Memory capex: how money becomes supply
Memory capex becomes supply after a one-to-two-year lag, then moves prices
The three lines
- Capex flows through construction, tool installation and yield ramp before becoming supply
- In past cycles, capex surges led gluts and price crashes by one to two years
- HBM's long-term contracts pre-sell output, breaking the old surge-to-glut reflex
Key questions
- What is semiconductor capex
- Spending on fab construction and production tools. In memory, today's capex determines supply one to two years out, which makes budget announcements a preview of future output.
- Why does memory capex crash prices
- Because the supply arrives all at once, a year or two later, whether or not demand kept up. That lag between spending and output is the engine of memory's boom-bust cycle.
- Is another memory glut coming
- Not automatically. HBM ships against long-term contracts — closer to build-to-order — so its capacity does not flood a spot market. Commodity DRAM additions, though, still ride on demand forecasts.
SK hynix just raised its annual capex near 60%, to roughly $35 billion. Samsung and Micron are spending at comparable intensity. Whether that news is bullish or terrifying depends entirely on a piece of machinery most coverage skips: the pipeline by which memory capex becomes memory supply. This is a reference piece on that pipeline — written to be linked to whenever the next budget headline lands.
1. How money becomes supply
Capex divides into three streams: buildings and cleanrooms, production tools (lithography, deposition, etch), and process conversions of existing lines. The slowest stream is a new fab — from groundbreaking through cleanroom completion, tool move-in and yield stabilisation, the distance from announcement to shipped wafers typically runs one to two years.
That lag is the engine of the memory cycle. Capacity ordered in a boom arrives together during whatever demand environment exists later; if demand has cooled, inventory piles up and prices crash. Investment cut in the bust pre-books the next shortage. Capex announcements read as previews of future supply because that is literally what they are.
2. The pipeline
| Stage | What happens | Distance to market |
|---|---|---|
| Announcement | annual budget disclosed | sentiment moves immediately |
| Construction | building and cleanroom works | 1–2 years to output |
| Tool move-in | production equipment installed | months to a year |
| Yield ramp | trial runs, process tuning | months |
| Volume output | supply reaches prices | cycle complete |
| Shell fill-out | tools into an existing empty fab | lag sharply shortened |
The last row is today's key variable. Installing tools in an already-built shell skips the construction stage entirely — which is exactly what SK hynix is doing at its M15X fab, meaning this cycle's capex converts to supply faster than the historical norm.
The structural break is contracts. HBM is built against multi-year supply agreements — effectively build-to-order — so its new capacity is pre-sold rather than dumped on a spot market. The old surge-to-glut reflex still applies to commodity DRAM additions; it no longer maps cleanly onto the HBM share of these budgets.
3. What is still open
One-to-two years is a norm, not a law; line type, tool lead times and process difficulty stretch it in both directions. Vendors do not disclose how capex splits between HBM, commodity DRAM and NAND, so the ratio that matters most is an estimate. And with contract volumes and prices private, the truly decisive number — how much of the new capacity is already sold — cannot be computed from public data.
For this week's concrete case, see the SK hynix capex piece; for the product at the centre of it, the standing HBM reference.