Podcast Notes | SemiAnalysis Analyst Breaks Down Current Pullback: Semiconductors Are Paying Debts, But Not Yet at the End of the Cycle

Organized & Compiled: 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)

Air Date: July 29, 2026

Disclosure: Doug O'Loughlin and Dylan Patel are both SemiAnalysis employees. SemiAnalysis 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.

Key Points Summary

Doug O'Loughlin returns to SemiAnalysis Weekly after a long absence, just as the semiconductor sector experiences a sharp drawdown after the "best first half in history." South Korea's KOSPI fell 40%, retail 2x leverage was wiped out, and SK Hynix missed expectations due to a shift to LTA slowing price increases. Doug compares the current situation to the 1980s Taiwan bubble, believing behavioral patterns are highly similar, but fundamentals remain healthy.

The two engage in a fierce debate around "How big is AI demand really." Dylan starts from SemiAnalysis's own experience: after coding agents launched, company AI spend grew 100x, users expanded from 9 to 90, and usage per person also grew 10x. Doug does not deny strong demand, but raises core concerns: the supply side can be calculated clearly, but the demand side is a "trillion dollar question," no one has the answer. More critically, scaling laws require chips to double, but physical and institutional bottlenecks like electricians, capital, and permits cannot double synchronously. Hyperscalers have issued $450B in debt this year, funds coming from pensions and annuities, but pension pools themselves are shrinking.

Highlights Summary

About Market Drawdown

"By the end of Q2, this was the best performance in semiconductor history. Then we started paying back debt. The faster something rises, the greater the gravity."

"Koreans have a 20-year record: buying at the top every time. Banks in 2007, SaaS in 2021, this time yolo'd themselves."

"KOSPI down 40%, 2x leverage people wiped out directly. Then self-fulfilling spiral: everyone watching accounts shrink, decide to sell, then exacerbate the decline."

About Memory Cycle

"SK Hynix shifting to more LTA, price increase slowed from 3x to 30 to 50%. Brains in the finance circle are broken, only look at rate of change. Once the second derivative comes down, they think the cycle is over."

"Semiconductor script is always the same: everyone double orders during shortage, factories see demand and expand production crazily. Then demand sneezes, supply still ramping, utilization drops from 100% to 50%, can only cut prices."

About AI Demand

"Demand curve is the trillion dollar question. Supply curve relatively easy to understand, but whether demand is 10x or 100x, no one knows."

"SemiAnalysis itself is a case: after coding agents launched, 9 tech users became 90 all-staff users, token usage per person also 10x. Company AI spend 100x."

About Supply Chain Bottlenecks

"US lacks 100,000 electricians. Mid-level electrician annual salary $250k, those willing to overtime can reach 400, 500k. Someone uses Cessna small planes to send electricians to remote sites."

"Hyperscalers issued $450B debt this year, second only to US government and China borrowing scale. These funds come from pensions and annuities, but pension pools won't double."

"TSMC directly and indirectly accounts for 20% of Taiwan GDP. If doubling again, Taiwan needs to have more children to gather enough workers."

About AI Politics

"AI unpopularity lower than ice cream, lower than politicians. This is not priced in. AI will become the scapegoat for cost of living issues in mid-term elections."

"ROSA Act passed 300 to 20 in House of Representatives, stuck in Senate. Corporate lobbying power is stopping legislation limiting Chinese remote access to GPUs."

Main Text

Best First Half in Semiconductor History, Then Starting to Pay Back Debt

Dylan: Stock market drawdown, AI names all falling. Today we either add fuel to the fire or give everyone a little comfort.

Doug: By June 30 Q2 end, 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 gravity. We are paying for the previous crazy momentum rally.

The situation in Korea is crazy. Every day stocks hit the daily limit down. There was a tweet saying "How do I do my job?" HR director lost all money, everyone is very depressed because stocks all fell. If you look back at Asian financial market history, this happens more often than you imagine.

My favorite book is about the Taiwan Great Bubble. Taiwan had a 100x bubble on a per capita basis, banks trading at 500x PE, everything was crazy.

Dylan: When was this?

Doug: Late 1980s.

Dylan: Do you think Korea's fundamentals are different from back then?

Doug: Fundamentals are good. But the problem is, things are never as bad as feared, nor as good as you imagine. SK Hynix missed expectations today because they shifted to more LTA. Ironically, when doing ADR roadshows they were still complaining Micron did LTA and got lower prices.

