Korean stocks accelerate de-leveraging? Tom Lee: The AI bubble isn’t even early—let alone late stage

Fundstrat’s Tom Lee pointed out on-site at the NYSE that the Korean stock market’s and the AI semiconductor selloff is a forced liquidation of leveraged capital, not a fundamental turnaround. He cited Cisco’s history: it was cut in half four times from 1993 to 2000, and ultimately surged 100x. He believes an AI bubble doesn’t even count as being at the late stage.

(Background recap: Bitmine buys another ten thousand ETH! Holdings reach 5.79 million ETH; Tom Lee: Ethereum momentum strengthened)

(Background addition: Bitmine scoops up 9,946 ETH to control 4.8% of the global supply! Tom Lee: From a technical standpoint, Ethereum at $2,500)

Table of Contents

Toggle

• Forced deleveraging: weak hands’ positions have been cleaned out
• Cisco cut in half four times, up 100x: AI bubble doesn’t even count as late-stage
• NVIDIA’s 16x PE isn’t expensive, but memory stocks have a bullwhip effect
• Ethereum vs Bitcoin: interest-bearing assets vs digital gold
• Crypto catalysts in line: CLARITY Act and the breakout power of Robinhood Chain

As the Korea Kospi index plunged in a single month and AI semiconductor stocks collectively flashed to a stop, just as the market panicked about an “AI bubble bursting,” Fundstrat co-founder Tom Lee **at the NYSE现场 during a Global Money Talk interview directly called out the core view: this is not a fundamental reversal, but a forced deleveraging event.

He warned that if retail investors exit now, they will ultimately chase back in at higher levels. Lee even brought up Cisco’s history from 1993 to 2000: cut in half four times and ultimately surged 100x. He believes the structural trend for AI is far from over.

Tom Lee thinks the past month’s selloff in the Korean stock market and the AI semiconductor sector was a “forced deleveraging” event. In recent years, the Korean market has introduced a large number of leveraged products, amplifying bidirectional volatility.

Based on data from U.S. investment banks’ lead broker-dealer information, the speed at which hedge funds have been selling their tech-stock long positions hit the fastest pace in nearly 10 years. The “weak hands” have already been washed out. He cited two well-known quotes from Peter Lynch and Charlie Munger to support his core advice: don’t trade in cycles within a structural trend; “money is sitting there waiting to come out, not made by buying and selling.”

Lee used Cisco’s 1993–2000 history to explain where AI sits. During that cycle, Cisco experienced multiple pullbacks of more than 40%. Each time, people announced, “Tech stocks are done.” But ultimately, it climbed from $0.80 to $80, a 100x move. The key difference is that the top in 2000 was a real bubble: the buyers were fiber-optic companies that used unrealistic DCF models to justify procurement. “Today’s buyer is a hyperscaler—serious companies buying equipment, stocking it, and taking big orders.”

On the impact of China’s AI models (Kimi K3), Lee admitted it’s an existential question “beyond my pay grade,” but pointed out that open-source models are essentially generic drugs: someone’s R&D costs still have to be paid for. He’s more focused on downstream AI opportunities (Mag 7, software, crypto), which are starting to outperform upstream semiconductor businesses—and his GRNY ETF is precisely built by sticking to this framework. This year it has outperformed its peers by 92%. For his macro view for the second half, Lee is betting that inflation will come in below expectations (the oil-price shock has already passed its peak; housing and wages are weakening). That would force the Federal Reserve to turn more dovish.

Forced deleveraging is not the end of the story

“The speed at which hedge funds are selling their tech-stock long positions is the fastest in the past 10 years. Weak hands have already been shaken out.”

“Every time the market rises straight up, it traps leveraged longs inside, and then they get forced to liquidate. This is what’s happening now.”

“No one can time the bottom precisely. But if you sell and step aside now, then wait for signals to re-enter, you’ll end up chasing at even higher levels.”

Sell the flowers, water the weeds: don’t trade in cycles in a bull market

“Peter Lynch said that selling the winner you hold is like cutting the flowers and watering the weeds.”

