Doug O'Loughlin makes a long-awaited return to SemiAnalysis Weekly, just as the semiconductor sector experiences a sharp pullback following its "best first half in history." South Korea's KOSPI has fallen 40%, wiping out retail investors with 2x leverage, while SK Hynix missed expectations as a shift towards more Long-Term Agreements (LTAs) slowed price increases. Doug draws parallels between the current situation and the 1980s Taiwan bubble, noting highly similar behavioral patterns but fundamentally healthy underlying conditions.
The two engage in a heated debate over "How big is the AI demand really?" Dylan points to SemiAnalysis's own experience: after launching a coding agent, the company's AI spending grew 100x, the user base expanded from 9 to 90 people, and per-person usage also grew 10x. Doug doesn't deny strong demand but raises core concerns: the supply side can be calculated, but demand is a "trillion-dollar question" with no clear answer. More crucially, while scaling laws call for doubling chips, physical and institutional bottlenecks like electricians, capital, and permits cannot double at the same rate. Hyperscale cloud providers have already issued $450B in debt this year, funded by pensions and annuities, while the pension pool itself is shrinking.
Key Insights Summary
On the 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 record: buying at the peak every time. In 2007, they bought banks; in 2021, SaaS; this time, they YOLO'd themselves."
"The KOSPI is down 40%. People with 2x leverage got wiped out. Then comes the self-fulfilling spiral: everyone sees their accounts shrink, decides to sell, and accelerates the decline."
On the Memory Cycle
"SK Hynix shifted to more LTAs, slowing price increases from 3x to 30-50%. The financial world's thinking is broken; they only look at the rate of change. The second derivative turns down, and they think the cycle is over."
"The semiconductor script is always the same: during shortages, everyone double-orders, fabs see insane demand and expand capacity. Then demand catches a cold, supply keeps ramping, utilization drops from 100% to 50%, and the only way to survive is to cut prices."
On AI Demand
"The demand curve is the trillion-dollar question. The supply curve is relatively understandable, but is demand 10x or 100x? No one knows."
"SemiAnalysis itself is a case study: after the coding agent launched, 9 technical users became 90 full-staff users, and each person's token usage also grew 10x. The company's AI spending grew 100x."
On Supply Chain Bottlenecks
"The US is short 100,000 electricians. A mid-level electrician earns $250k a year, up to $400k-$500k with overtime. People are using Cessnas to fly electricians to remote job sites."
"Hyperscale cloud providers have issued $450 billion in debt this year, second only to the US and Chinese governments in borrowing scale. This money comes from pensions and annuities, but the pension pool won't double."
"TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If it doubles again, Taiwan would need to have more babies to get enough workers."
On AI Politics
"AI is less unpopular than ice cream, less unpopular than politicians. This isn't priced in. In the midterms, AI will become a scapegoat for cost-of-living issues."
"The ROSA Act passed the House 300-20, stuck in the Senate. Corporate lobbying power is blocking legislation to restrict Chinese remote access to GPUs."
Main Text
Best First Half in Semiconductor History, Then Payback Begins
Dylan: The stock market is pulling back; all AI names are down. Today we either pour fuel on the fire or offer some comfort.
Doug: By June 30th, the end of Q2, this was arguably the best performance in semiconductor history. Then we started unwinding. Much of this can be attributed to technical factors: leverage, momentum reversal. But the reality is, the faster something rises, the stronger gravity pulls it down. We're paying for the previous insane momentum rally.
The situation in Korea is crazy. Stocks hit daily limits down. A tweet said, 'How do I do my job?' The HR director lost all their money, everyone is depressed because their stocks are down. If you look back at Asian financial market history, this happens more often than you think.
My favorite book is about the Great Taiwan Bubble. Taiwan saw a 100x bubble on a per-capita basis, banks trading at 500x P/E, everything went crazy.
Dylan: When was this?
Doug: Late 1980s.
Dylan: Do you think Korea's fundamentals now are different from then?
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 shifted to more LTAs. Ironically, during their ADR roadshow, they were criticizing Micron for taking lower prices with LTAs.
Memory prices rose about 3x last year. They can't rise another 3x next year, maybe 30-50%. But the financial world's thinking is broken; they only look at the rate of change. Historically in memory cycles, when the second derivative turns down, it's usually the end. Because the rate of change doesn't stop at 30%; it goes straight to negative 50%.
The script for this cycle is always the same: Everyone invests in building fabs, capacity comes online, and then they find, 'Oh my god, demand is so weak.' Because it was double-ordering, triple-ordering before. Fabs go from 100% utilization to 50%, and the only way to break even is to cut prices. That's the nature of the semiconductor market.
