Lanner earmarks almost its entire NT$10B raise to buy DRAM — the quarter memory reclassifies from component to commodity reserve.
What Taiwan's memory module makers did in the past 72 hours is rewriting the accounting logic of the industry. On June 3, Lanner (3135-TW) disclosed plans to raise NT$6-10B (US$190-310M) and said it would deploy almost the entire proceeds to buy memory inventory. The use of funds does not fit any traditional bucket — not capex, not M&A, not working capital. Lanner is issuing equity to buy DRAM chips.
The same day, Linghang posted Q1 EPS of NT$10.76, guided Q3 contracts at DRAM +30% and NAND +70%, and said order visibility now extends to 2028. Team Group (4967-TW) president Chen Ching-wen warned that SK Hynix and Micron had just lifted DDR4/DDR5 quotes 8.5-15% in a single move and that ~10% monthly hikes will continue through H2 2026, with tightness potentially lasting to 2030. At COMPUTEX, Phison CEO Pua Khein-Seng went further: 2027 will be worse than 2026 because the token-growth curve has structurally outrun NAND/DRAM capex.
All of it points to one re-categorization: memory is no longer a spec component. It is an asset.
Bloomberg's 2027 profit ranking exposes the value-chain inversion
The sharpest expression of this re-categorization is Bloomberg's June 2 forecast of 2027 global operating profit: Samsung #1 at $331B, Nvidia #2 at $329B, SK Hynix #3 at $243B — and TSMC #9 at $109B. The same TSMC that hit a record NT$2,440 on June 3 as the AI era's "toll road" earns less than half of SK Hynix's operating profit on a 2027 forward basis. The foundry premium is being crushed by the memory oligopoly's pricing power. Nanya's (2408-TW) May revenue of NT$27.67B (+730.14% YoY, +8.55% MoM) is not an outlier; it is the first quarterly data print of profit pool migrating from compute to memory. CounterPoint corroborated it from the demand side: Q1 global NAND revenue hit a record $46B, up 3.5x YoY.
Module houses turn from creditors into traders
Traditionally, Taiwan's memory module makers played two roles: buy DRAM chips from the big three, sell modules to PC/server OEMs. Margins were single-digit, cycles ran 12-18 months, and inventory was the largest P&L risk.
What's happening now is different. Lanner's NT$10B raise is not OEM demand fulfillment — it is forward speculation on the supply curve. Lanner's pitch is that AI servers consume 8x the memory of traditional servers and the big three are concentrating capex into HBM. Linghang has locked itself as a key SK Hynix-supplied module house and now claims visibility to 2028. Winbond (2344-TW) posted April EPS of NT$1.66 — 74% of its entire Q1 EPS in a single month — and limit-upped to a record NT$184.50 on June 2. The 00403A national-champion ETF holds NT$6.3B of Winbond stock.
What these companies are buying is no longer inventory for next quarter's shipments. DDR5 16Gb spot sits at $42.833 as of June 3, with Q3 contracts already published at DRAM +30% and NAND +60-70%. The accounting value of chips received in August will be 30-70% higher than the June acquisition cost. That is trading P&L, not module-business P&L. The threshold has been crossed: equity issuance is now the cheapest source of capital for buying memory.
The carbon-fee ceiling on supply
Taiwan's Environment Ministry collected NT$4.97B (US$155M) in its first carbon-fee cycle from 461 plants. Semiconductors paid ~NT$2.2B (>40% of the total), with TSMC alone responsible for 33 plants and the largest single share. Memory and foundry expansions now carry an explicit per-ton carbon cost. The marginal cost of new capacity has stepped up permanently. That is the ground beneath Phison's claim that 2027 is worse than 2026: new capacity now faces simultaneous political, fiscal, and environmental friction.
The tape is already pricing the re-categorization
TAIEX closed June 3 up 901 points at a record 46,459, and Goldman Sachs upgraded Taiwan to Overweight with a 51,000 index target. Foreign funds were net buyers of NT$10.28B for a third straight session, loading up on Wistron and Hon Hai as Nvidia-backplane names. The quieter signal is the May rebalance of the 00919 high-dividend ETF (1.26M holders): Realtek (2379) and Quanta (2382) added, UMC and Vanguard dropped. Legacy foundry out, AI backplane in.
Implication: the league table of capital sources is being rewritten
When Taiwan's module houses sell equity to buy inventory, they are sending one signal: memory has been re-categorized as a buy-and-hold asset. DDR5 spot at $42.833, Q3 DRAM contracts +30%, NAND +70%, ~10% monthly hikes, visibility to 2028 — every data point converges on the same conclusion. So long as the price per ton of chips is rising and the option value of future delivery exceeds the cost of issuance, equity is the cheapest way to fund memory accumulation.
While TSMC sits at #9 in 2027 operating profit, Nanya prints +730% YoY in a single month. That is the new hierarchy. The AI dollar flows from foundry to memory, and from memory to whoever is holding the most of it.
Key Sources: - Lanner Memory plans up to NT$10B fundraising to stockpile DRAM amid AI supercycle (cnYes, 2026-06-03) - Nanya Tech May Revenue Soars 730% YoY to NT$27.67B on DRAM Upcycle (cnYes, 2026-06-03) - Bloomberg 2027 Profit Forecast: Samsung Tops Nvidia, SK Hynix #3, TSMC #9 (cnYes, 2026-06-02) - Team Group: Memory shortage worst still ahead, tight supply may last to 2030 (cnYes, 2026-06-02) - Taiwan's first carbon fee haul hits NT$4.97B; chips pay over 40%, TSMC alone 33 plants (TechNews, 2026-06-03) - plus 5 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram