The AI Supercycle Collapse Scenario

I suddenly felt the urge, so I opened my laptop. This is a story I haven't been able to bring myself to tell until now—a raw, unfiltered piece. It’s a collection of worst-case scenarios that made me think, "Wait, something is off here," while I was writing articles. Written in one sitting, zero editing. I’m just going to write comfortably.

Hyperscalers’ Debt-Fueled Investing: A Ticking Time Bomb

Hyperscale capital expenditure (CapEx) was $357.5 billion last year. This year, it’s projected to reach $604.7 billion. In Korean Won, that’s 540 trillion and 911 trillion won, respectively. All of this astronomical amount of money is being poured into building data centers. They are dumping cash into erecting buildings, buying AI GPUs, DRAM, HBM, and energy storage systems (ESS), and laying down power equipment and cables. Their strategy is to go all-in right now and recoup everything later with AI.

Next year’s CapEx estimate increases even further to between $900 billion and $1 trillion. It makes sense, given that AI usage is continuously rising. $1 trillion is 1,500 trillion won. Since South Korea's annual budget is around 730 trillion won, this is more than double that amount.

This massive pile of money is sourced through existing cash and borrowing from others. The former is Free Cash Flow (FCF), and the latter is corporate bond issuance. The combined FCF of hyperscalers was $60 billion in the fourth quarter of last year alone. However, in the first quarter of this year, it chopped down to one-third, hitting the $20 billion level. It is a foregone conclusion that it will turn negative next year.

Hence, the debt-fueled investing has begun. Just last year, five major hyperscalers issued a total of $121 billion (177 trillion won) in corporate bonds. This was four times the past five-year average of $28 billion—the largest scale in history. This year, they will increase it even more to hit that $1 trillion mark. And people will willingly buy those bonds. Why? Because these are ultra-blue-chip companies that rank within the top ten in global market capitalization. Their credit ratings are exceptionally high. Whether AI actually makes money later or not isn’t an immediate problem. That’s exactly what they are doing right now.

Rising Interest Rates, Growing Burdens

As hyperscalers run out of cash and their reliance on borrowing grows, they become increasingly sensitive to interest rates. In other words, if interest rates go up, they have to tack on more interest for corporate bond buyers. And this is happening at a time when they already have countless places to spend money.

The possibility of rate hikes is growing larger by the day. This is due to concerns over U.S. inflation driven by Trump's tariffs and the prolonged war involving Iran. Initially, the market expected the Fed to cut the benchmark interest rate by 0.25% by the end of this year. But now, the probability of a rate cut has dropped to 0%. Instead, the probability of a rate hike has surged to 58%.

The new Fed Chairman, who is classified as a hawk (monetary tightener), is also a risk. In the past, the Fed utilized its balance sheet to buy massive amounts of government bonds and Mortgage-Backed Securities (MBS), injecting dollars into the market. As a result, interest rates were artificially lowered, and asset prices rose. It was asset market management that went beyond standard monetary policy.

The new Fed Chairman, Kevin Warsh, is basically someone who refuses to do this. He will only take responsibility for price stability and the financial system; he won't engage in propping up asset prices. If things go this way, the intensity of the Fed’s intervention will be low when the asset market shakes. Meaning, even if AI investments and the semiconductor supercycle are in jeopardy, the Fed won't tick-tock respond with policy fixes like they did in the past. This is because saving U.S. inflation, employment, and the overall financial system is more important than saving hyperscalers.

In short, the era where they built up CapEx by borrowing money at cheap rates is ending, and an era of rising interest rates and a non-intervening Fed is coming.

"AI... Does That Actually Make Money?"

But what if, in this exact situation, a catastrophic disaster happens where <<>>? I believe this will act as the trigger that collapses the fragile AI supercycle. Isn't it obvious? There is no scenario more mind-boggling than taking on massive debt to finish all investments, only to find out that the revenue is lackluster when you actually run the AI.

