July delivered the kind of market action that can quickly turn a crowded investment theme into a referendum on the theme itself.
The S&P 500 was essentially flat for the month, while the Nasdaq declined and semiconductor stocks experienced a much more severe correction. Several AI-linked semiconductor and memory names fell sharply, in some cases by more than 20%.
My view is straightforward: the selloff raised legitimate questions about leverage, positioning and expectations, but I do not believe it demonstrated that the underlying AI infrastructure cycle has broken.
The question is whether earnings, capital spending and demand deteriorated enough to justify the magnitude of the repricing.
So far, I do not think the evidence supports that conclusion.
A Leverage Failure Is Not Necessarily a Thesis Failure
One of the more dramatic examples from the selloff was Leopold Aschenbrenner’s AI-focused Situational Awareness fund.
The fund suffered enormous losses as highly leveraged AI positions moved sharply against it. Reported portfolio value fell from roughly $45 billion near its peak to around $10 billion in a matter of weeks.
I think the lesson here is important, but it is easy to draw the wrong one.
This was a cautionary tale about leverage.
When you combine highly volatile equities with extreme leverage, even a temporary correction can create forced selling, margin pressure and permanent capital impairment. Margin calls do not care whether a company’s three-year earnings outlook remains intact. They create sellers because capital has to be raised immediately.
I would not interpret that episode as proof that the AI investment thesis was fundamentally wrong. I would interpret it as proof that extreme leverage can turn an otherwise manageable correction into a catastrophic loss.
Three Forces Drove the Semiconductor Selloff
I see three primary drivers behind the pullback: concerns about Chinese memory competition, fears that hyperscaler capital spending had become unsustainable, and a momentum unwind.
Of the three, I think the momentum unwind has the best legs beneath it.
The semiconductor trade had become crowded. Once prices started falling, aggressive algorithmic strategies, retail traders, leveraged semiconductor ETFs and other momentum-oriented investors were all capable of reinforcing the move.
That creates an important distinction.
A stock can fall because its earnings outlook deteriorates. It can also fall because too many investors are trying to exit the same position at the same time.
Those are not the same thing.
There were exceptions. C3.ai, for example, entered the correction with weaker fundamentals after a significant run-up. In my view, the punishment there was justifiably so.
But across much of the semiconductor complex, I believe the magnitude of the selling was driven more by positioning and de-risking than by a comparable deterioration in underlying business conditions.
CXMT Is a Real Competitor, but HBM Is a Different Market
The second major concern involved China’s ChangXin Memory Technologies, or CXMT.
I do not dismiss CXMT. It is gaining share in conventional DRAM and represents a legitimate competitive threat in parts of the memory market.
But I think investors are making a mistake when they treat standard DRAM and high-bandwidth memory, or HBM, as if they were interchangeable products.
They are not.
HBM is one of the most technically demanding and strategically important components in advanced AI systems. Today, that market remains concentrated among SK hynix, Samsung and Micron.
CXMT can compete more directly in standard DRAM. It is not yet competing at the same scale or capability in premium HBM.
The distinction matters enormously for profitability.
If conventional DRAM pricing comes under pressure, I would expect Samsung to feel the greatest impact, Micron somewhat less and SK hynix the least, because of the relative importance of high-value HBM within their respective businesses.
The analogy I use is simple: a Honda Accord may be an excellent car, but it does not automatically replace a Mercedes for the buyer who specifically wants the Mercedes product.
Lower-cost competition can affect the fringes of a market without destroying the economics of the premium segment.
That is how I currently view CXMT’s impact on HBM.
The real question investors should be asking is how quickly CXMT can close the technology, qualification and manufacturing gap in advanced memory. That is the development I would monitor, rather than assuming standard DRAM competition automatically destroys the HBM profit pool.
The Hyperscaler Capex Panic Has a Problem: Hyperscalers Keep Spending
The next argument behind the selloff was that AI capital spending had become unsustainable and hyperscalers would soon pull back.
I understand the concern. These companies are spending extraordinary amounts of money on data centers, GPUs, networking, power and related infrastructure.
But the most recent forward guidance points in the opposite direction.
Alphabet raised its 2026 capital-expenditure outlook and indicated that spending should increase significantly again in 2027.
Meta continues to talk about maximizing available capacity through 2026 and 2027 while expanding compute resources into 2028 and beyond.
Amazon has indicated that its AI infrastructure investment is accelerating rather than slowing. Its internally designed chip business has already exceeded a $25 billion annual run rate.
Microsoft’s disclosures are also important because they challenge another bearish argument: that AI demand is concentrated among a handful of frontier-model developers.