The KOSPI is down 40% now. People with 2x leverage got wiped out. Then comes the self-fulfilling spiral: Everyone sees their accounts shrink, decides to sell, and accelerates the decline.
Chinese Memory: Might Spoil the Party, But Demand Still Exceeds Supply
Dylan: Recently, Chinese memory entered the ecosystem with CXMT, YMTC's big IPO. What's your take?
Doug: They have capacity. Even with low yields, it doesn't matter. Chinese companies aren't competing on margins or EPS.
CXMT is now clearly the fourth player, but this is a shortage environment; they can still make money. Apple has started using CXMT memory because Micron was 'price gouging.' No one cries in the casino, Tim Apple. You have to pay market prices.
CXMT might spoil the party, but the reality is demand still exceeds supply. The real trillion-dollar question is: where is the demand curve? The supply curve is relatively understandable. The demand curve we don't know. We know coding agents and chatbots mean more demand, but we don't know if it's 10x or 100x. Supply will ramp blindly until one day it hits the demand curve.
Coding Agent as Inflection Point: SemiAnalysis's Own 100x AI Spending
Dylan: I think demand is clearly very strong and will last a long time. I just look at my own company's internal usage. If you believe future demand will flatten or even decline, you have to believe models won't get better. I don't see any signs of stagnation, only signals in the opposite direction.
Doug: Let me play devil's advocate. What's the biggest bear argument? The pace of technological improvement might outpace the rate people adopt it. Suppose the killer app for AI is data entry; Kimi K3 is enough. We build faster cars, better products, but the real demand curve is satisfied by a product we already have.
It's like the dot-com bubble: back then they said 'demand doubles every 90 days,' but fiber optic technology improved 2-3x per year. In the end, the capacity of a single fiber became 500,000 times what it was, and then people said, 'Wait, we don't need this much fiber.'
Dylan: I don't agree, but it's worth discussing. My counter: there are still 100 to 1000x more people not using any model now. Second, AI use cases go far beyond coding. It can do video generation, drug discovery, materials science. Someone's using AI for superconducting components; how much is that worth? Worth a lot of GPUs.
And coding itself isn't just 'centering a div.' It represents a whole class of tasks with economic value far higher than front-end debugging. Sam Altman talks about RSI (Recursive Self-Improvement), Anthropic has new models coming. The coding agent in the Claude 4.5 version was a clear inflection point: You cross a certain intelligence threshold, and a whole new market opens up. What you couldn't do yesterday, you can do today.
Doug: You are the prototypical user. Last year at this time, less than 10 people on the SemiAnalysis tech team used a coding agent. Then you and Dylan said, 'Everyone in the company needs to learn to use this.' Now we have 90 users.
Dylan: From 9 to 90, that's 10x. Then within 3-4 months, per-person usage also grew about 10x. The company's AI spending grew 100x. The question now is, will every company do this? Probably not at our intensity, but many companies have lots of work to cut.
H100 Won't Become Scrap, But Models Are Getting Bigger
Doug: I think old chips will become worthless. Everyone says 'H100 is an appreciating asset,' but one day inferencing a model might require 100 H100s. Then you'll say, 'Retire the old girl, buy a B300.' The real confirmation signal is when pricing diverges between the B200 and B300.
Dylan: I completely disagree. The fundamental reason: No one will pull out H100s and replace them with B300s. Data center designs are completely different. You can't swap Hopper for Blackwell or Rubin in the same server room; you have to tear down and rebuild the whole thing. So to justify retiring an entire Hopper data center, you must first prove the revenue from those chips is below the operating cost. It's not variable cost; it's sunk cost.
Doug: In a frictionless world, you're right, but we live in a world with increasing friction. The friction for new compute includes power permits, land, approvals.
Dylan: Right, I agree. The scenario for GPU price declines is if model progress stalls; the scenario for price increases is if model progress continues. There's also an X-factor: government intervention on frontier labs. If they restrict who can use the newest, best chips, demand gets compressed, and prices for older chips follow down.
Capital and Electricians: The Physical Ceiling of Scaling Laws
Doug: What worries me most isn't demand; it's the physical bottlenecks on the supply side. First, electricians. The US is short 100,000 electricians. A mid-level electrician earns $250k a year, up to $400k-$500k for those willing to work 18-hour shifts. A website tracks electrician job postings; on the Wayback Machine, you can see hourly wages rising from $15-$20 to $50, $100, $200. Training an electrician takes 18 months. The number of people needed to double this, we've never trained that many.
