July has been an absolutely terrible month for AI markets, the worst in the industry's record by a long shot.

In fact, most of the “AI stocks” are now considerably down from all-time highs, an interesting situation considering most of them are posting their best economic results in history.

Which is to say:

We already went into my reasons why I believe markets simply misunderstand AI, in particular with respect to its inability to turn a profit, which we already disproved.

But markets aren’t totally stupid, and they “see” things that are, indeed, pretty dramatic. And worse, they aren’t yet seeing what’s to me the real danger.

This newsletter will explain in full detail what went wrong with AI markets this month and point you directly to my biggest worries. Simply put, AI is amazing and full of potential, but there are very worrying signs.

Let’s dive in.

The story begins before July. Semiconductor stocks had just completed one of the most extraordinary quarters in market history. The Philadelphia Semiconductor Index rose close to 90% during Q2, its best quarter on record.

Several companies doubled in a matter of months, while memory, semiconductor equipment, networking, power, cooling, and almost anything connected to AI infrastructure became part of the same trade.

At the same time, the trade had become extremely crowded. Semiconductor funds received approximately $12 billion of inflows over two weeks in June. Unbeknownst to many, the event was led by “smart money,” which came pounding into global semis and hardware like Winnie the Pooh on honey.

These times were like the calm before the storm. And then came July 1st.

A report claimed that Meta was considering selling some of its data-center capacity to external customers through a cloud business.

While Meta later said compute remained scarce and that third parties had offered to pay premiums for access to its infrastructure, markets obviously did not wait for that clarification.

Instead, what they actually heard was this:

The Philadelphia Semiconductor Index fell 6.3% in one session. KLA fell approximately 12%, Applied Materials 10%, Lam Research close to 10%, and Micron and SanDisk more than 10%.

The selloff continued the following day. Global markets were relatively stable, and European stocks even rose, but semiconductor stocks fell another 5.5%. South Korea’s market also suffered one of its worst sessions in years.

The next warning came from Samsung. On July 6, Samsung estimated that its quarterly operating profit had increased approximately nineteenfold. Under normal circumstances, that would have been an extraordinary result.

But Samsung’s stock fell almost 7%.

Roughly $80 billion in market value disappeared even though the company had just reported one of the largest profit increases in corporate history.

Funnily enough, this was a “logical?” reaction in principle.

Memory stocks are known to be aggressively cyclical; they have one good year in every four, literally swinging from negative gross margins three years ago to making as much money that they could buy up the company back in just four years, like Micron, which, by the way, is a one-trillion-dollar company.

This leads to fascinating behavior; when memory stocks are scoring record results, they sell, because markets are expecting the downturn to come soon.

But markets were not just punishing specific stocks; they were correcting the entire trade. Markets had already priced in shortages, higher memory prices, growing HBM demand, increasing hyperscaler investment, and rising margins. Therefore, record earnings could still disappoint when they did not imply an even faster rate of acceleration.

Hilariously irrational expectations, but that’s just the market and why you should be humble with your investments. As John Maynard Keynes once said, “The market can stay irrational longer than you can stay solvent.”

Then came TSMC, the world’s semiconductor fab, with 92% of advanced chips being manufactured on this small island 160km off the coast of Mainland China, on which China holds an historic claim (that it intends to accomplish militarily if necessary and could be carried out as soon as next year, by the way).

The company reported approximately $40 billion of quarterly revenue and a gross margin close to 68%. It guided to even higher revenue for the next quarter, raised its full-year growth outlook, and increased expected capital spending to between $60 billion and $64 billion, all extremely bullish signs.

But TSMC’s shares still fell.

ASML produced a similar result. The company increased its 2026 revenue forecast, reported extremely strong orders, and announced plans to expand capacity by approximately 30%.

But the stock also declined.

This one is particularly hilarious because it’s a literal monopoly posting nothing but record numbers; ASML is the least replaceable company on Earth. And yet, it fell on the news. The real fear was naturally whether spending would slow down (spoiler, it’s not).

And yet, everything kept falling.

Companies like ASML, or other fab equipment like Applied Materials, Lam Research, or KLA, do not depend only on semiconductor sales. They depend on semiconductor manufacturers deciding to build more factories and install more equipment.

Chip demand can remain strong for TSMC. But if TSMC doesn’t build new fabs, equipment demand can weaken.

Naturally, weaker performance for the biggest companies in the space rapidly trickled down to those stocks exposed to them. This led us to equipment stocks to take massive hits. Applied Materials and Lam Research lost approximately 30% during the month, while KLA fell close to 40% at one point; alongside memory stocks, the segments with the worst performance.

May I remind you: these companies were not reporting collapsing orders. These companies are quite literally printing money.

But if you have to pinpoint the metric that is scaring the living daylights out of everybody, that is Hyperscalers (Meta, Google, Microsoft, Amazon, and Oracle) and their “interesting” relationship they are developing at a very rapid pace with debt.

Hyperscalers’ relationship with debt is like that couple that you can see is going too fast. You know a wall is coming any day, and you can tell the situation is toxic. And well, Hyperscalers and debt are developing a relationship that could be just as toxic.

Towards the end of the month, we started to receive real data from Hyperscalers, as these companies started posting their earnings.