Microsoft has reported rapidly expanding forward commitments from customers outside the major frontier labs, with longer-dated revenue schedules increasing substantially.
To me, that is a critical data point.
If AI infrastructure demand were being driven solely by a few model developers, I would be much more cautious about extrapolating current spending trends. But if demand is broadening across enterprises, cloud customers, application developers and other workloads, the infrastructure cycle becomes much more durable.
NVIDIA recently highlighted another striking forward indicator: Amazon’s commitment to deploy an additional two million NVIDIA GPUs through NVIDIA’s fiscal second quarter of 2029.
I find that difficult to reconcile with the idea that AI infrastructure spending is about to collapse.
None of this means every dollar of capex will produce an attractive return. That is a different question.
Depreciation costs will rise. Power constraints are real. Hardware becomes obsolete quickly. Some infrastructure will undoubtedly earn poor returns.
But there is an important difference between questioning the eventual return on capital and claiming that the spending itself is about to disappear.
Right now, I see little evidence of the latter.
Valuation Requires Looking at Earnings, Not the Starting Price
This is where I think behavioral bias has become especially important.
A common argument goes something like this: these semiconductor stocks have already risen dramatically, therefore they must now be overvalued.
Investors can become anchored to where a stock traded at the beginning of the year. Once the share price rises 100% or 200%, the instinct is to assume that valuation has become excessive.
But price is only one half of the equation.
If earnings and cash flow rise at the same pace, or faster, the valuation multiple can remain unchanged or actually decline.
Parts of the memory complex are trading at what I would describe as “Companies manufacturing premium memory are trading at forward P/Es even lower than utility stocks. That’s like charging Men’s Warehouse prices for Armani suits!”
Earnings growth has been extraordinary. Micron is a clear example.
Revenue and profitability have increased dramatically as AI-related memory demand and pricing have strengthened. When earnings are rising several hundred percent, a large share-price gain does not automatically imply that the stock has become more expensive on a fundamental basis.
A stock can rise 200% and still become cheaper if earnings rise faster than the share price.
That sounds obvious, but I think much of the current “overvalued” argument ignores it.
The mistake is anchoring to the original share price rather than recalculating what the business is earning today and what it may earn over the next several years.
That is a heuristic fallacy. The headline return looks extreme, so investors assume the valuation must also be extreme.
Sometimes it is. Sometimes it is not.
You have to dig into the details.
What I Expect From Here
Semiconductor stocks are higher-beta assets. They are naturally more volatile than the broad equity market, and the seasonal soft patch often lasts through September and sometimes into October.
That means I would not be surprised to see continued volatility.
But over the final two or three months of the year, I expect investors to spend more time looking at the multi-year earnings forecasts, hyperscaler spending commitments, HBM supply constraints and cash-generation potential of the leading memory companies.
If those fundamentals remain intact, I believe the valuation disconnect becomes increasingly difficult to ignore.
Corporate behavior may become important as well.
SK hynix is already buying back its own shares while signaling that management believes the stock is undervalued.
I pay attention to that.
Management teams have imperfect information like everyone else, but they understand their order books, capacity plans, customer demand and cash-flow outlook better than most outside investors.
If public markets continue to assign low valuations while earnings remain strong, I would not be surprised to see additional companies become more aggressive with buybacks.
What Matters Most Now
I do not think July should be dismissed. The selloff exposed genuine risks.
Positioning had become crowded. Leverage had become excessive in some corners of the market. Chinese memory competition is advancing. Investors are right to question whether hundreds of billions of dollars of AI infrastructure spending will ultimately generate acceptable returns.
Those are legitimate concerns.
But they need to be separated from the claim that the underlying AI infrastructure cycle is already deteriorating.
That’s speculation that’s not supported by the current data.
Hyperscaler capital spending continues to rise. Cloud commitments remain strong. Advanced-memory demand remains supply constrained. Earnings at several critical suppliers have expanded dramatically.
July may eventually prove to have been the beginning of a broader reassessment of AI economics.
It may also prove to have been a violent positioning reset inside a multi-year infrastructure buildout that remains fundamentally intact.
My focus is on the evidence that will tell us which interpretation is correct.
I am watching HBM pricing and capacity, hyperscaler capex revisions, data-center utilization, AI revenue growth, customer concentration, Chinese technological progress and corporate capital allocation.
Most importantly, I am separating price action from business performance.
Markets can be irrational in both directions. Don’t assume a stock is cheap because it fell or expensive because it rose. Understand what the underlying business is earning, how durable those earnings may be and what expectations are already embedded in the price.
That is where I believe the signal is today.
