SemiAnalysis Analyst Breaks Down the Current Pullback: Semiconductors are Repaying Debts, but Haven't Reached the Cycle's End
- Core Thesis: The semiconductor sector experienced a severe pullback after its "best first half ever," with South Korea's KOSPI index falling by 40%. Despite robust AI demand (e.g., internal AI spending at SemiAnalysis has grown 100x), the supply side faces structural constraints like electrical technician shortages, capital bottlenecks, and physical limitations, creating uncertainty about the long-term sustainability of scaling laws.
- Key Elements:
- Market Pullback and Bubble Analogy: The KOSPI has dropped 40% from its peak, wiping out 2x leveraged investors. Doug compares this to the 1980s Taiwan bubble (where bank stocks traded at 500x P/E), noting that while fundamentals are healthy, behavioral patterns are similar.
- Memory Cycle and Second Derivative: SK Hynix's shift to long-term agreements (LTAs) has slowed price increases from 3x to 30-50%. The market, focusing solely on the change in rate, has amplified the panic. The semiconductor cycle is inherently a self-fulfilling spiral of overcapacity following double ordering.
- AI Demand Evidence: An internal case at SemiAnalysis shows that after deploying a coding agent, users expanded from 9 to 90, per-capita token usage grew 10x, and overall AI spending increased 100x, demonstrating the potential for exponential demand growth.
- Supply-Side Physical Bottlenecks: The U.S. faces a shortage of 100,000 electrical technicians, with hourly wages already up 5-10 times. Hyperscalers have issued roughly $450 billion in bonds this year, but the structural shrinkage of pension funds means they cannot support another doubling of investment.
- Risk of AI Politicization: AI could become a scapegoat for the "cost of living" issue in the midterm elections. The ROSA Act passed the House 300:20, but corporate lobbying has prevented the Senate from passing legislation restricting Chinese remote access to GPUs.
- ROI Concerns: The cumulative capital expenditure across the AI ecosystem is $1 trillion, generating annual revenue of about $150 billion. At a 50% profit margin, the return is only 7.5%. Future revenue would need to triple to $500 billion, and doubling from there will be extremely difficult (requiring a $5 trillion investment for only $500 billion in additional revenue).
Compiled & Edited by: Odaily TechFlow
Guest: Doug O'Loughlin, Analyst at SemiAnalysis (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)
Air Date: July 29, 2026
Disclaimer: Doug O'Loughlin and Dylan Patel are employees of SemiAnalysis, a paid research firm for the semiconductor industry whose business model depends on industry health. The following content faithfully presents the original dialogue and does not constitute investment advice.
Key Takeaways
Doug O'Loughlin makes a long-awaited return to SemiAnalysis Weekly, right as the semiconductor sector experiences a sharp pullback following its "best first half ever." South Korea's KOSPI is down 40%, retail investors using 2x leverage have been wiped out, and SK Hynix missed expectations due to a shift towards more LTAs, slowing the pace of price increases. Doug compares the current situation to the Taiwan bubble of the 1980s, arguing that the behavioral patterns of the bubble are highly similar, though the fundamentals remain healthy.
The two engage in a heated debate around "how big is AI demand really." Dylan draws from SemiAnalysis's own experience: after deploying a coding agent, the company's AI spending grew 100x, with users expanding from 9 to 90 and usage per person increasing 10x. Doug doesn't deny strong demand but raises a core concern: the supply side can be calculated, but the demand side is a "trillion-dollar question" with no clear answer. More critically, scaling laws demand that chips double, but physical and institutional bottlenecks like electricians, capital, and permits cannot double in sync. Hyperscalers have already issued $450B in debt this year, funded by pensions and annuities, yet the pool of pension capital is itself shrinking.
Highlight Reel
On the Market Drawdown
"By the end of Q2, it was the best performance in semiconductor history. Then we started paying the piper. The faster things go up, the stronger the gravity."
"The Koreans have a 20-year track record: buying at the top every single time. Bought banks in 2007, bought SaaS in 2021, this time they YOLOed themselves."
"KOSPI is down 40%. People with 2x leverage are completely wiped out. Then it becomes a self-fulfilling spiral: everyone watches their portfolio shrink, decides to sell, and exacerbates the decline."
