Key Points Summary
Doug O'Loughlin makes a long-awaited return to SemiAnalysis Weekly, as the semiconductor sector goes through a "best-ever first half of the year" followed by a sharp drawdown. The Korean KOSPI drops 40%, retail investors with 2x leverage get liquidated, and SK Hynix misses expectations due to a shift to LTA causing a slowdown in price increases. Doug compares the current situation to the 1980s Taiwan bubble, seeing a high degree of similarity in bubble behavior, but the fundamentals remain healthy.
The two engage in a heated debate about "How Big is the AI Demand." Dylan, drawing from SemiAnalysis's own experience, notes that after the deployment of encoding agents, the company's AI expenditure grew 100-fold, the team expanded from 9 to 90 people, and individual usage increased tenfold. Doug does not deny the strong demand but raises a key concern: while the supply side can be calculated, the demand side is a "trillion-dollar question" with no clear answer. More critically, scaling laws require chip doubling, but physical and institutional bottlenecks such as electrical work, capital, permits cannot double in sync. The hyperscale cloud providers have already issued $450 billion in bonds this year, with funds coming from pension and retirement pools, which are themselves shrinking.
Key Insights
On Market Drawdown
"By the end of Q2, this was the best-ever performance in the history of semiconductors. Then we started to pay the piper. 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-ed themselves."
"KOSPI drops 40%, those with 2x leverage get wiped out immediately. Then it's a self-fulfilling spiral: everyone watches their accounts shrink, decides to sell, then exacerbates the downturn."
On the Memory Cycle
「SK Hynix is shifting towards more LTA, with price increases slowing from 3X to 30 to 50%. The minds in the financial sector are all messed up, only looking at the rate of change. Once you take the second derivative, you think the cycle has ended.」
「The semiconductor script is always the same: when there is a shortage, everyone double-orders, and factories see the demand skyrocket and ramp up production like crazy. Then when demand sneezes, supply is still ramping, and the utilization rate drops from 100% to 50%, forcing prices to be cut.」
About AI Demand
「The demand curve is a trillion-dollar question. The supply curve is relatively easy to understand, but whether the demand is 10X or 100X, no one knows.」
「SemiAnalysis is a case in point: after launching the agent encoding, 9 technical users turned into 90 full users, and the token usage per person also increased by 10X. The company's AI expenditure increased by 100X.」
About Supply Chain Bottlenecks
「The U.S. is short of 100,000 electricians. Mid-level electricians earn $250,000 annually, and those willing to work overtime can make $400,000 to $500,000. Some are using Cessna small planes to fly electricians to remote sites.」
「Major cloud hyperscalers issued $450 billion in debt this year, second only to the borrowing of the U.S. government and China. This money comes from retirement and pension funds, but the retirement fund pool will not double.」
「TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If this doubles, Taiwan would need to have more children to have enough workers.」
About AI Politics
「AI's unpopularity is lower than ice cream, lower 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 Act passed in the House by 300 to 20, but is stuck in the Senate. Corporate lobbying power is hindering legislation to restrict Chinese remote access to GPUs.」
Body
Semiconductor's Best First Half Ever, Then It's Debt Repayment Time
Dylan: The stock market is retracting, and all AI names are falling. Today, we either add fuel to the fire or provide some comfort to everyone.
Doug: By the end of Q2 on June 30, this was 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 are paying the price for the previous crazy momentum rally.
The situation in Korea is insane. Stocks are hitting limit down every day. There was a tweet saying, "How do I do my job?" The HR director lost all their money, everyone is very depressed because all the stocks tanked. If you look back at the history of the Asian financial markets, this kind of thing happens more frequently than you think.
One of my favorite books is about the Taiwan Bubble. Taiwan saw a bubble of 100x on a per capita basis, bank trades at a P/E of 500x, everything went crazy.
Dylan: When was this?
Doug: Late 1980s.
Dylan: Do you think the fundamentals in Korea now are different from back then?
Doug: The fundamentals are good. But the problem is, things are never as bad as you fear, nor as good as you imagine. SK Hynix missed expectations today because they shifted more to LTA. Ironically, when they were doing the ADR roadshow, they were mocking Micron for doing LTA at a lower price.
Memory prices tripled last year, it's impossible for them to triple again next year, maybe they'll rise by 30 to 50%. But the minds of the finance industry are all broken, they only look at the rate of change. Historically, in the memory cycle, when the second derivative comes down, that's usually the end. Because the rate of change doesn't stop at 30%, it goes directly to negative 50%.
The script for this cycle is always the same: everyone invests in factories, capacity comes online, and then they realize, "Wow, demand is so low." Because before it was double ordering, triple ordering. Factories go from 100% utilization to 50%, the only way to break even is to cut prices. That's the nature of the semiconductor market.