“Charlie Munger put it even better: money isn’t made from buying and selling. Money is made by waiting for it to come out, sitting there.”

Forced deleveraging: weak hands’ positions have been cleaned out

“If there are structural themes, you should buy into them and then forget about them.”

NVIDIA’s 16x PE isn’t expensive, but memory stocks naturally have a bullwhip effect

“NVIDIA’s forward PE is 16x, not the twenties. It has a CUDA moat and an almost certain upgrade roadmap, so it should be re-rated as a growth stock at 25–30x.”

“Memory and semiconductor equipment are two layers removed from end customers, so there’s a risk of a bullwhip effect: hyperscalers may place repeat orders because they expect price hikes, and memory manufacturers in the future may over-expand production.”

“Cyclical stocks have the lowest PE at the top of the cycle. That doesn’t mean it’s a sell signal. You only need to bet that earnings forecasts will keep getting revised upward.”

Cisco cut in half four times, up 100x: AI hasn’t reached the late stage yet

“Cisco rose 100x from 1993 to 2000. Along the way, it was cut in half at least four times, and every time someone said tech stocks were finished.”

“If this were truly the late stage of an AI bubble, people should be shouting, ‘This is the bottom—buy semiconductors aggressively now.’ But what are they doing? They’re疯狂地卖出.”

“In 2000, Cisco’s 200x PE meant the buyers were the fiber companies that used a 6% discount rate to run a 10-year DCF. Today’s buyer is a hyperscaler—they’re not those hippies digging up earth and selling IRUs.”

Ethereum vs Bitcoin: interest-bearing assets vs digital gold

Cisco cut in half four times, up 100x: AI bubble doesn’t even count as late-stage

“Bitcoin is a store of value, and the ecosystem wants to ‘mummify’ it into digital gold. Ethereum is an interest-bearing asset, with a staking yield of about 3%.”

“BitMine’s current staking yield is about $6 million per week and $300 million per year. If ETH hits $5,000, that figure is close to $1 billion per year.”

“Our perpetual preferred stock only needs to pay $30 million in dividends per year. Staking yield alone can cover it.”

Crypto catalysts in line: CLARITY Act + Robinhood Chain + institutions entering the field

“Since the end of June, Ethereum has outperformed memory stocks by 72 percentage points. Some people lost 40% on memory, but made nearly 30% on Ethereum.”

“Robinhood Chain is built on Ethereum, not on other chains. Daily trading volume has already exceeded $1 billion, and Robinhood could make $1 billion just from this chain in a single year.”

“The CLARITY Act has advanced to the finish line—it will establish a single federal regulator for the entire crypto economy. Japan and Russia have already passed similar bills; the U.S. has to catch up.”

Gold isn’t suddenly unattractive—it just ran up too much and needs a breather

“Over the past 3 years, gold’s gains are at a level equivalent to five standard deviations within all 12 centuries of history. Of course, it needs to digest.”

“In an AI world, gold’s role as a hedge and store of value won’t change. I suggest everyone holds some—maybe 1%.”

NVIDIA’s 16x PE isn’t expensive, but memory stocks have a bullwhip effect

Tom Lee: Korea has done extremely well over the past few years—not just 2026; it’s been that way for several years. The underlying logic is that the amount of semiconductors and memory used per unit of global GDP output has been continuously rising. This means that as an economy and stock market, Korea will be far more important in the next decade than it was over the past 30 or even 50 years. Earnings should perform very well.

But over the past few quarters, the Korean market has introduced a large amount of leveraged products, which has amplified bidirectional volatility. When the market rises straight up, it traps leveraged longs inside, and then they get forced to liquidate. That’s what’s happening now: a forced deleveraging. But it doesn’t mean the underlying story is over. So I think this pullback will prove to be one of the best opportunities to buy semiconductor and AI stocks. From that, the Korean stock market, AI stocks, memory stocks, and semiconductors will ultimately create much higher highs than before.