On the Memory Cycle
"SK Hynix shifted to more LTAs, and the pace of price increases slowed from 3x to 30-50%. The financial world is completely broken, only looking at the rate of change. As soon as the second derivative comes down, they think the cycle is over."
"The semiconductor playbook is always the same: during a shortage, everyone double orders. Fabs see the demand and ramp up production crazily. Then demand sneezes, supply is still ramping, utilization drops from 100% to 50%, and the only option is to cut prices."
On AI Demand
"The demand curve is the trillion-dollar question. The supply curve is relatively understandable, but whether demand is 10x or 100x, nobody knows."
"SemiAnalysis itself is a case study: after deploying a coding agent, we went from 9 technical users to 90 full-time users, and token usage per person also increased 10x. That's a 100x increase in company AI spend."
On Supply Chain Bottlenecks
"The US is short 100,000 electricians. A mid-level electrician makes $250k a year, and those willing to work overtime can make $400k or $500k. People are flying electricians to remote job sites in Cessnas."
"Hyperscalers have issued $450 billion in debt this year, second only to the borrowing scale of the US government and China. This money comes from pensions and annuities, but the pension pool isn't going to double."
"TSMC directly and indirectly accounts for 20% of Taiwan's GDP. If it doubles again, Taiwan will need to have more babies just to have enough workers."
On AI Politics
"AI is less unpopular than ice cream, less unpopular than politicians. This is not priced in. AI is going to be a scapegoat for cost-of-living issues in the midterm elections."
"The ROSA Act passed the House 300 to 20, but is stuck in the Senate. Corporate lobbying is blocking legislation that would restrict remote GPU access for China."
Main Content
The Best First Half in Semiconductor History, Then Paying the Piper
Dylan: The stock market is pulling back, and all the AI names are down. Today, we're either going to pour gasoline on the fire or offer some comfort.
Doug: By June 30th, the end of Q2, it was probably the best performance in semiconductor history. Then we started to unwind. A lot of it can be attributed to technical factors: leverage, momentum reversals. But the reality is, the faster things go up, the stronger the gravity. We're paying for the crazy momentum rally.
The situation in Korea is crazy. Stocks are hitting limit-down every day. There was a tweet saying "How am I supposed to do my job?" The head of HR has lost all their money, and everyone is depressed because all the 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 had a 100x bubble on a per capita basis. Banks were trading at 500x P/E. Everything was insane.
Dylan: When was this?
Doug: Late 1980s.
Dylan: Do you think the fundamentals in Korea are different now?
Doug: The fundamentals are good. But the problem is, things are never as bad as you fear, and never as good as you imagine. SK Hynix missed expectations today because they shifted to more LTA contracts. Ironically, during their ADR roadshow, they were complaining about Micron taking lower prices for doing LTAs.
Memory prices went up about 3x last year. There's no way they go up 3x again next year, maybe 30 to 50%. But the financial world is completely broken, only looking at the rate of change. Historically in memory cycles, once the second derivative comes down, it's usually the end. Because the rate of change doesn't stop at 30%; it goes straight to negative 50%.
The playbook for this cycle is always the same: everyone invests in fabs, capacity comes online, and then they realize "Oh my god, why is demand so low?" Because it was double-ordering, triple-ordering. Fab utilization drops from 100% to 50%, and the only way to recoup costs is to cut prices. That's the nature of the semiconductor market.
KOSPI is down 40% now. People with 2x leverage are completely wiped out. Then it becomes a self-fulfilling spiral: everyone watches their portfolio shrink, decides to sell, and that exacerbates the decline.
Chinese Memory: Could Spoil the Party, But Demand Still Outstrips Supply
Dylan: Recently, Chinese memory has entered the ecosystem, with CXMT and YMTC doing big IPOs. What do you think?
Doug: Historically, whatever China touches, they turn into a commodity. They have capacity, and even if yields are low, it doesn't matter to them. Chinese companies aren't competing for profit margins or EPS; their shareholder is the government. The government incentivizes production, and provinces compete with each other for GDP.
CXMT is clearly the fourth player now, but this is a shortage environment, so they can still make money. Apple has started using CXMT memory because Micron was "price gouging." Nobody is crying in the casino, Tim Apple. You have to buy at market price.
