Nanya Tech quadruples its 2027 capex - Taiwan bets big on the DRAM comeback

The third-largest global DRAM player announces a 2027 capex quadrupled. This is not a bet on HBM - it's a bet on "old-school" memory that AI is putting back at the center.

In plain terms

Nanya Technology, the third-largest global DRAM manufacturer (behind Samsung and SK Hynix, ahead of Micron in some segments), announces quadrupling its capex for 2027—a jump of several tens of billions of NT$ compared to initial forecasts. The bet: a sustainable rise in DDR5 and LPDDR5X DRAM prices, driven by AI on the server side AND by the rebound in PC/enterprise server demand.

The affair

Nikkei Asia (July 11, 2026) reports that Nanya is directing this capex towards two axes: (1) increasing production at its Nan-Ke fab (Tainan) on 1a-nm and 1b-nm nodes; (2) preparing a new building dedicated to HBM3E—a segment in which Nanya was previously absent. The context: spot DDR5 has strongly rebounded in recent quarters, long-term contracts are following with a delay but they are following.

What's happening: HBM is absorbing an increasing share of DRAM wafers at SK Hynix and Samsung. As a result, "classic" DRAM (DDR5, LPDDR5X) is structurally undersupplied. Nanya, which had almost no HBM, finds itself with an unexpected moat in the "commodity" segment—hence the prices, hence the capex.

Recap of the thread (see #1000): SK Hynix recently raised a major amount on Wall Street on the same thesis—the memory is becoming a rare asset, priced accordingly. Nanya is playing the same book, smaller and later.

Under the hood

• Target node: 1a-nm (~0.014 μm half-pitch) and 1b-nm—equivalent to Micron 1β and Samsung D1a.
• Product mix targeted for 2028 (projection, to be confirmed by the company): predominantly DDR5, a pillar of LPDDR5X, an emerging share in HBM3E.
• Capex/revenue ratio: significant increase in 2027 vs. 5-year average. Financial signal: the market has validated, the stock reacted well on the day of the announcement.
• Cycle risk: DRAM = cyclical commodity, a quadrupled capex is paid for with negative cash flow 2 years before return.

So what

Three readings. One: standard DRAM is becoming a tight asset class—each capex expansion takes 18-24 months to come out, the high price window has room. Two: Micron must respond, and probably Samsung too on the non-HBM part—expect similar announcements by Q4. Three: on the cloud/enterprise side, those who signed long-term memory contracts at the beginning of 2025 (at bottom prices) made the best deal of the decade. To watch: the movement of spot DDR5 in the next 90 days, and Micron's next announcement on its FY27 capex.