly constitutes 20% of Taiwan's GDP. If TSMC doubles or triples, Taiwan needs to produce more children to provide enough workers."

On 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 be the scapegoat for living costs."

"The ROSA bill passed the House 300 to 20 but is stalled in the Senate. Corporate lobbying power is preventing legislation that would restrict Chinese remote access to GPUs."

Body

The Best First Half in Semiconductor History, Then Started Paying Off Debts

Dylan: The stock market is retracing, and all AI names are falling. Today we either fan the flames or provide some comfort to everyone.

Doug: By June 30, the end of Q2, this is probably the best performance in semiconductor history. Then we started to unwind. Much of this can be attributed to technical factors: leverage, momentum reversal. But the reality is, the faster something rises, the greater the gravitational pull. We're paying for the previous crazy momentum rally.

Conditions in Korea are insane. Every day there are stocks hitting the limit down. There was a tweet saying, "How do I do my job?" The HR director lost all his money, and everyone is depressed because all stocks are falling. If you look back at the history of Asian financial markets, such events occur more frequently than you think.

One of my favorite books talks about the great bubble in Taiwan. Taiwan had a 100-fold bubble on a per capita basis, with banks trading at a 500-times P/E ratio, everything was going crazy.

Dylan: When was this?

Doug: In the late 1980s.

Dylan: Do you think the fundamentals in Korea now differ from that time?

Doug: The fundamentals are good. But the problem is, things are never as bad as the fear, nor as good as you imagine. SK Hynix missed expectations today because they turned towards more LTA. Ironically, they were complaining during their ADR roadshow about Micron getting lower prices for LTA.

The memory prices rose about 3 times last year; there’s no way they can rise 3 times again next year, likely around 30 to 50%. But the finance sector has all gone haywire, only focusing on the rate of change. In the history of memory cycles, a drop in the second derivative usually signifies the end. Because the rate of change doesn’t stop at 30%, it can drop straight to negative 50%.

This cycle's script is always the same: everyone invests in building factories, the capacity comes online, then they find, "Oh my god, why is demand so low?" Because previously there were double orders, triple orders. The factories go from 100% utilization down to 50%, and the only way to break even is to cut prices. That’s the essence of the semiconductor market.

The KOSPI is now down 40%. Those with 2x leverage went completely to zero. Then it becomes a self-fulfilling spiral: everyone sees their accounts shrink, decides to sell, which intensifies the decline.

Chinese Memory: They Might Ruin the Party, But Demand Still Exceeds Supply

Dylan: Recently, Chinese memory entered the ecosystem, with CXMT and YMTC going public. What are your thoughts?

Doug: Historically, China has cheapened whatever they touch. They have the capacity; even with low yields, it doesn’t matter. Chinese companies are not competing on profit margins or EPS; the shareholders are the government, which incentivizes production, and provinces compete with each other for GDP.

CXMT is now clearly the fourth in the market, but this is a shortage environment, and they can still make money. Apple has started using CXMT's memory because Micron is engaging in "price gouging." Nobody cries at the casino, Tim Apple. You have to buy at market price.

CXMT may ruin the party, but the reality is that demand still exceeds supply. The real trillion-dollar question is: where is the demand? The supply curve is relatively easy to understand. We don't know about the demand curve. We know that coding assistants and chatbots mean more demand, but we don’t know if it’s 10 times or 100 times. Supply will ramp blindly until one day it bumps into the demand curve.

Coding Assistants as the Inflection Point: SemiAnalysis's Own 100 Times AI Spending

Dylan: I feel the demand is clearly very strong and will continue for a long time. Just looking at my company's internal usage is enough. If you think future demand will plateau or even decline, you have to believe the model will not improve anymore. I see no signs of stagnation, only signals in the opposite direction.

Doug: Let me be the devil's advocate. What is the biggest short argument? The pace of technological advancement may outstrip people's ability to use it. Suppose the killer application of AI is data entry; Kimi K3 is sufficient. We make faster cars and better products, but the real demand curve is satisfied by a product we already have mastered.