Memory prices rose about 3x last year, impossible to rise 3x again next year, maybe rise 30 to 50%. But brains in the finance circle are broken, only look at rate of change. Historically in memory cycles, once the second derivative comes down it's usually the end. Because the rate of change won't stop at 30%, it will drop directly to negative 50%.

This cycle's script is always the same: everyone invests building factories, capacity comes online, then discover "Oh my, why is demand so little". Because previously it was double orders, triple orders. Factories from 100% utilization drop to 50%, the only way to break even is cut prices. This is the essence of the semiconductor market.

KOSPI is now down 40%. 2x leverage people wiped out directly. Then self-fulfilling spiral: everyone watching accounts shrink, decide to sell, then exacerbate the decline.

China Memory: Might Mess Up the Party, But Demand Still Greater Than Supply

Dylan: Recently China memory entering the ecosystem, CXMT, YMTC large IPOs, how do you see it?

Doug: Historically China touches what and turns it into cabbage price. They have capacity, even if yield is low, doesn't matter. China companies are not competing for profit margin or EPS, shareholders are the government, government incentivizes production, provinces compete for GDP.

CXMT is now clearly market fourth, but this is a shortage environment, they still make money. Apple has started using CXMT memory, because Micron is "price gouging". No one crying in the casino, Tim Apple. You have to buy at market price.

CXMT might mess up the party, but the reality is demand still greater than supply. The real trillion dollar question is: Where exactly is demand? Supply curve relatively easy to understand. Demand curve we don't know. We know coding agents and chatbots mean more demand, but don't know if 10x or 100x. Supply will blindly ramp, until one day crashes into the demand curve.

Coding Agents are the Inflection Point: SemiAnalysis Itself 100x AI Spend

Dylan: I think demand is obviously very strong, and will last a long time. Just looking at my own company's internal usage is enough. If you think 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 see signals in the opposite direction.

Doug: Let me play devil's advocate. What is the biggest bear argument? Tech progress speed might exceed the speed at which people use it. Assume AI's killer app is data entry, Kimi K3 is enough. We build faster and faster cars, better and better products, but the real demand curve is satisfied by a product we already mastered.

This is like the internet bubble: back then said "demand doubles every 90 days", but fiber optic tech improved 2 to 3x every year. Finally one fiber optic's performance became 500,000x the original, then everyone said "Wait, seems we don't need this much fiber".

Dylan: I don't agree, but this is worth discussing. My rebuttal: there are 100 to 1000x more people now not using any models. Second, AI use cases are far beyond coding. It can also do video generation, drug discovery, materials science. Someone uses AI to do superconducting components, worth how much? Worth lots of GPU money.

And coding itself is not just "centering a div". It represents a whole class of economic value far higher than frontend debugging tasks. Sam Altman is talking about RSI (Recursive Self-Improvement), Anthropic has new models coming. Coding agents in that Claude 4.5 version are a clear inflection point: You cross a certain intelligence line, a whole new market appears. Things you couldn't do the day before, you can do the next day.

Doug: You are prototype users. This time last year SemiAnalysis tech team had less than 10 people using coding agents, then you and Dylan said "Company everyone must learn to use this". Now we have 90 users.

Dylan: From 9 to 90, 10x. Then within 3 to 4 months, per person usage also about 10x. Company AI spend 100x. Now the question is, will every company be like this? Maybe not 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 inference for one model needs 100 H100s. Then you say "Let the old girl retire, buy a B300". The real confirmation signal is pricing differentiation between B200 and B300.

Dylan: I completely disagree. The most fundamental reason is: no one will take H100s out and swap for B300s. Data center designs are completely different. You can't in the same server room swap Hopper for Blackwell or Rubin, you have to tear the whole thing down and rebuild. So to justify retiring a whole Hopper data center, you must first prove those chips' revenue is already below operating costs. This is not variable cost, it is sunk cost.

Doug: In a frictionless world you are right, but the world we live in friction is getting bigger. Friction for new build compute includes power permits, land, approvals.

Dylan: Yes, I agree. GPU price decline scenario is model progress stagnates, rise scenario is model progress continues. There is also an X factor: government intervention on frontier labs. If limit who can use the latest best chips, demand will be compressed, old chip prices will also follow and drop.

