Unpacking Gavin Baker's "Memory Crunch" Thesis

Atreides CIO Gavin Baker argues the memory “crunch” will persist, pointing to a hardware bottleneck for on-device AI (e.g., Apple needing more DRAM), constrained supply (TSMC disciplined capacity vs prior oversupply gluts), and rising Agentic AI memory-retrieval demands. Micron’s momentum is cited as confirmation, with 345% YoY revenue growth and a beat on EPS expectations, while Wall Street expects triple-digit EPS growth through next year. The article also notes memory names (Micron, SanDisk) and the Roundhill Memory ETF are retreating toward the 10-week moving average, framing it as a potential re-entry/buy zone.

Analysis

The cleanest expression is not a generic AI beta trade; it is a memory-scarcity trade versus the rest of semis. If DRAM/NAND pricing stays tight, MU and SNDK should keep converting revenue into margin at a pace that typical semiconductor multiples do not fully reflect, while NVDA/ALAB/COHR face a less obvious second-order headwind: higher memory content raises system cost and can delay end-customer deployments at the margin. AAPL is the most plausible loser on the device side because more local inference increases component intensity before it increases monetization, which can compress gross margin or force a slower feature rollout.

The key catalyst path is over the next 1-3 months: spot memory pricing, channel inventory, and commentary from hyperscaler capex budgets. If there is no inventory buildup and TSM continues to ration capacity, the cycle can stay extended longer than the market expects; if one large cloud customer pauses orders, the tape can turn fast because memory equities still trade like commodities, not annuity streams. Six to 18 months out, the structural risk is that supply response catches up and collapses the spread between price and cost, which would hit the highest-beta names first.

Contrarian view: consensus may be underestimating how much of this move is already in the stocks. The stronger the rally, the more vulnerable MU/SNDK become to any hint of EPS revision deceleration or guide conservatism. What would falsify the thesis is a sustained rollover in DRAM/NAND spot prices, TSM signaling less disciplined capacity, or a visible slowdown in hyperscaler capex. At current levels, the better risk/reward is to buy pullbacks in the strongest balance sheets rather than chase strength after gap-ups.

Key Decisions for Investors

• Long MU on a retest of the 10-week moving average; trade for 1-3 month momentum with a tight stop if DRAM spot pricing weakens or the next guide fails to raise expectations.
• Pair trade: long MU / short AAPL for 1-3 months as a relative-value expression of rising memory intensity versus device-margin pressure; exit if Apple shows stronger-than-expected on-device AI monetization.
• Hold or add to SNDK only on pullbacks, not breakouts; use the position as a higher-beta expression of the same scarcity theme, but keep sizing smaller than MU because NAND is more prone to abrupt inventory correction.
• Underweight NVDA relative to MU over the next earnings cycle if memory costs keep rising faster than AI pricing power; the risk is not demand destruction but gross-margin ceiling risk.
• Watch TSM commentary and hyperscaler capex prints as the main falsifiers; if either turns less disciplined, trim memory exposure quickly because the rerating can unwind within days.