Yet, warning signs have been detected all over the place since the end of last year. While overall revenue is growing, the growth rate of AI revenue—the very place they poured cash into—has slowed down or even decreased. Let's look at an example. Microsoft's stock price plummeted 9.99% the day after its Q4 earnings announcement last year. It was the largest drop since March 2020. On that single day, $350 billion in market cap evaporated.

Revenue was $81.27 billion, up 16.7% year-over-year, and earnings per share (EPS) was $4.1. Both beat market expectations. However, revenue growth for Azure, their AI cloud service, stopped at just 38%. An item that used to grow by 40% to 50% every quarter finally slipped below 40%. Furthermore, Q4 CapEx was $37.5 billion, a 66% increase year-over-year. This means that while revenue grew by 16%, they increased investment by 66%. Paid users for Copilot, which integrates AI into software, stood at 15 million. It sounds like a lot, but this is actually just about 3% of Microsoft 365 subscribers (450 million).

Shareholders thought: "Is this really the right return on investment?"

The Downfall of Samsung and Hynix Happens When CapEx is Cut

So, how will this affect Samsung Electronics and SK Hynix? The reason Samsung and Hynix's stock prices are rising right now is because of memory semiconductors. The structure is that HBM made by these two companies is loaded into Nvidia's AI GPUs, and those AI GPUs go into hyperscalers' data centers. Samsung and Hynix's DRAM and HBM are also loaded into the custom chips produced internally by hyperscalers.

In short, memory demand is overflowing. Because only three companies—Samsung, Hynix, and Micron—make it, a supplier-dominant market has formed. Memory became a seller's price market, and Samsung and Hynix's revenue and operating profits began to skyrocket. This is the basic logic driving the assumption that Samsung and Hynix's stock prices will continue to rise.

But let's think about it. The fact that Samsung and Hynix are selling memory at increasingly higher prices means that hyperscalers are buying memory at increasingly higher prices. They have to expand their already monstrous CapEx even further. But money doesn't just fall from the sky, and there are limits to FCF and corporate bond issuance. Up until now, they competitively increased CapEx out of fear of falling behind if they didn't buy, but now they might start wondering, "Can we actually handle this price?" In other words, controversies regarding the efficiency of AI infrastructure investment will begin to arise.

Even if it's not strictly because of memory prices, a moment will eventually come when hyperscalers say, "We’ve built enough. We'll cut back on investments now and start making a profit." In both cases, CapEx decreases, and memory demand decreases. Naturally, the earnings of suppliers will take a hit. This includes Nvidia, which makes AI GPUs, and Samsung, Hynix, and Micron, which make HBM and DRAM. The countless semiconductor material, component, and equipment (Sojoobang) companies linked to them, the secondary battery companies selling energy storage systems to data centers, the companies laying cables, and so on, will all start to see their stock prices fall.

Think we still have a long way to go? There is a year and a half left at best. Securities firms' forecasts for Samsung and Hynix's revenue and operating profit show them skyrocketing through this year and next, only to turn negative starting in 2028.

Let’s summarize.

Hyperscalers at the absolute peak of the AI supercycle are pulling off mind-blowing CapEx. In the midst of this, benchmark interest rates have gone up, increasing the interest burden. The government says they won't take responsibility even if the business goes under. To make matters worse, memory prices keep getting more expensive. Let's cut back on CapEx. It’s not like we can increase it forever anyway. "Thanks for everything. We don't need memory anymore."............ Now, let's start making some money!!

"Boss, but it's not making as much money as we thought???"

Closing Thoughts

There is an old adage in the stock market: when everyone can explain the reason why a stock price is rising, that is the peak. I feel like that time is right now.

I am not an AI pessimist. Rather, I think AI can replace humans in virtually all fields. However, AI changing the world and companies making money through AI investments are two completely different stories. Right now, investors seem to treat the two as the same thing. This is exactly why I wrote this piece.