First - before you read anything, you should read the GTC writeup by SemiAnalysis. I am very proud of the quick and professional work, and I heavily helped the writing there first. That is my technical overview of the event.
I will not repeat SemiAnalysis's (and my) work there. I believe it’s pretty much as good as it gets. There is only literally one thing I want to add as an observation. GPUs are like CPUs of old. They are improving quickly and support a broad range of accelerated workloads.
Trainium is competing against the afterimage of what Nvidia is doing. AMD is struggling to catch up to H100s, let alone Rubin. The problem with Nvidia's relentless pace is that it is ruthless. If you beat them once in a corner case of software, hardware, networking, or rack-scale, you must beat them again next year and the year after.
The number of companies with the resources to do that might be zero. That’s Jensen’s march of progress of tick-tock with rack, GPU, and networking improvements. Intel flattened the competition with this same playbook, and we see it repeat in real-time. History doesn’t repeat, but it does rhyme.
Intel peaked at a 90% market share in the Wintel area, similar to Nvidia’s today in accelerated computing. Now, on to Micron.
Micron Earnings and Margin Woes
Micron had a solid result, but shares faded. The reason why?
First, I’ll admit: Mark Murphy is the saddest-sounding person ever. Maybe this is a good method for managing expectations, but almost every question asked, "If HBM is so good, why are you so bearish on margins?" Shouldn’t the new memory paradigm be very good for margins? The answer was an emphatic maybe.
This question was asked over and over. Every time this was asked, there was a bearish pushback. Medhi Hosseini’s question almost perfectly sums it up:
I guess what we're struggling here, given all the secular trends that you detailed here, we're struggling with better understanding the longevity and earning opportunities that these secular trends are offering.
But we're just trying to understand, okay, as the mix increases to include more secular trend as capacity is allocated to higher margin, how is it going to play out in terms of operating profit? Your revenues have already exceeded prior peak with earnings, and sustainability of earnings would also happen.
Translation: Micron has exceeded past peak revenue, and you’re not willing to guide up margins in any meaningful way; it sounds like we will never see new margins high.
Mark Murphy’s answer was this dithering, but no clear “things are getting better.”
Mehdi. So the mix transformation that's occurring and the value-added content that's being added to certain products, almost a new industry being created, and the combination of these AI demand that we're seeing on HBM and low-power DRAM and so forth. But that will -- that's occurring now, and that's having a positive effect. But the overall industry profitability -- while those products are at a premium, the overall industry inventory levels are still elevated from the severe downturn that occurred.
Inventory gluts are still weighing on margins, which means they are not improving as much as they have in past cycles.
Now they have been improving. They ticked up a bit in this last quarter. But with volumes being down in DRAM and modest growth in NAND, but the -- but that volume growth continues or resumes. The shipments increased in the third quarter. And as we've said, our inventory levels tied on the leading edge, but our inventory levels overall decrease. And that's why we provide those numbers. Because that's associated, we believe, with more favorable market conditions. And in the case of DRAM, we should be below our target levels of inventory by year-end.
It’s never been this high before. And improving utilization only helps margins after you work through inventory.
So, no matter how good HBM is, inventory continues to be a drag. This was repeatedly said during the call. Meanwhile, Micron is trading at 2x+ book and is expecting a meaningful recovery. That is the setup for disappointment.
Things would be great with HBM ripping and potential share gain as Nvidia's swing producer. However, the company's lack of willingness to walk up margins confuses me. Ironically, the one place I was most concerned about has slightly turned, as NAND pricing has turned somewhat positive.
What’s New is Old in Memory
What I’m terrified of is the constraint of capacity conversations out of NAND players, which sounds quite familiar to a previous and now-defunct version of memory: HDD. NAND feels like hard drives, while DRAM is becoming closer to NAND, and HBM is replacing DRAM's previous profitability spot. There’s an odd circularity, but it's the same old industry in some ways for all the talk of new memory; it sounds a lot like the old memory.
The next step is spinning NAND companies into standalone companies (already happening) and then consolidation. As the supply consolidates, the industry becomes rational. That sounds good in an ex-YMTC world, but probably what’s new is now old. NAND is just the new HDD memory!
Cyclicals and the Economy
I admit I am more confused about cycles than I ever have been. Because of this, I will share more specific thoughts about the paywall.