Tom Lee: First of all, timing is never worth it. If you hold these stocks, you should keep holding. If you sell and step out now, waiting for signals before you re-enter, you’ll ultimately chase in at even higher levels. No one can time the bottom precisely.

However, the signal you want to see—large-scale deleveraging—has already happened. If you look at broker-dealer data from U.S. investment banks, the speed at which hedge funds are selling their tech-stock long positions is the fastest in the past three years—actually possibly the fastest in the past decade. They’ve already been through a massive round of deleveraging. We’re seeing some very high-profile forced-liquidation stories in Korea as well. So if someone is being forced to sell (“weak hands”), they’re already out. My view is that we’re quite close to the bottom.

But people still make mistakes: trying to guess the top and the bottom. In fact, the people who make the most money are those who keep holding. Peter Lynch said something very famous: selling the winner you hold is like cutting the flowers and watering the weeds. Charlie Munger put it better: “Money isn’t made from buying and selling. Money is made by waiting for it to come out.” If this is a structural theme involving more semiconductors and memory, you should buy it and then forget about it.

Tom Lee: Actually, it’s 16x.

Tom Lee: NVIDIA has proven it has a certain degree of recurring revenue, because of the CUDA platform, and there’s almost a certain upgrade roadmap that keeps people buying. It should be re-rated as a growth stock. I think a reasonable valuation range is between 25x and 30x.

As for memory and semiconductor equipment—these are two layers removed from end customers, and they carry bullwhip-effect risk. In simple terms, these industries are more cyclical because they don’t have purely visible order flow. Assume the end market is like a consumer using a certain AI lab service, such as ChatGPT or DeepSeek. They subscribe to an AI lab, which uses a hyperscaler. The hyperscaler buys chips from NVIDIA, and NVIDIA then orders from suppliers. The suppliers are too far from the end user. In the layers of handoffs in between, there is a lot of repeated ordering. If hyperscalers expect memory and chips to become more expensive, they may place double the quantity in advance to lock in prices. Then memory manufacturers in the future might over-expand production.

This happens in every cycle; of course there’s risk that it happens again this time. So the more cyclical the product, the lower the PE at the top of the cycle—that’s normal. You should expect PE to compress. But that isn’t a sell signal; you only need to bet that earnings forecasts will continue to be revised upward. In a machine-to-machine world, robots need far more memory and storage than humans do. Humans eat and have nervous systems; robots need memory and storage. Economies are becoming more and more memory-intensive and semiconductor-intensive. So I think earnings forecasts will keep getting revised upward. But don’t compare the memory PE to NVIDIA’s.

Tom Lee: The AI story will eventually turn into a bubble—that’s inevitable. Whenever there’s a story driven by structural demand and the market underestimates volatility, people make risk-adjustment decisions that are wrong: they underestimate risk.

Ethereum vs Bitcoin: interest-bearing assets vs digital gold

But I don’t think we’re in the late stage of an AI bubble. The reason is simple: when the stock market started dropping, most people announced the top. If it really were a bubble, people would say, “This is the bottom,” and then go crazy pouring money into semiconductors. But they didn’t—they’re selling like crazy.

Look at Cisco. From 1993 to 2000—7 years—but really it was only one cycle: the internet construction cycle. Cisco’s breakout point was $0.80. By 1997 it rose to $9, a 10x move. Then in 1997 it pulled back 40%. At the time there was the Asian financial crisis, and everyone said, “Cisco’s story is over.” What happened? By 1998, it rose from $5 to $18, double the prior high. Then something happened again in 1998—Greenspan’s “irrational exuberance” speech, Russia’s default, and the collapse of Long-Term Capital Management. Cisco fell from $18 to $9, down 42%. At the time, many people announced that tech stocks had topped. I remember that period clearly: many people danced on the graves of tech stocks saying this trade was over. Then Cisco kept rising—from $9 all the way to $80 in 2000. From 1993 to 2000, it rose a total of 100x. And from the prior high in 1998, it took only 18 months to rise 5x.