KOSPI is now down 40%. Those with 2x leverage are completely wiped out. Then it's a self-fulfilling spiral: everyone watches their account shrink, decides to sell, and exacerbates the decline.
Chinese Memory: Might Crash the Party, but Demand Still Outstrips Supply
Dylan: Recently, Chinese memory has entered the ecosystem with CXMT and YMTC's major IPOs. What's your take?
Doug: They have the capacity, so even with low yields, it doesn't matter. Chinese companies are not in it to compete on profit margins or EPS.
CXMT is now clearly the fourth player in the market, but in this shortage environment, they can still make money. Apple has already started using CXMT's memory because of Micron's "price gouging." No one's crying in the casino, Tim Apple. You have to buy at market prices.
CXMT might crash 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. Supply will blindly ramp up until one day it hits the demand curve.
Coding Agents as a Turning Point: SemiAnalysis' Own 100x AI Spend
Dylan: I think the demand is clearly very strong and will last a long time. Just looking at the internal usage in my own company is enough. If you believe future demand will plateau or decrease, you have to believe the models won't get any better. I see no signs of stagnation, only signals in the opposite direction.
Doug: Let me play devil's advocate for a moment. What's the biggest bear argument? The speed of technological progress may outpace how quickly people utilize it. Suppose AI's killer app is data entry, Kimi K3 is sufficient. We're making faster and better products, but the true demand curve is being met by a product we already have mastered.
This is like the Internet bubble: they said, "Demand doubles every 90 days," but fiber optic technology improved by 2 to 3 times each year. Eventually, the performance of the last fiber optic became 500,000 times its original capacity, and then everyone said, "Wait, it seems we don't need that much fiber optic."
Dylan: I disagree, but it's worth discussing. My rebuttal is: there are still 100 to 1000 times more people who are not currently using any models. Second, AI use cases go far beyond coding. It can also do video generation, drug discovery, material science. Someone is using AI for superconducting devices; how much is that worth? It's worth a lot more than a GPU.
Moreover, the code itself is not just about "centering a div." It represents a whole class of economic value far beyond front-end debugging tasks. Sam Altman talks about RSI (Recursive Self-Improvement), and Anthropic has a new model coming out. The Code Proxy was a clear turning point in version 4.5 of Claude: you cross a certain smart line, and a whole new market appears. What you couldn't do the day before, you can do the next day.
Doug: You are the prototype user. This time last year, the SemiAnalysis tech team of less than 10 people was using the Code Proxy, and then you and Dylan said, "Everyone in the company must learn to use this." Now we have 90 users.
Dylan: From 9 to 90, 10x. And then within 3 to 4 months, everyone's usage has also increased by about 10x. The company's AI spending has increased by 100x. Now the question is, will every company do this? Maybe not at our intensity, but many companies have a lot of work to cut.
The H100 Won't Become Scrap Metal, But Models Are Growing
Doug: I think old chips will become worthless. Everyone says "H100 is an appreciating asset," but one day you will need 100 H100s to reason about a model. By then, you will say, "Let the old lady retire, buy a B300." The real confirmation signal is when there is price differentiation between B200 and B300.
Dylan: I completely disagree. The most fundamental reason is: no one will take out an H100 and replace it with a B300. Data center designs are completely different. You can't swap a Hopper for a Blackwell or Rubin in the same server room; you have to tear down the whole thing and rebuild it. So, to justify retiring a whole Hopper data center, you first have to prove that the revenue from those chips is below the operating cost. This is not a variable cost; it's a sunk cost.
Doug: In a frictionless world, you are right, but the world we live in has increasing friction. The friction of adding new compute includes power permits, land, approvals.
Dylan: Yes, I agree. The scenario of GPU prices falling is when model progress stalls, and the scenario of rising prices is when model progress continues. There is also an X factor: government intervention in cutting-edge laboratories. If there are restrictions on who can use the latest and best chips, demand will be compressed, and the prices of old chips will also fall.
Capital and Electricians: The Physics Ceiling of Scaling Laws
Doug: What worries me the most is not demand, but the physical bottleneck on the supply side. The first one is electricians. The U.S. is short of 100,000 electricians. Mid-level electricians make $250,000 per year and are willing to work 18 hours a day for $400,000 to $500,000. There is a website that tracks electrician hiring, and on the Wayback Machine, you can see the hourly wage increase from $15 to $20 to $50, $100, $200. It takes 18 months to train an electrician. To double the needed workforce, we have never trained that many before.