Second, capital. Hyperscale cloud providers issued about $450 billion in debt this year, the largest ever. Second only to the US and Chinese governments. Someone has to buy these bonds. To make them buy more, you need to offer higher interest rates. And the source of this money is largely pensions and annuities. The pension pool is structurally shrinking. Pensions have largely shifted to 401(k)s, and 401(k)s don't buy bonds. So you essentially have to believe everyone needs twice as much insurance, which doesn't make sense.
Scaling laws say, 'Great, let's double the model size.' But not everything can double or triple in sync.
Dylan: Wait, are you saying pensions are funding data center construction?
Doug: Yes. Annuities are purchased before retirement; the baby boomer generation is retiring, so this asset pool is relatively large. But can it double? Triple? I don't think so. Life insurance is another source. But you'd have to believe everyone needs twice the insurance. No one buys double life insurance.
Dylan: That's interesting. Pensions are structurally shrinking, but there's a lot of money there.
Doug: Another example is Taiwan. TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If TSMC doubles or triples again, Taiwan would need to have more babies to get enough workers. Taiwan only has one game to play. Taiwan's GDP grew 25% this year, it's TSMC baking chips. But to double again, there aren't enough people.
AI Politicization: The Midterm Election Scapegoat
Dylan: A lot of people dislike AI; this isn't priced in. How could it be priced? I think the midterm elections.
Doug: AI is probably the fifth priority, not top three. Healthcare, cost of living come first. No one will run on an AI platform.
Dylan: But AI will become a sub-proxy for cost-of-living issues. Not 'Do we support AI?' but 'Care about the economy, blame tech bros and AI.' The ROSA Act passed the House 300-20, stuck in the Senate. Corporate lobbying is blocking it.
Doug: If it's not a top-three priority, lobbying power will win over public sentiment.
Dylan: But AI is already a scapegoat people use for other issues. Climate change, housing, inflation—AI and tech bros get pulled out.
Doug: There's an interesting poll: People who hate data centers usually don't live near them. Those who live nearby, especially younger people, have positive attitudes because of jobs. I visited a data center near Buffalo; the locals were super supportive. Data centers in remote areas are actually good; they broaden economic participation. A one-gigawatt data center needs about ten thousand people. Seventy gigawatts is 700,000 jobs. This starts to affect votes.
Endgame: Five Trillion Investment, Fifty Billion Revenue
Doug: The future promised by a tech boom always arrives; the problem is the timing of cash flows. You spend a trillion, get a hundred billion back; it will indeed become a trillion one day. But that might be five years later, and by then you say, 'Dude, I'm out of money.'
Assume the entire AI ecosystem currently has an ARR of $150 billion, with cumulative CAPEX of one trillion. A 15% return on revenue, assuming 50% margins, is 7.5%. Not bad, but not super profitable either. You have to believe $150 billion can become $500 billion, which is possible. Then $500 billion can support two to three trillion in CAPEX. But doubling again is very hard.
OpenAI and Anthropic believe final pre-training is coming because they want to IPO. The model from final pre-training is indeed good, revenue grows fast, but not fast enough to pay the bills. You built a house you can't afford. Spent five trillion in investment, get fifty billion in revenue; that's a decade's worth of money.
Dylan: You're talking about revenue, not profit. And you say this knowing how high the margins are for these companies' current services.
Doug: Right, we're not there yet. We're still on a narrow path, seeing how revenue can match up. Hyperscale cloud providers have other businesses printing cash; if they wanted to stop CAPEX, profits would shoot up immediately. But as you put more in, the stakes get higher, the path gets narrower. At some point, you practically require everyone to use it. The problem is that decision-makers and actual adopters are two completely different worlds. Zuck thinks everyone will wear Meta glasses and burn a trillion tokens a day in the metaverse, but a grandma in Nebraska doesn't even use the new iPhone.
Dylan: Revenue doesn't rely on grandma. It relies on enterprises, banks, telecoms, retail companies, defense, and intelligence agencies. I see every bank, every telco, every retailer using this in daily work. The more interesting constraints are on the supply side: Can you install enough GPUs, can you hire enough people to sell them?
Doug: Right, the supply-side problems are more interesting and harder. Electricians, capital, permits—these things can't double per scaling laws. But given time, they'll arrive. They might indeed issue a trillion in bonds next year. The real problem is the path narrows, stakes get higher, and then you require everyone to be using it. This adoption curve takes time.