That’s when Cisco truly topped. In 2000, I was a technology analyst. Why was that top real? Because nobody believed the valuations: Cisco’s 200x PE. The underlying issue was that the fiber companies (CLECs) that laid fiber were Cisco’s real buyers. And to justify the CLECs’ valuation, you had to use a 6% cost of capital and a 30x exit multiple to run a 10-year DCF. Those assumptions were completely unrealistic. It was a farce. If someone asks whether it’s the same story today? No. Today, the number of people using AI is still small, but AI has already demonstrated very high productivity. The companies buying these devices today (hyperscalers) are extremely serious. They’re not those hippies digging up earth and selling IRUs—they’re actually buying equipment, stocking it on racks, and taking big orders. So I think we’re nowhere near the bubble stage.

Tom Lee: Honestly, the answer to this question is beyond my pay grade. What we already know is that AI is actually extremely capital-intensive. Even just maintenance costs a lot, the equipment ages, and there’s also token consumption. Open-source models are indeed cheaper, but part of the reason is they’re like generic drugs: they don’t have R&D costs, and many are distilled models. These models are open-source, but it’s impossible for them to be truly free—someone always has to foot the bill.

Second, as Elon Musk said, we’re moving toward “the singularity.” AI and robots might create such enormous productivity that everything becomes almost free. That would be disruptive even to capitalism itself. So the answer is: I don’t know. Like there’s Linux and there’s Windows, there’s Android and iOS—this logic holds. But is that negative for hyperscalers? I don’t think so. These are all very serious companies. They can choose not to participate, like Apple once did. Apple generated a lot of free cash flow that way, but it also faced criticism because it wasn’t AI-forward enough. For investors, this existential question is hard to answer. They should focus instead on where the opportunities are.

Tom Lee: Our strategy has been executing effectively this year. Our Granny Shots ETF (GRNY) has outperformed the S&P 500 by about 120 basis points since the beginning of the year, ranking in the top ten among peer funds, beating more than 92% of fund managers. The reason we can do that is because we’ve stuck to long-term themes: downstream AI, network security, and Fed monetary policy staying loose. Even if the market becomes hawkish, we still bet that the Fed will turn dovish. Small-cap performance has been unusually strong this year, and that is actually the biggest dovish signal from the Fed.

We haven’t made many changes for the second half. Earnings growth is accelerating, and we’re in earnings season now—so the AI story is very complete. But we’re willing to buy downstream: Mag 7, software, crypto. These are downstream narratives for AI, and they’ve already started to outperform.

As for the Fed, bond markets are currently very hawkish, believing the Fed must raise rates. Our bet is that inflation will be below expectations. People focus too much on oil as an inflation driver, but the oil price shock has already happened. I believe oil’s impact on inflation has already passed its peak. The underlying drivers of inflation are weakening: housing is soft, and wages haven’t truly accelerated. That will eventually put the Fed in a position to cut rates.

Tom Lee: More precisely, it’s close to $5 billion.

Tom Lee: When AI and memory stocks are rising, people criticize our fund, saying we don’t have heavy weights in AI and semiconductors. We do have exposure, but not heavy weighting. Then when memory and AI saw a 40% pullback, our fund ended up outperforming, because our anchored, longer-term ideas in the AI trade paid off.

Crypto catalysts in line: CLARITY Act and the breakout power of Robinhood Chain

Tom Lee: Let me clarify first: BitMine is the world’s largest Ethereum holdings institution, holding about 5.78 million ETH, and also the world’s largest ETH holder. But we’re actually not the first Ethereum treasury company—we’re probably the fourth or fifth—but we’re the one that became the largest. We just crossed the milestone of our one-year anniversary of operations.

We have three core differences from MicroStrategy.

First, the underlying asset. Bitcoin is a store of value, and the Bitcoin ecosystem wants it to be “mummified,” not introducing change, always doing digital gold. Ethereum is different: it’s an interest-bearing asset, with a staking yield of about 3%. Ethereum’s ecosystem is continuously developing. It’s the largest ecosystem in crypto, bigger than Bitcoin. The entire ecosystem aims to make Ethereum the financial settlement layer for Wall Street, with the whole financial rails ultimately running on stablecoins on top of Ethereum.