It’s like the internet bubble: back then, they said, "Demand doubles every 90 days," but fiber optics improved by 2 to 3 times each year. The last fiber's performance became 500,000 times that of the original, and then everyone said, "Wait, it seems we don’t need that much fiber."

Dylan: I disagree, but it’s worth discussing. My counter is: there are 100 to 1000 times more people currently not using any models. Secondly, AI use cases are far beyond coding. It can also do video generation, drug discovery, materials science. Some are using AI to create superconducting components, which are worth quite a bit. Worth a lot of GPUs.

Moreover, coding itself is not just "center a div." It represents a whole class of economic value far exceeding the tasks of front-end debugging. Sam Altman talks about RSI (recursive self-improvement), and Anthropic has new models coming out. The coding assistant in the Claude 4.5 version is a clear inflection point: you cross a certain line of intelligence, a whole new market emerges. What you couldn’t do yesterday, you can do the next day.

Doug: You are a prototype user. Last year at this time, SemiAnalysis’s tech team had fewer than 10 people using coding assistants, and then you told Dylan, "Everyone in the company must learn to use this." Now we have 90 users.

Dylan: From 9 to 90, 10 times. Then in about 3 to 4 months, each person's usage is also about 10 times. Company AI spending is 100 times. Now the question is, will every company do this? Maybe not at this intensity, but many companies have a lot of work they can cut.

The H100 Won’t Become Scrap Metal, But Models Are Getting Bigger

Doug: I think old chips will become worthless. Everyone says, "The H100 is an appreciating asset," but one day, reasoning a model will require 100 H100s. At that time, you’ll say, "Let's retire the old lady, buy a B300." The real confirmation signal will be pricing differentiation between B200 and B300.

Dylan: I totally disagree. The fundamental reason is: no one will tear down an H100 to replace it with a B300. The design of data centers is completely different. You can't swap a Hopper for a Blackwell or Rubin in the same rack; you must tear the whole thing down and rebuild it. So to justify retiring an entire Hopper data center, you first have to prove that those chips' revenues have fallen below operating costs. This isn't variable costs; it’s sunk costs.

Doug: In a frictionless world, you’re right, but we live in a world where friction is growing. The friction of building new computing power includes electric permits, land, and approvals.

Dylan: Right, I agree. The scenario of GPU prices dropping is stalling model progress; the scenario of prices rising is continuously progressing models. There's also an X factor: government intervention in cutting-edge labs. If they restrict who can use the latest and greatest chips, demand will be compressed, and the prices of old chips will fall as well.

Capital and Electricians: Physical Ceilings of Scaling Laws

Doug: What worries me most isn’t demand but the physical bottlenecks on the supply side. The first is electricians. The U.S. is short 100,000 electricians. Mid-level electricians earn $250,000 a year, and those willing to work 18 hours can make $400,000 to $500,000. There’s a site tracking electrician hiring where you can see hourly wages rising from $15-20 to $50, $100, or $200. Training an electrician takes 18 months. We’ve never trained that many to double the requirement.

The second factor is capital. Hyper-scale cloud vendors have issued about $450 billion in debt this year, the largest ever recorded. This is second only to government borrowing from the U.S. and China. Someone has to buy that debt. To get them to buy more, you have to offer higher interest rates. And largely, the source of this money is pensions and annuities. The pension pool is structurally shrinking. Many pensions have shifted to 401k plans, which don’t buy bonds. So you basically have to believe that everyone needs double the insurance, but that doesn’t make sense.

Scaling laws say, "Great, we can make the model twice as big." But not everything can be doubled or tripled in sync.

Dylan: Wait, are you saying pensions are paying for data center construction?

