AI Memory & Storage

The question

Is the memory and storage complex in the early or late innings of a supply-constrained AI-driven upcycle, and are the survivors of the rerating priced for trough earnings or for mid-cycle earnings?

Base rates

The reference class is memory and storage cycles since 1990. Historically, DRAM cycles have lasted 8-14 quarters peak-to-trough with peak operating margins of 35-45% and trough margins negative. Base-rate peak-to-trough share-price moves are roughly 3-5x. HDD cycles are shallower but longer, with nearline refresh driving 2-year step-ups roughly every 5-7 years. The historical base rate for owning the surviving oligopolists through the mid-cycle phase is attractive -- roughly the 70th percentile of technology subsectors -- but the base rate for correctly identifying mid versus late is poor, roughly coin-flip.

Two features differentiate this cycle. First, HBM is a structurally tighter supply equation than commodity DRAM because wafer conversion from DDR to HBM is slow, capacity is customer-qualified, and yields improve only gradually. Second, nearline HDD capacity-per-drive is growing faster than hyperscaler unit demand, which compresses the traditional glut mechanism. Both facts argue the reference-class average understates this cycle's length.

Why the consensus view is wrong (or incomplete)

Consensus models this cycle as a normal memory upswing and extrapolates a 2027 glut using the historical 8-14 quarter clock. The causal mechanism we think that misses is that HBM demand is not a function of PC or handset refresh -- it is a function of accelerator shipments, and accelerator shipments are themselves supply-constrained by foundry and packaging capacity. That creates a capacity-capacity loop where DRAM supply cannot oversupply because its downstream customer is itself rationed. The loop only breaks when either accelerator capacity expands faster than DRAM or accelerator demand collapses.

On the storage side, the mechanism is training-set and checkpoint volume. Model weights and the associated telemetry scale faster than compute does; nearline HDD and hyperscale flash are the destination layer. Consensus underwrites enterprise refresh; it does not fully underwrite the AI payload.

Position construction

Four equal-weight positions at 25% each. The concentration is deliberate and the rationale is explicit: the hard 1-year-versus-SPY filter disqualifies most of the memory and storage cohort. What remains are the names that have already demonstrated earnings power in the current cycle and have cleared a high performance bar. Equal weighting reflects the fact that we are not trying to handicap which sub-segment wins; we are trying to own the survivors of the rerating.

MU is the HBM lever and the cleanest way to own the DRAM tightness. STX and WDC are the nearline HDD duopoly; both have generated meaningful free cash flow through the cycle and trade at low multiples on rising earnings power. PSTG is the flash-based enterprise storage franchise with a subscription economic model that damps cyclicality and provides a differentiated exposure versus the commodity complex.

Concentration acknowledgment

Four names is unusually concentrated. We accept this because a) the hard screen disqualifies the rest of the cohort, b) equal weighting across four distinct sub-segments (HBM, two HDD vendors, flash systems) provides meaningful intra-book diversification despite the small name count, and c) the alternative -- adding names that failed the screen to pad the position count -- would dilute the thesis without improving the risk profile. The left-tail protection comes from the fundamental cash generation of the survivors, not from holding count.

Asymmetric payoff

Base case: HBM tightness extends through 2027, nearline refresh continues, and the book's weighted earnings compound 20-30% over two years with modest multiple expansion. Total return roughly 30-50% over two years. Bear case: sell-side is right, glut arrives in 2027, multiples compress from mid-cycle to trough; return -30 to -40%. Bull case: HBM tightness extends into 2028 and nearline pricing firms on capacity discipline; return 60-90% over two years.

At a 60% base, 25% bear, 15% bull, expected value is roughly +15 to +22% annualized. The payoff profile is more symmetric than most of our books because memory is genuinely cyclical; we are comfortable holding it because the survivors' balance sheets are cleaner than in prior cycles and the downside is more bounded than the reference class implies.

Three things that would change our mind

1. HBM spot pricing declining more than 15% quarter-over-quarter for two consecutive quarters, with no corresponding accelerator demand signal.
2. A leading memory vendor guiding DRAM capex up more than 25% on a single call, breaking the current supply discipline regime.
3. Hyperscaler storage capex guidance showing a year-over-year decline for two consecutive quarters, indicating the AI payload thesis is not translating into nearline demand.

What we are explicitly NOT betting on

We are not buying the broader semiconductor complex -- that exposure lives in the picks-and-shovels book. We are not buying enterprise storage software or hyperconverged infrastructure; the historical base rate for those subsectors during memory upcycles is that they lag the hardware vendors, and the AI-specific catalyst is weaker. The book is a narrow bet on memory bit growth and nearline capacity, not a general bet on data infrastructure, and the concentration reflects that narrowness.