Capital and Electricians: The Physical Ceiling of scaling laws

Doug: What worries me most is not demand, it is supply side physical bottlenecks. First is electricians. US lacks 100,000 electricians. Mid-level electrician annual salary $250k, those willing to work 18 hours can reach 400, 500k. There is a website tracking electrician hiring, Wayback Machine can see hourly wage from 15, 20 dollars rose to 50, 100, 200 dollars. Training an electrician takes 18 months. People needed to double again, we never trained that many.

Second is capital. Hyperscalers this year issued about $450B debt, largest in history. Second only to US government and China government. Someone has to buy these bonds. To let them buy more, must give higher interest rates. And these funds source, largely is pensions and annuities. Pension pools are structurally shrinking. Pensions already largely shifted to 401k, 401k doesn't buy bonds. So you basically have to believe everyone needs double insurance, but this doesn't make sense.

scaling laws say "Great, we make models 2x bigger". But not everything can 2x 3x amplify synchronously.

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

Doug: Yes. Annuities concentrated purchase before retirement, baby boomers 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 double insurance. No one will buy double life insurance.

Dylan: This is interesting. Pensions structurally shrinking, but indeed lots of money there.

Doug: Another example is Taiwan. TSMC directly and indirectly accounts for 20% of Taiwan GDP. If TSMC again doubles triples, Taiwan needs to have more children to gather enough workers. Taiwan only has one game playing. Taiwan GDP this year rose 25%, is TSMC baking chips. But double again, people won't be enough.

AI Politicization: Scapegoat for Mid-term Elections

Dylan: Many people don't like AI, this is not priced in. How could it be priced in? I think it is mid-term elections.

Doug: AI is probably fifth priority, not top three. Healthcare, cost of living rank front. No one will run for election with AI as platform.

Dylan: But AI will become a sub-proxy for cost of living issues. Not "do we support AI", but "care about economy, blame tech bro and AI". ROSA Act in House of Representatives 300 to 20 passed, stuck in Senate. Corporate lobbying stopping it.

Doug: If not top three priority, lobbying power will win over public opinion.

Dylan: But AI is already people's scapegoat on other issues. Climate change, housing, inflation, AI and tech bro will all be dragged out.

Doug: There is an interesting poll: people hating data centers usually don't live near data centers. And those living nearby, especially young people, attitude is positive, because employment. I visited a data center near Buffalo, locals super supportive. Data centers built in remote areas is actually a good thing, it makes economic participation broad. One gigawatt data center approximately needs 10,000 people. 70 gigawatts is 700,000 jobs. This starts affecting votes.

Endgame: 5 Trillion Investment, 50 Billion Revenue

Doug: Tech boom future will always realize, problem is cash flow timing. You spend 1 trillion, get 100 billion, it indeed one day will become 1 trillion. But might be 5 years later, by then you say "Bro, I'm out of money".

Assume whole AI ecosystem currently ARR is 150 billion, cumulative CAPEX 1 trillion. 15% revenue return, by 50% profit margin calculate is 7.5%. Not bad, but not super profitable. Have to believe 150 billion can become 500 billion, this can be done. Then 500 billion can support 2 to 3 trillion CAPEX. But double again very hard.

OpenAI and Anthropic believe final pretraining is coming, because they want to IPO. Final pretraining out models indeed very good, revenue growth very fast, but growth speed not enough to pay bills. You built a house you can't afford. Spent 5 trillion investment, revenue 50 billion, that's ten years money.

Dylan: You say this is revenue not profit. And you say this when, you know these companies now provide service profit margin how high.

Doug: Yes, we haven't reached that step yet. Now still on narrow road, can see how revenue matches up. Hyperscalers have other businesses crazily making cash, if they want stop CAPEX, profits immediately can print out. But as you invest more and more, stakes higher and higher, road narrower and narrower. To some point you actually have to require everyone is using it. Problem is decision makers and actual adopters are two completely different worlds. Zuck thinks everyone will wear Meta glasses in metaverse every day burn trillion tokens, but Nebraska grandma won't even use new iPhone.

Dylan: Revenue doesn't rely on grandma. Relies on enterprises, banks, telecom, retail companies, defense and intelligence agencies. I see every bank, every telecom company, every retailer using this in daily work. More interesting constraints on supply side: can you install enough GPUs, can you recruit enough people to sell.

Doug: Yes, supply side problems more interesting and harder. Electricians, capital, permits, these things can't according to scaling laws double. But give them time, it will arrive. They next year indeed might issue 1 trillion bonds. The real problem. The problem is the path will become narrow, stakes will become high, then you have to require everyone is using it. This adoption curve needs time.