Remember CXMT?
If not, don’t worry: I didn’t either. I had to use my own “Search” box to see if I’d written about the company before. As it turns out, I have. And quite recently.
That I didn’t recall an article I wrote less than three months ago testifies rather loudly to the notion that getting old “generally sucks,” as I put it earlier this week (at least I can still remember what I wrote yesterday). Or it’s a testament to the sheer amount of content I create. Or to permanent damage inflicted by over a decade spent drowning my synapses in hard liquor. (All three. It’s all three.)
Anyway, CXMT is a Chinese company that went public this year in what counted as the second-largest mainland IPO ever. They make what I could apparently use: Memory. Indeed, CXMT is the fourth-largest DRAM maker behind only Samsung, Micron and SK Hynix. The shares are viewed as one of the only pure-plays on Xi Jinping’s bid to make China a player in the global AI infrastructure buildout.
I bring the company up on Tuesday because according to itself, CXMT’s on the brink of rolling out server memory with higher-density cells. Specifically, the company says it reached a key threshold that’d theoretically allow it to put half again as many cells in the same space.
Note “theoretically.” As with a lot of claims made by Chinese tech firms that aren’t Tencent, Alibaba and Baidu, it wasn’t immediately possible to assess the veracity of CXMT’s claim. But if the company’s telling the truth, it means they hit an industry milestone before Samsung and SK Hynix — at least in terms of bringing the technology to market.
That’d be notable enough on its own, but it’s made even more compelling (or concerning, depending on who you are) in the context of restrictions which limit Chinese companies’ access to ASML’s bleeding-edge equipment.
In layman’s terms: If CXMT’s claim is true and it can manufacture this kind of memory consistently, at scale, then it’s yet another piece of evidence to support the contention that export controls won’t work to prevent China from making great leaps forward (sorry) in the AI race.
This is insult to injury for semis, which were already backfooted. Samsung and SK Hynix both dropped precipitously, dragging the Kospi down a fifth consecutive session, a stretch during which each successive daily decline was larger than the last.
The figure above shows you the SOX. It too is down five days running for a cumulative loss of more than 8%. At the lows in late-July, the SOX was down almost 30% from the prior month’s nosebleed highs.
Remember: “There’s a lot riding on the chips.” S&P earnings growth is highly sensitive to semi margins, and although Micron hit it out of the park two weeks ago, the shares are down double-digits since its earnings report.
Between them, Micron and Nvidia are seen accounting for a third of index EPS growth this year, and — get this — almost 40% of overall S&P 500 profit growth 2027.
In the latest installment of BofA’s Global Fund Manager survey, “long global semis” (which basically just means Samsung, Micron, SK Hynix and Nvidia) was again identified as the most-crowded trade.
As the figure shows, chip stocks took 45% of the vote, which the bank’s Michael Hartnett was keen to note is well below the 82% share in July, when the space was on the way to correcting sharply following a stratospheric run-up.
The good news for the “good” guys (i.e., for the US and South Korean memory oligopoly) is that there’s no guarantee CXMT actually achieved commercial production capacity for the new memory.
And as Bloomberg put it Tuesday, the company’s breakthrough anyway “shouldn’t pose an immediate threat to Samsung, SK Hynix and Micron’s dominance in HBM,” even as it demonstrates how China’s “incrementally closing the gap across the broader memory sphere.”
Those Micron and Nvidia stats are incredible.
It will be interesting to see how long it takes for sufficient supply to come online and eventually undermine the pricing power they have right now. Assuming these companies don’t operate as a cartel (a big if in today’s world), they’ll inevitably overbuild and margins will collapse. However, they likely have a few years of windfall profits unlike anything we’ve ever seen.
I still contend that the law of large numbers will put a cap on the upside in revenue and profit growth going forward, but that doesn’t change the fact that this is the most impressive profit bonanza of all time and that AI is here to stay and will fundamentally change how work gets done.
We are already seeing Nvidia’s largest customers doing an end-around on them and designing their own GPU chips. Just as Apple did before them, the largest customers want to take those margins for themselves. The hyperscalers do not need CUDA either. Since they all outsource all of their production to TSMC, there is no moat there.
Memory may have a longer runway, though the news out of China over the last week suggests that high margins, as always, are attracting new competitors.
“AI is here to stay and will fundamentally change how work gets done.”
Maybe. But so did VisiCalc, Lotus 123 and Excel. The boost to productivity took a while, but it was real and was measurable. In many ways I’d suggest that it was pretty comparable to what AI will deliver, but without the data security risks AI Agents are introducing.
My dear wife is 56 now, and I am 62. We first began to notice the lapses in our memories when we were all locked-down for COVID, perhaps because of all that time we were spending together. We now like to joke that at least between the two of us, we still “collectively” possess one fully functioning brain. I believe that one of the secrets to aging gracefully is to understand and accept your limitations in this regard. I try to never insist I am 100% right about anything anymore, and I also try to genuinely consider the advice and opinions of others much more now. It also helps to be more forgiving of yourself and others, and to be thankful for any help you may receive. In a way, being truly thankful becomes a prayer of sorts.
What are you going to do when you actually start getting old? There is a running joke between my wife and I where we are talking about someone and it almost always ends with, well, he’s dead now you know.
“The problem with most people my age is they’re already dead.”
— Yogi Berra
One of my favorite quotes of all time. Along with this one:
“If you’re not humble to this world, this world will thrust humbleness upon you.”
— Mike Tyson
The Ancient Greeks had a minor goddess called “Hubris”
For what it’s worth, AI’s strength is calculating (not surprising as computers are counting machines). It’s not good at valuing in a moral or right/wrong situation. So the mad scientists behind AI will need an algorithm or two for calculating right and wrong Good luck with that.
Some interesting rumblings in the neocloud space today. I’m curious how the hyperscalers feel about the free fall in neocloud bonds. Would the hyperscalers prefer to pick up some compute capacity on the cheap if CoreWeave or Oracle go under or would they prefer keeping them around to offload the risk of the buildout? I’m guessing they’d like to keep them around for now as I’m not sure anybody wins in the scenario where this becomes systemic and undercuts the AI trade.
Food for thought , thank you