The AI trade is becoming increasingly selective as investors ask which companies own true bottlenecks—and which merely benefit from temporary shortages.
KEY TAKEAWAYS
The selloff in memory stocks wasn't a rejection of AI demand. It reflected growing concern that new Chinese and Korean capacity could eventually erode the scarcity that has supported extraordinary memory pricing and profits.
ASML represents a very different kind of scarcity. Customers are trying to reserve equipment years in advance, while there is currently no equivalent second supplier for the most advanced lithography machines.
TSMC provides the next test of AI pricing power. The important question is not simply how much it earns today, but how far ahead customers are reserving advanced manufacturing capacity and what they are willing to pay.
AI bottlenecks keep migrating. The constraint has moved through processors, memory, electricity and infrastructure, and capital itself may increasingly become the scarce input.
Technology demand and investment returns are not the same thing. The CLEC and fiber boom demonstrated that demand can explode while returns collapse if formerly scarce capacity becomes abundant.
MY HOT TAKES
The AI trade isn't breaking—it is maturing. Investors are beginning to distinguish between companies benefiting from temporary shortages and businesses controlling genuinely difficult-to-replicate assets.
Pricing power matters more than simply having AI exposure. A company sitting beside a bottleneck can look spectacular until new capacity arrives and reveals that its extraordinary economics were cyclical rather than structural.
ASML may currently possess one of the cleanest examples of structural scarcity in the AI ecosystem. Multi-year customer demand is very different from profits created primarily by a temporary supply shortage.
Capital could become AI's next important bottleneck. If financing becomes more expensive or selective, companies capable of funding enormous AI investments from internally generated cash flow gain a significant strategic advantage.
Investors should stress-test AI holdings against flat capital spending, not just a crash. The companies retaining pricing power when spending merely stops accelerating may ultimately prove to be the strongest businesses in the chain.
You can quote me: “Being right about a technology and being right about an investment are two very different things–and the gap between them is where portfolios get hurt.”
Medallion. Ok, New Yorkers, raise your hand if you remember when a yellow-cab medallion was considered one of the safest investments in the city. I know that this may sound like an obscure reference, but please stay with me. Families literally borrowed against them. Banks lent against them as though they were brownstones. Drivers worked brutal hours for years–YEARS–to own one outright, because a medallion was the closest thing this city had to a guaranteed retirement plan. And the logic was airtight. The city capped the number of medallions, New Yorkers never stopped needing to get crosstown in the rain, so the value could only go one way. By 2014, a little metal badge bolted to the hood of a car was changing hands for something like $1.3 million. Think about that–$1.3 million for the right to pick people up on Lexington Avenue. Scarcity WAS the asset. Then a phone app turned "a ride" into a commodity, and here is the part people forget: nobody stopped needing a ride. The demand was real. It always was. What evaporated was the scarcity–and with it, roughly 90% of what a medallion was worth. The lesson I took from it, and one I've watched play out more than once in almost four decades on Wall Street, is that scarcity is never permanent. It just moves. I bring this up because this week the stock market got a sharp reminder of exactly that, and it came from the hottest corner of the AI trade. 🚕
Over the weekend, a Chinese memory chipmaker most investors had never heard of–ChangXin Memory Technologies, or CXMT–said it had made progress on a new chip architecture that packs more memory cells into a smaller space, and that it plans to launch a new server memory product before the end of the year. That was all it took. On Tuesday, Samsung fell about 7% and SK Hynix about 10%, dragging South Korea's KOSPI down 5.5% in a single session. In the US, Micron, SanDisk, Seagate and Western Digital all took it on the chin, and the broader semiconductor index logged its fifth straight losing session. What we saw was a commonly themed story–once investors sensed competition coming, everybody headed for the same door at the same time. 🤣
Let me translate all this, because the details matter here. DRAM (dynamic random-access memory–the short-term working memory that every computer, phone and AI server needs) has historically been one of the most cyclical businesses in all of technology. When supply is tight, prices soar and memory makers print money. When everyone builds new factories at once, prices collapse and the very same companies post losses. Boom, bust…repeat. The AI boom created a newer, more exotic product–HBM, or high-bandwidth memory, stacked right next to the AI processors themselves–along with a genuine shortage of it, and that shortage turned memory stocks–the once-poster children of commodity hardware producers–into too-hot-handle growth stocks. Their profits have been driven as much by constrained supply as by demand. And THAT is the point. Investors this week were not questioning whether AI needs memory. It obviously does. They were questioning how long memory makers keep their pricing power once new Chinese and Korean factories come online in 2027 and 2028. In fairness, Chinese producers still have real technological ground to cover in the most advanced memory like HBM, and that distinction matters–a lot. But markets don't wait for the competition to arrive. They instantly price the possibility.
