Growth requires strength—the strength to remain steadfast and the strength to reach higher. "Ask the Cultured," the knowledge news that makes you smarter the more you listen, empowers you to become a cultured person growing alongside top SBS Premium contributors.
Contributor Lee Hyo-seok
CEO of HS Academy
Former Head of Asset Strategy at SK Securities
Former Hedge Fund Manager at Kyobo AXA Investment Managers
Rule #1 of AI Investment Is 'Volatility'.. How Conviction Fueled the Leverage Race
I think the number one rule of AI investment is volatility. We experienced tremendous volatility in our domestic market in July. It is tough and painful. But if you say, "I hate volatility so I'll stop investing in AI," then you simply cannot invest in AI. In my view, this volatility will continue until the narrative surrounding AI comes to an end.
Why is that? First, AI is just too certain. It is a technology that is unquestionably clear in every aspect—changing the world, changing individual lives, and changing corporate profits.
Second, every economic actor is desperate. "They say in the AI era you can't earn money just by working. So I have to invest now." There is a powerful FOMO (fear of missing out). For individuals, that strong FOMO turns into desperation. The same goes for companies. Why are people talking about overinvestment? Because of the fear that if they lose the AI race, their company will go under.
The U.S. government also faces its own issues. In the struggle for hegemony against the Chinese government, if they lose this game, it is over. There is a fear that they might lose global supremacy.
Individuals, corporations, and governments alike are all desperate. Meanwhile, it is certain that the world will change. So they resort to leverage. What the U.S. government considers the absolute last line is YCC (Yield Curve Control, a policy of keeping long-term interest rates within a specific range). "We tried leaving interest rates to the market to stabilize them, but people keep selling U.S. Treasuries. Let's just peg the interest rate. Make it so the 10-year yield cannot move above 4%."
In the 1940s, the U.S. government did this when it looked like they might lose World War II. When things get dangerous, they go as far as YCC. That is total leverage. From the U.S. government's perspective, it means pulling in as much money as possible. YCC is something mentioned in research related to the U.S. Federal Reserve. "What if this creates problems later?" "Does a future problem matter? If we lose here, it's over!" It is this mindset that amplifies volatility.
And this volatility won't end. It has no choice but to continue until a real AI bubble forms and bursts. But is it a bubble right now? There are rumors that even Google might issue new shares, and Big Tech companies are borrowing money. I believe Big Tech will continue borrowing much more money for quite some time. Companies are making a huge fuss and ramping up investment to seize leadership in AI.
Are Banks and Big Tech the Same? AI Investment Viewed Through 'Leverage'
A bank's core business itself is leverage. When you deposit money in a bank, from the bank's perspective, it is raising funds. With those raised funds, it issues loans. By lowering deposit rates—meaning borrowing costs—and raising lending rates—the operational yield—the bank makes money.
Usually, do savings accounts pay 1 or 2%? Hardly anything. If the borrowing cost is 2% and the lending rate is 5%, the bank puts up none of its own capital while making a 3% margin, so it would naturally want to scale this infinitely. That is why banking is a licensed industry, and regulators step in to cap approved banks, telling them, "Only leverage up to 10 times," or "Only up to 8 times" of accepted deposits. Otherwise, they would inevitably do it endlessly. This is the essence of leverage.
Big Tech is the exact same. Companies like Google and Meta raise capital much like taking deposits. They issue bonds and offer paid-in capital increases. That is all fundraising. How much does operating with that raised money yield? In other words, how much money can they make? Big Tech firms build AI data centers. Are these AI data centers actually profitable? How profitable? That is the crucial question. "So in the end, what is their borrowing cost?" If it is 5% or 6%, then when Big Tech builds and operates AI data centers, "How much return are they expecting to make?"
Q. Because the rate of return must be higher than borrowing costs to yield a profit.
Exactly. If borrowing costs and return rates become similar, loans have to be recalled. Then everyone collapses. "Let's stop investing in AI and take our money back," and everyone goes broke. When leverage turns into deleveraging, that is when the market bubble bursts. That is why the return on investment for AI data centers was so critical in the recent Big Tech earnings reports.
Amazon played a decisive role in quelling fears about returns this time. "We are spending a lot right now building AI data centers, but once construction is finished, we expect to pay back all borrowed money within three years." Recovering it in three years—so what does that mean for the rate of return? Roughly speaking, dividing the raised capital by one-third implies an assumption of a 33% annual return. And since the AI data center will still be there after three years, total earnings will exceed that. Elon Musk took it a step further, saying, "We can recoup it in one year."
So people start asking, "Just how lucrative is the AI data center business?" A borrowing cost of 5–6%? That becomes laughable. The mindset becomes, "If we can expect returns of over 30%, or even 40–50%, then of course we should raise funds." Of course, Amazon and Tesla build quickly and well, so their revenues are solid, but people wonder, "Who knows what will happen to Oracle?" or "Oracle could still go under right now."
However, the way market leaders are moving forward or their exact profit rates aren't fully precise. Since this is inferred from the three-year timeline, we cannot definitively say it is 33%, but we can assume it is extremely high. Just as banks say, "Are we making money off deposit and lending rates? Then we should do more," people are saying, "We should build more AI data centers."
"Three Years' Worth Already Sold Out": Why Big Tech Is Rushing to Expand Capacity
The next consideration is this: what AI data centers sell is called compute power. Companies like Anthropic and OpenAI have already signed three-year contracts to buy compute. From Big Tech's standpoint, they have to build as fast as they can. The data center capacity Elon Musk currently has is around 1 to 1.5 gigawatts. He needs to rapidly build up to 2 gigawatts by the end of this year and expand to 10 gigawatts next year. Right now, slightly over 1 gigawatt is under construction. But business looks so promising that he plans to finish 2 gigawatts by this year and reach 10 gigawatts next year. Once Big Tech earnings came out, the mood shifted dramatically.