Second, MicroStrategy takes a relatively passive approach to Bitcoin: holding, and creating digital credit. BitMine is deeply involved in the Ethereum ecosystem. We led investments in three companies spun out from the Ethereum Foundation, all focused on strengthening Ethereum (improving price and/or strengthening the ecosystem. We’re helping shape Ethereum’s future.

Third, balance sheet complexity. MicroStrategy’s balance sheet is intentionally designed to be complex: it has convertible notes, four categories of preferred stock, and common stock—these components can compete with each other under certain circumstances. Because Bitcoin has no native yield, they have to sell shares to pay dividends. BitMine’s capital structure is extremely simple: only common stock, and we’ve recently issued perpetual preferred stock. The current staking yield is about $6 million per week and about $300 million per year. If ETH rises to $5,000, annual staking yield will be close to $1 billion. And the perpetual preferred stock only needs to pay $30 million in dividends per year. Staking yield alone can easily cover it.

So you should think of BitMine as a company highly embedded in the Ethereum ecosystem. We’re betting on Ethereum becoming not only the settlement layer for Wall Street, but also the settlement layer for communication between robots.

Tom Lee: If you can directly buy ETH and stake it yourself, that’s fine. But there are two types of investors who can’t. First, institutional investors: asset managers that manage massive capital can’t directly hold ETH tokens because that requires maintaining crypto wallets. But they can buy stocks. BitMine has been included in the Russell 1000 large-cap index and trades on the NYSE. It’s currently the only large-cap Ethereum stock available for major fund managers to buy. The Russell 1000 is the largest and most widely used benchmark index. If you’re managing a large fund out of Boston and want Ethereum exposure, you can only buy BitMine.

Second, investors who want to use options and derivatives, or investors who want higher ETH exposure. When it’s up, BitMine outperforms ETH, and it also has a rich options and perpetual contract market. In the stock-investor world, the market is $240 trillion; in the native crypto-investor world, it’s only a few hundred billion. Betting on the stock world to buy BitMine may be a better choice.

Tom Lee: What you described is what it looks like in a bear market. In a bear market, nobody talks about stocks. Back then when Apple was falling, nobody talked about Apple either. Price drives sentiment, and when crypto is in a drawdown, everyone is bearish at the bottom. That’s exactly the mechanism that forms the bottom: deleveraging and resetting expectations.

But there are plenty of crypto catalysts. Since the end of June, Ethereum has outperformed memory stocks by 72 percentage points. Someone lost 40% on memory, but made nearly 30% on Ethereum. CLARITY Act has just “crossed the finish line”—it will establish a single federal regulatory body for the entire crypto economy. The U.S. currently has regulation by individual states, so the rules are fragmented. Japan has already passed a similar version, and Russia just passed one as well. The U.S. has to catch up. This will usher in an era of institutional adoption of cryptocurrencies—a market even larger than anything in crypto history.

If you’ve been in crypto for a long time, what you’ve experienced is the “enthusiast stage,” which I call the Ethereum 1.0 era: the meme coin and NFT age. The future market is stablecoins and the payments rails. Robinhood wants to tokenize everything. They can build on any blockchain, but they chose Ethereum and launched Robinhood Chain. That’s already a breakthrough success: daily trading volume exceeds $1 billion, and Robinhood could make $1 billion just from this chain in a year—that’s a huge success for them. Every company on Wall Street is watching what Robinhood does, and then realizing: tokenizing assets on Ethereum can make a lot of money.

Tom Lee: Yes.

Tom Lee: Gold’s increase over the past 3 years is roughly at the level of about five standard deviations across the entire 12-century history. It obviously needs to digest. Maybe there’s still around 10% downside room, but that’s about it. At Fundstrat, we’ve always suggested allocating some gold—like 1%. Because no matter what happens in the future—debt uncertainty, AI’s impact on society, or social instability—gold’s hedge function as a store of value won’t change. In an AI world, this is especially important. So I think everyone should hold some gold, but it’s risen too much and needs to rest.

Original video link