Now here is what makes this week so interesting. This very morning, while memory stocks were still licking their wounds, ASML–the Dutch company that makes the extreme ultraviolet lithography machines (the room-sized tools that print circuits onto the most advanced chips) that every leading-edge chip factory depends on–reported record third-quarter sales of €11.7 billion with a 57.5% gross margin. Then it said something you almost never hear from a manufacturer: its planned supply for 2027 AND 2028 will not fully meet customer demand, and some customers are asking to place orders for 2029 and even 2030. Read that again. 👈👀 Customers are trying to reserve machines four years out. There is no second supplier of these machines. If you want to make the most advanced chips in the world, you stand in ASML's line, and you pay ASML's price. Now THAT is a medallion nobody has figured out how to copy–at least not yet.
And the next link down the chain reports tomorrow. TSMC (Taiwan Semiconductor Manufacturing–the contract factory that physically makes the chips NVIDIA, Broadcom, AMD and Apple design) is expected to post a roughly 65% jump in third-quarter net profit, after guiding revenue of $44.6 billion to $45.8 billion. TSMC doesn't own a monopoly the way ASML does, but at the leading edge it comes awfully close, and its customers have very few places to go if they want the most advanced chips built at scale. What I will be listening for is not the profit number–that is largely known already–but what management says about how long its customers are booking capacity, and at what price. That is pricing power in plain English, and it is exactly what the market is now auditing, link by link.
So the market isn't selling AI. It's reading the labels. Think of the AI economy as a chain–the machines that make chips, the foundries (contract factories) that manufacture them, the companies that design them, the memory and storage that feed them, the hyperscalers spending hundreds of billions of dollars on data centers, and, at the very end but increasingly important, the people lending the money. At every link, the question is the same: who owns something scarce, and how long will it remain scarce? Last week I wrote that the AI trade was becoming selective, with investors starting to discriminate among companies exposed to the very same theme. This week, that sorting got a lot less polite. And bottlenecks have a habit of moving–first it was AI processors, then memory, then electricity and the physical plumbing of data centers. The winners in each phase looked like geniuses, right up until the bottleneck moved down the line.
Where might it move next? Listen to Broadcom CEO Hock Tan, whose company designs custom AI chips and has said nearly all of its AI revenue comes from just six major customers. Speaking in Singapore the other day, Tan said many AI investors don't understand what they are funding and compared their expectations to a casino–his exact word, OUCH. More importantly, he and Intel's CEO argued that the next big constraint on AI may not be chips or power at all–it may be capital. That rhymes with what I wrote yesterday, when JPMorgan and Goldman Sachs moved to limit their exposure to a $5 billion AI infrastructure loan as investors demanded higher yields to hold AI-linked debt. When the scarcest input becomes money, the companies that fund their own growth from free cash flow own the medallions. The ones that need to borrow are the drivers carrying the loans.
I have seen this movie before. In the late 1990s, I was doing tech M&A during the CLEC boom (the upstart phone companies that were going to take on Ma Bell). The demand for bandwidth was real–the internet DID change everything. The fiber in the ground was real. But once everybody had laid fiber, nobody could charge for it, and the value migrated to whoever owned something that couldn't be duplicated. I sat across the table from more than a few management teams who were absolutely convinced that owning the pipe meant owning the profit. It didn't. The capacity got built, the prices fell, and the stocks fell with them–even as the traffic running through those pipes kept growing year after year. Being right about a technology and being right about an investment are two very different things–and the gap between them is where portfolios get hurt.
So what should investors be asking themselves this week? Do I own the bottleneck–or the commodity sitting right next to it? Am I paying a growth-stock multiple for earnings that might turn out to be cyclical? How many of my companies depend on the same handful of customers to keep spending? And if AI capital spending simply leveled off–not collapsed, just leveled off–which of my holdings would still have pricing power? None of these have one-line answers. TSMC reports tomorrow, and its outlook will give us the next important clue about where the scarcity sits today.
Here's the thing about a taxi medallion–the plate itself never changed. Same metal, same yellow hood, same city, same people standing on the corner with their arms in the air. What changed was whether anybody else could do the same job. That is exactly the question the market is now asking of every link in the AI chain–not whether AI is real, but whether what you own will still be scarce when the competition shows up. Scarcity is never permanent. It just moves. Make sure you know where it's moving–before the next guy with an app does. 😉
YESTERDAY’S MARKETS
Stocks finished slightly lower yesterday as investors weighed the first major bank earnings reports of the season against a selloff in memory chip stocks. The S&P 500 slipped 0.02%, while the Dow Jones Industrial Average fell 180 points, or 0.35%, to 51,670. The Nasdaq Composite declined 0.15%. Memory chipmakers fell for a second straight day on concerns about Chinese competition, and Treasury yields eased as oil prices declined.
NEXT UP
CPI / Consumer Price Index (September) may have risen to 3.6% from 3.4%.
Fed Beige Book will be released this afternoon; it will give us an anecdotal peek into economic conditions across the Fed districts. There is a wealth of information in this report for those who really want to know what’s going on.
Fed speakers today: Bowman, Barkin, and Goolsbee.
Important earnings: Progressive, Blackrock, Bank of America, Fastenal, Morgan Stanley, and State Street.