SanDisk

Thesis

Written 2026-08-21. This is a momentum position, and I want that on the record before anything else on the page.

I bought this as a momentum trade on the AI memory cycle, straight after a sharp drop, and I have no view on the business. I have not formed an opinion on its earnings quality, its management or its competitive position, and I am not going to construct one now to make the trade look considered. It was a price and a flow, not a company. I have already trimmed it heavily at a profit, and I intend to exit the rest. It does not fit the long-term shape of the book.

What follows is the analysis I did after the fact, which is the honest order.

The trade works while capital expenditure is rising and demand exceeds supply, and it stops when that reverses. That is the whole mechanism. Memory has run the same loop for three decades: prices boom, manufacturers over-invest, supply floods in, prices collapse. Nothing about AI repeals that; it changes the size and the timing.

The scale of the build-out is the reason I do not want to own this for years. Micron is ramping capital expenditure past $25bn in fiscal 2026 and has outlined roughly $200bn of planned capacity expansion. Samsung and SK Hynix are expanding on a comparable scale. Every dollar of that is capacity arriving later to serve demand priced today, and it is the same dollar that ends the cycle. SK Hynix's M15X fab starts mid-2027; Samsung's P5 is projected for 2028.

There is a second-order risk that is easy to miss. Shortages cause customers to double-order. Backlogs then look like durable demand right up until capacity catches up, at which point they vanish at once. A sold-out order book is evidence of scarcity, not of underlying demand.

The two memory positions I hold are not the same trade. Forecasts have DRAM and NAND diverging in 2027: DRAM staying structurally short into 2028 as HBM absorbs capacity, while NAND turns to sufficiency in the second half of 2027 as new supply arrives against soft consumer demand. That puts a date on one of these positions well before the other.

NAND is the leg that turns first, and this position is pure NAND. Current forecasts have the NAND sufficiency ratio turning positive in the second half of 2027 as new capacity comes online against weak consumer electronics demand — while DRAM stays short into 2028. Of the two memory positions I hold, this is the one with the earlier expiry.

The evidence of the boom is real in the meantime. NAND pricing is projected to rise 75–100%, supply agreements are being struck with floor prices near $0.29 per gigabyte, and demand for next generation BiCS8 parts has lifted average selling prices materially. The drop I bought into was a broad risk-off move across memory and AI names rather than anything operational.

What would make me wrong, and when I sell

I have already halved this position. What remains is a question of when, not whether.

• Any sign of NAND pricing flattening, which the forecasts put in the second half of 2027 but which the market will price long before it arrives.
• Consumer electronics demand staying weak while capacity lands. That is the specific combination the sufficiency forecast rests on.
• Hyperscale spending slowing. The whole trade is downstream of one capital expenditure cycle.
• Supply agreements renewing at or near their floor prices, which would say the pricing power is already gone.

Being wrong here mostly means being late. A cyclical that has run this hard does not give much warning, and I would rather leave return on the table than test that.

Latest movements

• 2026-08-06 — Opened at $1,277.12.
• 2026-08-18 — Trimmed 50% at $1,614.19, twelve days after opening. Realized +26.4%.