Selloff of US Memory Stocks Creates a Compelling Entry Point

But isn't that a second derivative deceleration? Yes, of course, and we have highlighted over and over that this would be inevitable. Per SIA, DRAM pricing was up 70% q/q in 1q and over 40% in 2q. There's simply no possibility of extending that momentum given that memory revenues are now over $200 bn this quarter vs. $46 bn this time last year — and if it did it would be demand destructive as DRAM revenues would catch up to cloud capex. Everyone knows this, and knew this at the share price peak a few weeks ago.

Aren't the long-term agreements capping upside? Our view is that the long-term agreements are important to the stocks, but more because they provide evidence of what our checks are showing rather than because they will be hard and fast contracts in any and all environments. That said, the structure of the agreements is to some extent aiming to provide duration but cap the amplitude of the cycle. The LTA does not ensure that the customer pays the peak price, it's somewhere in between.

That said, we believe that Micron's disclosure on its earnings call that the 2q price could be the ceiling on some of the recently closed deals is something of a conservative framing. Per our industry checks, those are likely deals that were agreed in principle some time ago, with long lead times for legal sign-off, etc., and that new deals being agreed upon now will have a higher ceiling.

It's still safe to say that an LTA meet in the middle, providing both counterparties some certainty and protection, and it may disappoint those who had very high near-term earnings expectations.

What about de-speccing? We noted recently that we believe NVIDIA has cut the main LPDDR5 memory content in the racks fairly materially. Further, NVIDIA also talked about the need to drive for more memory efficiency given token growth of 10x per year dramatically outpacing memory supply, which is more than just de-speccing. There could be a general re-engineering of compute, working, and storage designed to optimize around memory constraints. That's a factor that could dampen pricing at the margin. If companies can't ship anything because there isn't enough DRAM, there is a quarter-end auction that has a powerful impact on price. But at the same time, our view is that it provides a very durable cycle view, as memory usage will certainly scale up with supply.

The underlying premise behind those potential actions is that we have a memory shortage that will last for several years, and rather than the AI business coming to a halt waiting for more memory, we will make do. But that longer-term shortage is the signal.

Higher DRAM capex will drive more supply, to be sure, but we aren't talking about a PC smartphone and server market growing 3-5% as the primary demand driver. AI spending growth north of 50% — perhaps well north — based on processor company commentary, matters more each year as AI is a bigger portion of the pie. HBM4 complexity will absorb capacity, with content doubling next year with Rubin Ultra — but there is also exceptional demand for low-power DDR5 for racks, and very high demand in enterprise storage.

In the case of NAND, capex has continued to be unbelievably benign; we do expect spending to climb next year, but not in a way that expands supply.

Our view continues to be that looking for sell signals from prior cycles misses the point. Memory is not just constrained by AI demand — memory is increasingly one of THE major primary constraints on AI demand, along with space and power. For three years, we heard others make that case, and we disagreed with it, as there was clearly slack in the DRAM supply, but that slack is gone. AI is consuming so much DRAM that there isn't enough left over for other sectors, and everywhere we look we see indications that it is a true bottleneck. It's holding back PC builds and smartphone builds. Cloud customers are paying premiums to the expected 2q price for 6-week expedites; do we think those customers are paying those premia to stockpile memory in a warehouse?

Duration is in many ways the most important debate here, rather than amplitude of peak earnings. Several years of earnings that are climbing from the current run rates are likely more conducive to high valuations than a single very strong year, and most of the actions here — longer-term agreements, customer engineering efforts — point to stronger duration.

As a result, we think that the selloff in the stocks has created a strong entry point. We think that the best value in the market comes from the compute names, notably NVDA and AVGO, but memory is catching up quickly given this deceleration, and we think this should provide a good entry point for the stocks.