Think of memory chips as bread. The big bakeries switched their ovens to the expensive loaves that AI servers want (HBM and DDR5). The plain loaves that every television, router and car still needs are now baked by only a few small bakeries, and ESMT is one of them. Prices of the plain loaves have doubled or tripled, and new ovens take two to three years to build. That gap is the window.
Source: ESMT company filings and Q2 2026 results (31 Jul 2026); TWSE monthly revenue filings (Aug 2026); 10 Mar 2026 investor conference; TrendForce sector data. Figures marked as estimates are my own.
Why I chose it
1. The purest play on the legacy-DRAM squeeze
Samsung, SK Hynix and Micron have moved wafer capacity to the high-bandwidth memory and DDR5 that AI servers consume, winding down the older DDR2, DDR3, DDR4 and 2D NAND that everything else still needs. ESMT's entire product set is exactly those abandoned chips. The supply exit is structural, not a normal cyclical dip: replacement capacity does not arrive before 2028, and China's CXMT is walking away from DDR4 too.
In plain termsThe giants stopped making the cheap chips to chase the AI ones, and ESMT makes the cheap chips that are now scarce.
2. It captures the price faster than anyone
ESMT reprices most products quarterly and fills only 60 to 70% of its orders, so a rising price reaches its profit-and-loss within a quarter, while the large integrated makers are locked into annual or multi-year contracts. The result is visible in the numbers: gross margin went from 3.9% in early 2025 to 68.7% in Q2 2026. The company also discloses monthly earnings, which lets the market mark the cycle in near real time.
In plain termsWhen prices jump, ESMT's profits jump the same quarter, not two years later.
3. Explosive earnings on a clean balance sheet: a Cassandra inversion
My research usually looks for an equity held down by a debt overhang. ESMT is the opposite. It holds NT$12.9 billion of net cash, about 16% of its market value, with no rated bonds and its one 2023 convertible already converted to shares. There is no capital-structure problem to wait out. The discount to peers is pure cycle scepticism: the market is pricing the earnings for the cycle, not the company.
In plain termsThe cheap valuation is not about hidden debt. It is the market betting the boom won't last.
4. Management trades its own stock like an informed insider
In the same weeks, ESMT issued a NT$2 billion zero-coupon convertible that only turns into shares above NT$308, roughly 12% above the price, while a subsidiary bought about 2.2% of the company in the market at an average of NT$270. In other words, it raised money cheaply above the share price and bought its own stock below it. A team that knows the cycle better than anyone is monetising its own volatility, which is a signal worth reading.
In plain termsThe people who know this business best are selling expensive options and buying cheap shares at the same time.
What would break it
This is a price cycle on flat volume, so the risks are all about how long the price holds. Each signal below is observable in public data, and several have dates attached.
| Signal | What I would see | Where it shows first |
|---|---|---|
| The price plateau turns into a decline | Revenue rises on price, not volume. When the price stops rising, revenue stops; if it falls, the 156-day inventory and take-or-pay wafers turn the 2023 loss pattern back on. This is the dominant risk. | First non-seasonal month-on-month revenue decline |
| Q4 margin compression | Foundry Powerchip raised DRAM wafer prices 45% in July; that cost reaches ESMT's P&L from November, so Q4 gross margin compresses from the 70%+ Q3 peak. | Q3 call guidance; Q4 results |
| DDR5 loosens (the 2027 threat) | If DDR5 supply loosens as the AI server agreements are filled, the downgrade cascade that pushed demand into legacy chips reverses, and legacy demand can fall faster than supply is re-cut. This is the 2023 pattern; I carry it as a 25% bear weight. | TrendForce 1Q27 contract forecast (~late Dec) |
| Foundry allocation is lost | ESMT owns no fab; it depends on Powerchip's wafer allocation, and rivals Etron and AP Memory bid for the same wafers. Powerchip has already sold one fab to Micron for AI back-end work. | Powerchip customer-mix disclosure; Q3 call |
| China substitution and export rules | A meaningful share of demand is China-facing [end-market split unverified]. China's GigaDevice and CXMT aim to raise domestic legacy-DRAM share, and export-control regimes could disrupt flows. | GigaDevice DRAM share; annual-report geographic mix |
| The register turns | About 40% of the register is foreign money, realised volatility is 70 to 85%, and the stock was on the exchange's disposition list in August. A Taiwanese mid-cap can reverse violently on the first negative print. | Foreign-flow data; TWSE attention and disposition notices |
Known noise, not a thesis breaker: the new convertible dilutes only above NT$308, and the subsidiary's ~6.6 million shares (about 2.2%) can be resold into the market without a cancellation vote. Part of the volatility around each print is these mechanics, not the fundamentals.
How I value it
The cycle turns in H2 2027, the 2023 loss pattern returns with no penalties, and the shares fall to a trough multiple.
Prices plateau at elevated levels and wafer cost bites but does not break the model.
The shortage extends into 2027 as Samsung and Powerchip signal, and the earnings hold.
The dividend test. On my FY2026E earnings a normal Taiwanese 50% payout would be about NT$46 a share, a 16.7% yield at the current price. The board's dividend proposal, normally announced in March 2027, is the single event that would convert peak earnings into a valuation floor. A token payout would confirm the market's suspicion that the cash is a war chest for the next downturn, and would cap the shares at cycle multiples. It is the first question I would put to the company.
Entry discipline and dated checkpoints
The base value sits above the current price, but this is an 80%-volatility name with no listed options to hedge it, so the levels below frame where the risk and reward improve rather than chasing the print.
Glossary
- Fabless / IDM / foundry
- A fabless company designs chips but hires a foundry (Powerchip, SMIC, TSMC) to make them. An IDM (Samsung, Nanya) designs and makes its own in its own fabs.
- DRAM
- The working memory of every computer, phone, TV and router; it loses its content when power is off.
- DDR2 / DDR3 / DDR4 / DDR5
- Successive generations of DRAM. DDR2 and DDR3 are "legacy"; DDR4 is older mainstream; DDR5 is current for PCs and servers.
- HBM (high-bandwidth memory)
- Stacked DRAM placed next to an AI accelerator; it uses far more wafer area per bit, which is why it drains legacy capacity.
- NOR flash / NAND
- Memory that keeps its content without power. NOR is small, fast-read code storage for cars and microcontrollers; NAND is bulk storage.
- Legacy memory
- Older chip generations the majors are discontinuing but that industrial, automotive and consumer products still require.
- Downgrade cascade
- When makers cannot get a newer chip, they redesign products to use an older generation, pushing demand down into the legacy pool.
- Contract vs spot price
- Contract prices are negotiated (quarterly for ESMT, multi-year for the IDMs); spot is the daily open-market price.
- Wafer / foundry allocation
- A silicon disc on which chips are made; a fabless firm depends on how many wafers the foundry allocates to it.
- Take-or-pay
- A wafer contract that obliges the buyer to pay for capacity whether or not it is used; a hazard if demand falls.
- Convertible bond
- Debt the holder can exchange for shares at a set conversion price (NT$308 on ESMT's new bond).
- Net cash
- Cash minus total debt; ESMT has far more cash than debt, so it carries net cash rather than net debt.
- P/E and P/B
- Price to earnings and price to book value. Memory stocks show a low P/E at cycle peaks and a high or undefined one at troughs.
- TrendForce
- The Taiwan-based research house whose quarterly memory contract-price forecasts are the industry's reference.