And the first piece of news wasn’t very encouraging. Alphabet reported another quarter of strong cloud growth, but free cash flow turned negative after capital spending for the first time in the company’s public history. At the same time, the company increased its annual capital expenditure plan by approximately $15 billion.

The market had previously treated higher hyperscaler spending as automatically positive for semiconductors. But this time, investors focused on the other side of the transaction.

Google has massive cash reserves and all (around $55 billion plus ~$180 billion in highly liquid assets), but it’s definitely a risky situation, and Google could very well have a change of heart.

In my view, investors are really starting to push for proof that there’s a return to be made in all of this while also acknowledging that Hyperscaler free cash flows can no longer sustain this entire thing.

This opens what’s clear to me as one of the main cans of worms in all of this: credit.

Back in May, I gave a speech to marketing executives, first covering the state of the AI trade, which was booming at the time, to situate them economically.

At that time, I told them directly that, if I had to look somewhere to guess when the party would be over, that should be Hyperscaler credit spreads, the difference between the cost of capital for these companies versus the US Government.

The rationale was pretty simple: at the time, Hyperscalers were financing themselves on extremely advantageous terms, almost as advantageous as the ones the USG enjoys (besides Oracle, of course).

Consequently, considering that free cash flow was already flirting with the red, cheap financing signaled that Hyperscalers didn't need to worry about dwindling cash flows, as they could always borrow very cheaply when needed.

But if the cost of debt started to rise, Hyperscalers wouldn’t have it so easy to continue spending.

Nonetheless, as of today, the five Hyperscalers (or dare I say only four, because Microsoft has yet to issue debt) have emitted $202 billion in corporate bonds. This year alone.

In other words, despite having the cash not to require debt, they were issuing debt simply because the conditions were so incredibly good. How amazing would it be to be in that position, right?

The issue with this strategy is that, at some point, debt markets would notice and decide, well, how about we ask for more? The more debt you issue, the more the costs of new debt will be for you.

And while these companies had done all that was possible to keep their pristine credit ratings, things started to change. The credit default swap market, a market that lets investors hedge the risk of default from these companies, usually inversely correlated with how much debt you have, saw CDS spreads start to rise, even above those on average investment-grade loans.

This is a fascinating graph to look at because these are some of the most liquid, most secure entities you can lend money to, and they are still trading as if they had worse credit than the average company. Absolutely nuts, which puts into perspective how “unconvinced” markets are about this entire thing.

But growing debt is hardly the only issue. Another important one with Hyperscalers (and other big companies in the trade like NVIDIA) is the famous circular financing and ‘other-income’ “situations.”

On the former, the AI market remains largely incestuous; “I give you money, you buy ‘me things’ with ‘me money’.” Not the ‘bestest’ of trends, I have to say.

The latest example is OpenAI and NVIDIA’s deal, where NVIDIA will act as a crucial financial backstop of up to $250 billion for OpenAI’s $500 billion data center (which, at current prices, is likely around 8-10GW of power).

Why? It’s pretty simple: This guarantee would let lenders underwrite the OpenAI lease against NVIDIA’s stronger credit.

And if that wasn’t enough, NVIDIA is also discussing a separate $350 billion in financing of the chips at the data center. These giant corporations are basically trading financial exposure like your kid trades Pokémon cards while the real payers, customers, arrive.

I don’t think I need to explain why this is NOT the type of stuff a healthy industry would see (we’ll get later into the reasons why this is happening in the first place).

The other element that spooked markets was the “other income” section of these companies' 10-Qs. As shown below, back in Q1, the section that holds income the company has accrued from activities separated from the company’s operations was already at almost 40% for the average Hyperscaler.

And this week, Google’s 10-Q showed that more than 70% of the income for the quarter was not from its business, but unrealized gains from its investments (mostly the SpaceX IPO, of which Google was a massive shareholder).

Particularly striking are Amazon and Google’s cases, which, if stripped of their “other income”, would have considerably inflated price-to-earnings ratios:

I’m a big Google investor myself, and I can tell you this is definitely not the type of stuff I want to see. And while all this is happening, while Hyperscalers are seeing rising costs of their AI spending, which could rise even further if the USG eliminates certain tax breaks, the only constant thing among all the change is that AI CapEx not only does not remain constant, but it keeps rising.

Nonetheless, Amazon raised its CapEx spending projections for the year to $220 billion on Thursday, adding an extra $20 billion, which looks like a small amount if not for the fact that it is 5 times larger than the average net income in the S&P 500, the 500 largest companies in the US.

Macroeconomic trends aren’t helping either. The yield on the US 30-year Treasury reached approximately 5.24% on Friday, its highest level since 2007.

Higher rates hurt semiconductor stocks in two ways.

But rates were an amplifier; they do not seem like the original cause. Semiconductor stocks began collapsing while broader markets were relatively stable.

Nonetheless, the S&P 500 is just 1% away from all-time highs as I write these words, which means people aren’t taking money out of the market; they rotated away from AI.

So, what happened?

But there was one more thing that truly illustrates the situation perfectly. One event that will go down in the history books took place this week, a lesson that markets are not to be played with no matter how smart you think you are.