Doug: Yes. Annuities accumulate before retirement, and the baby boomer generation is retiring, so this asset pool is relatively large. But can it double? Can it triple? I don’t think so. Life insurance is also a source. But you have to believe everyone needs double the life insurance. No one will buy double life insurance.

Dylan: This is interesting. Pensions are structurally shrinking, but indeed there's a lot of money sitting there.

Doug: Another example is Taiwan. TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If TSMC doubles or triples again, Taiwan needs to produce more children to provide enough workers. Taiwan only has one game to play. Taiwan's GDP rose 25% this year, thanks solely to TSMC frying chips. But if they double, there won't be enough people.

AI Becoming Political: A Scapegoat for Midterm Elections

Dylan: Many people dislike AI, and this hasn't been priced. How could it be priced? I think it's the midterm elections.

Doug: AI is probably the fifth priority, not in the top three. Healthcare and living costs are higher priorities. No one is running on an AI platform.

Dylan: But AI will become a secondary issue for living costs. It’s not about "do we support AI," but rather "care about the economy, blame the tech bros and AI." The ROSA bill passed the House 300 to 20 but is stalled in the Senate. Corporate lobbying is blocking it.

Doug: If it's not a top three priority, the lobbying power will win over public opinion.

Dylan: But AI has already become a scapegoat on other issues. Climate change, housing, inflation; AI and tech bros will be dragged out.

Doug: There’s an interesting poll: people who dislike data centers typically don't live near them. Those who do live nearby, especially younger people, have positive attitudes because of jobs. I visited a data center near Buffalo, and locals were super supportive. Building data centers in remote areas is good; it broadens economic participation. A one-gigawatt data center needs about ten thousand people. Seventy gigawatts equals 700,000 jobs. This begins to impact voting.

Conclusion: Five Trillion Investment, Fifty Billion Income

Doug: The future of a technological boom will always materialize; the question is the timing of cash flow. You invest one trillion and get back one hundred billion; it truly can become one trillion one day. But it might be five years later, and by then you say, "Dude, I'm out of money."

Assuming the entire AI ecosystem currently has an ARR of $150 billion and a cumulative CAPEX of one trillion. A 15% return on income, calculated at a 50% profit margin, is about 7.5%. It’s not bad, but it’s not super profitable either. You have to believe $150 billion can grow to $500 billion, which can be done. Then $500 billion can sustain two to three trillion of CAPEX. But doubling it again becomes difficult.

OpenAI and Anthropic believe that eventually pre-training will come along because they want to IPO. The models produced from pre-training are indeed excellent, and revenue growth is rapid, but the growth rate isn't enough to pay the bills. You’ve built a house you can’t afford. You’ve spent five trillion, and earned five hundred billion; that’s ten years' worth of money.

Dylan: You’re saying that’s income, not profit. And when you’re saying this, you know how high the margins are on the services these companies provide now.

Doug: Yes, we’re not there yet. We are still on a narrow path, looking at how the revenue matches up. Hyper-scale cloud vendors make cash from other businesses; if they want to halt CAPEX, profits can be printed immediately. But as you invest more, the stakes get higher, and the path gets narrower. At some point, you actually have to demand that everyone is using it. The problem lies in the decision-makers and actual adopters being in two completely different worlds. Zuck thinks everyone will wear Meta glasses in the metaverse and burn trillion tokens daily, while a grandma in Nebraska can’t even use a new iPhone.

Dylan: Income isn't reliant on grandma. It's based on enterprises, banks, telecoms, retail companies, defense, and intelligence agencies. I see every bank, every telecom company, every retailer using this in their daily work. The more interesting constraints are on the supply side: can you equip enough GPUs, can you hire enough people to sell.

Doug: Yes, the supply-side issues are more interesting and more complex. Electricians, capital, licensing—these cannot be doubled according to scaling laws. But give them time, it will come. They might indeed issue one trillion in bonds next year. The real problem remains that the path narrows, the stakes rise, and then you have to demand that everyone is using it. This adoption curve takes time.