Unfortunately, many people suffered losses due to leveraged ETFs, saying, "If it just goes back up, I'll sell everything and leave. I'm sick of this volatility." But the issues everyone worried about have been resolved. People said, "I'm worried because cash flow is turning negative right now," but the results showed there is no need to worry.
After Infrastructure Comes Models and Apps: The Real AI Showdown Begins Now
Q. Are there any companies showing relative warning signs among frontier AI companies like OpenAI or hyperscalers?
What I view as the most important question in the recent Big Tech earnings is: "Are the guys buying compute doing okay?" Representative companies are Anthropic and OpenAI. When the stock market underwent a major correction recently, one rumor was that Leopold (founder of hedge fund Situation Awareness) was liquidated because he used too much leverage. But when he was liquidated, the company he refused to sell until the very end was Anthropic. Anthropic is such an incredible company that he said, "I can't sell this even if I die." AI hasn't even truly started yet.
Jensen Huang | NVIDIA CEO (Jan. 21, 2026)
From an industrial perspective, AI is essentially like a "five-layer cake."
The first layer of the cake is energy. You need plenty of energy. The second layer is semiconductor chips. The third layer is infrastructure—AI data centers. AI data centers are making money right now.
At first, energy producers made a lot of money, and then chipmakers made a lot of money. Energy producers are still making money. If infrastructure companies make money, will chipmakers stop making money? They will keep earning. But now the focus is moving up to infrastructure. Massive money is being made in infrastructure.
Next will be companies that build AI models, like Anthropic and OpenAI. Going forward, AI model companies will make a fortune. Once building models is established, then come applications—things like KakaoTalk. When the mobile revolution occurred, who made the most money? Apple, Google, and Facebook. Did companies like Google or Facebook help lay down mobile infrastructure? They had zero interest in that, but once the mobile era arrived, they built applications and became the biggest earners. Anthropic and OpenAI won't just say, "I'm going to stop at building models." Naturally, they will build applications to create paid services using AI. AI hasn't even truly started yet.
The dilemma is that energy, infrastructure, and semiconductors consume far more capital than expected. People thought it would be over in two to three years, but as more AI data centers and more compute are needed, they have to spend even more money. "Even if we have to use leverage or cross-shareholding, we must keep going." They must raise capital until applications emerge. If the gap between borrowing costs and return rates narrows along the way, the bubble bursts. That is the market's current worry.
So what we should look at is: "Is Big Tech making good money?" Unless they are lying, it seems there's no need to worry. Then the question becomes, "Is Anthropic making good money?" In terms of ARR (annualized recurring revenue), Anthropic surpassed $1 billion for the first time in January 2025. One year later, in January 2026, it hit $9 billion, then $30 billion in April, and $47 billion in May—a 47-fold increase in just a year and a half. Investors are looking at this astronomical revenue growth. There is already talk that Anthropic turned a profit this quarter. Then it will raise more capital, consume all the compute, and try to maintain its top position.
What about OpenAI? Some say it is falling behind Anthropic in model competition and that its profitability is lower. However, I believe OpenAI's potential should not be underestimated—it has 1 billion users. Whenever it decides to, it can make money. Just running ads on that platform—how much would that be? It can't do that immediately, of course, because subscribers would leave. It can't do that, but having operated a massive service with 1 billion users is a competitive advantage in itself.
Chinese firms are impressive these days, but they have never operated a service with 1 billion users. OpenAI is handling all the backend work required to run something of that scale. OpenAI used to disclose its revenue, but stopped talking about it since April. Why? Because it lost to Anthropic and remains in second place. After staying quiet, it recently said, "Among April, May, and June, June's revenue exceeded the entire quarter's total." OpenAI's revenue growth rate is also steep.
Will Anthropic's IPO Be the Signal? Could the 'Real AI Bubble' Begin?
Anthropic will likely hold its IPO between September and October. Things that were uncertain will be proven by numbers, and I think this could mark the true beginning of the AI bubble. So far, stock prices have surged for companies building infrastructure. Their share prices rose in exact proportion to the money they made. The reason why neither SK
Hynix nor NVIDIA can be called a bubble is that their valuation rose precisely as much as their earnings. Valuation multiples didn't inflate; they went up strictly according to earnings.
However, AI model companies don't rise based on profits. They rise on dreams. Hard numbers aren't there yet; they rise watching the rate of revenue growth. That is a bubble. That's why when Anthropic goes public is a key focal point, but one thing I worry about is Chinese companies. Hearing talk like "they made it cheap," including Kimi K3, they say, "Let me open-source all AI models so everyone can use them." If they make them open-source, prices will drop across the board, making things tough for Anthropic. But the very first person advocating this is Jensen Huang.
Jensen Huang | NVIDIA CEO (June 1, 2026)
We open-source our models and open-source our data. We even open-source how we trained them, so you can develop Cosmos into your own model.
So it seems OpenAI is also trying to open-source some models, and Anthropic will likely start open-sourcing lower-end products. Once models are all open-sourced, infinite competition ensues. Then even more compute power will be needed. The moment infinite competition hits, even more infrastructure is required.
AI investment inevitably involves volatility. What should investors do? From now on, leveraged investments are especially off-limits. To survive the cycle heading toward an AI bubble, holding your position while investing is paramount, so it would be wise to continue investing while monitoring what the industry worries about and what bodes well.
Subscribe to the "Ask the Cultured" YouTube channel to watch more content.
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.