The second one is capital. The hyperscale cloud providers issued around $450 billion in debt this year, the largest in history, second only to the U.S. and Chinese governments. Someone has to buy this debt. To get them to buy more, you have to offer a higher interest rate. The source of this money is largely pension funds and annuities. The pension fund pool is structurally shrinking. Pensions have largely shifted to 401(k)s, and 401(k)s do not buy debt. So basically, you have to believe that everyone needs double the insurance, but that doesn't make sense.
Scaling laws say, "Great, we'll just make the model twice as big." But not everything can be scaled up two or three times in sync.
Dylan: Wait, are you saying pensions are footing the bill for data center construction?
Doug: Yes. Pension funds are buying in bulk ahead of retirement, and the baby boomer generation is all retiring, so that asset pool is quite large. But can it double? Can it triple? I don't think so. Life insurance is also a source. But you have to believe that everyone needs double the insurance. No one is going to buy double the life insurance.
Dylan: That's very interesting. Pensions are structurally shrinking, but there is indeed a lot of money there.
Doug: Another example is Taiwan, China. TSMC directly and indirectly accounts for 20% of Taiwan, China's GDP. If TSMC doubles or triples again, Taiwan, China will need to have more babies to have enough workers. Taiwan, China is playing only one game. Taiwan, China's GDP grew by 25% this year, thanks to TSMC making chips. But if they double again, there won't be enough people.
AI Politicization: The Scapegoat of Midterm Elections
Dylan: Many people dislike AI, and this has not been priced in. How could it be priced? I think it will be in the midterms.
Doug: AI is probably fifth on the priority list, not in the top three. Healthcare and the cost of living come first. No one is going to run for office with AI as their platform.
Dylan: But AI will become a subagent of the cost-of-living issue. It's not about "Do we support AI?" but about "Economic concerns, tech bros, and AI." The ROSA Act passed the House 300 to 20, but it's stuck in the Senate. Corporate lobbying is blocking it.
Doug: If it's not in the top three priorities, lobbying power will prevail over public opinion.
Dylan: But AI has already become a scapegoat for people on other issues. Climate change, housing, inflation – AI and tech bros will be brought up.
Doug: There's an interesting poll: People who hate data centers usually don't live near one. Those who do, especially young people, are positive because of job opportunities. I visited a data center near Buffalo, and the locals were very supportive. Having data centers in remote areas is actually a good thing; it broadens economic participation. A one gigawatt data center requires about ten thousand jobs. 70 gigawatts equals 700,000 jobs. This is starting to affect votes.
Endgame: $5 Trillion Investment, $500 Billion Revenue
Doug: The future of technological prosperity will always be achieved; the issue lies in the timing of cash flow. You spend a trillion dollars but only receive a hundred billion; it will indeed someday become a trillion. But maybe in five years, by which time you'll be saying, "Dude, I'm broke."
Assuming the entire AI ecosystem currently has an ARR of $150 billion with a cumulative CAPEX of $1 trillion. A 15% return on income, based on a 50% profit margin, is 7.5%. Not bad, but not super lucrative either. You have to believe that $150 billion can become $500 billion, which is achievable. Then $500 billion can support two to three trillion in CAPEX. But going for another round of doubling is very difficult.
OpenAI and Anthropic believe that ultimate pretraining is on the horizon because they are planning to IPO. Models pretrained in the end are indeed very good, with rapidly growing income, but the growth rate is not fast enough to foot the bill. You've built a house you can't afford. You've spent $5 trillion on investment, with $500 billion in revenue, but that's ten years' worth of money.
Dylan: You said this is revenue, not profit. And when you said that, you knew how high the profit margins are for the services these companies are offering now.
Doug: Right, we're not there yet. We're still on the narrow path, trying to align the revenue. The hyperscale cloud providers have other cash cow businesses, and if they want to turn off CAPEX, profit can immediately pop out. But as you invest more and more, the stakes get higher, and the path gets narrower. At some point, you actually have to ask everyone to be using it. The problem is, decision-makers and actual adopters inhabit two completely different worlds. Zuck thinks everyone will be wearing Meta glasses and burning trillions of tokens in the metaverse every day, but grandma in Nebraska won't even know how to use a new iPhone.
Dylan: Revenue doesn't depend on grandma. It depends on enterprises, banks, telcos, retailers, defense, and intelligence agencies. I see every bank, every telco, every retailer using this in their day-to-day operations. The more interesting constraints are on the supply side: Can you get enough GPUs, can you hire enough people to sell.
Doug: Yes, the issues on the supply side are more intriguing and more challenging. Electricians, capital, licenses—these things don't scale on scaling laws. But give them time, it will come. They might actually issue a trillion in bonds next year. The real problem is that the path will narrow, the stakes will get higher, and then you have to ask everyone to be using it. This adoption curve takes time.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