CXMT could spoil the party, but the reality is that demand still outstrips supply. The real trillion-dollar question is: Where exactly is the demand? The supply curve is relatively easier to understand. We don't know the demand curve. 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 Agents Are the Inflection Point: SemiAnalysis's Own 100x AI Spend
Dylan: I think demand is clearly very strong and will continue for a long time. I only need to look at my own company's internal usage. If you believe future demand will plateau or decline, you have to believe the models will stop getting better. I don't see any signs of stagnation, only signals pointing the other way.
Doug: Let me play devil's advocate. What is the biggest bear argument? The rate of technological progress might outpace the rate at which people use it. Suppose AI's killer app is data entry, and the Kimi K3 is good enough. We build faster and faster cars, better and better products, but the true demand curve is already satisfied by a product we already have.
It's like the internet bubble: they said "demand doubles every 90 days," but fiber optic technology improved 2 to 3 times every year. Eventually, a single fiber had 500,000 times the capacity, and people said "Wait, it looks like we don't need all this fiber."
Dylan: I disagree, but it's worth discussing. My counterpoint is: there are 100 to 1000 times more people who aren't using any models right now. Second, the use cases for AI go far beyond coding. It can do video generation, drug discovery, materials science. Someone is using AI for superconducting components. How much is that worth? It's worth a lot of GPUs.
And coding itself isn't just "centering a div." It represents a whole category of tasks with economic value far higher than front-end debugging. Sam Altman talks about RSI (Recursive Self-Improvement), and Anthropic has new models coming out. The coding agent was a clear inflection point with Claude 4.5: you cross a certain intelligence threshold, and a whole new market appears. What you couldn't do the day before, you can do the next day.
Doug: You are prototype users. This time last year, SemiAnalysis's tech team had less than 10 people using coding agents. Then you and Dylan said "Everyone in the company needs to learn how to use this." Now we have 90 users.
Dylan: From 9 to 90, that's 10x. Then, within 3 to 4 months, usage per person also went up about 10x. So company AI spend went up 100x. The question now is, will every company be like this? Probably not to our intensity, but many companies have a huge amount of work that can be cut.
H100s Won't Turn to 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 "Let the old girl retire, buy a B300." The real confirmation signal will be pricing divergence between B200 and B300.
Dylan: I completely disagree. The most fundamental reason is: nobody is going to rip out H100s and replace them with B300s. The data center designs are completely different. You can't put Hopper in the same room and swap it for Blackwell or Rubin; you have to tear the whole thing down and rebuild. So to justify retiring an entire Hopper data center, you first have to prove that the revenue from those chips is less than the operating costs. That's a sunk cost, not a variable cost.
Doug: In a frictionless world, you'd be right, but the world we live in has increasing friction. The friction for building new compute includes power permits, land, and approvals.
Dylan: Right, I agree. The scenario where GPU prices fall is where model progress stalls. The scenario where they rise is where model progress continues. There's also an X factor: government intervention in frontier labs. If they restrict who can use the newest, best chips, demand gets compressed, and older chip prices will also fall.
Capital and Electricians: The Physical Ceiling of Scaling Laws
Doug: What worries me most isn't demand, but the physical bottlenecks on the supply side. First, electricians. The US is short 100,000 electricians. A mid-level electrician makes $250k a year. Those willing to work 18-hour days can make $400k or $500k. There's a website tracking electrician job postings; on the Wayback Machine, you can see hourly rates go from $15 or $20 to $50, $100, $200. It takes 18 months to train an electrician. We've never trained enough for what it would take to double capacity again.
Second is capital. Hyperscalers have issued roughly $450 billion in debt this year. The largest ever. Second only to the US government and the Chinese government. Someone has to buy this debt. To get them to buy more, you have to offer higher interest rates. And the source of this money is largely pensions and annuities. The pension pool is structurally shrinking. Pensions have already largely shifted to 401ks, and 401ks don't buy bonds. So you basically have to believe everyone needs twice the insurance, which doesn't make sense.
Scaling laws say "Great, let's make the model twice as big." But not everything can be scaled up 2x or 3x simultaneously.
Dylan: Wait, are you saying pensions are paying for the data center buildout?
Doug: Yes. Annuities are bought heavily before retirement. The baby boomers are all 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 twice the insurance. Nobody is going to buy two life insurance policies.
Dylan:
