DDR4 Is Now More Expensive Than DDR5: Nanya's Record Quarter Shows What the AI Boom Costs PC and Phone Buyers
Taiwan's Nanya Technology this week delivered the most profitable quarter in the company's history — an 85.7 percent gross margin and earnings per share roughly 51 times the year-ago level — without producing a single HBM chip.
Taiwan's Nanya Technology (TWSE: 2408) reported on Monday the most profitable quarter in the company's history: an 85.7 percent gross margin and earnings per share roughly 51 times last year's level. The remarkable part is that Nanya, holding roughly 1.6 percent of the global DRAM market measured by revenue, produces no HBM at all — the memory type AI datacenters are hungry for. The company is profiting precisely because the AI boom is pushing ordinary memory out of the fabs. This article explains the mechanism behind the numbers, why ordinary PC and mobile customers are paying the bill, how long the squeeze is expected to last — and how confident we can really be in the figures.
The record quarter in numbers
Nanya reported on Monday its results for July–September: net income of NT$87.6 billion (roughly US$2.74 billion) on sales of NT$113.6 billion (roughly US$3.55 billion), according to secondary reporting of the company's quarterly release. Revenue rose 61.9 percent from the second quarter and 611.7 percent from the same quarter last year (db833994).
For a company of Nanya's size, such numbers are historic. Goldman Sachs' analysis of the quarter, as reported via Futu News, estimates that Nanya's average selling price (ASP) per chip rose more than 60 percent from the prior quarter and more than 500 percent from a year ago — driven mainly by contracts with six-month to one-year terms, signed earlier when DRAM prices were lower, being renewed at market prices during the quarter. Goldman puts Nanya's operating margin at 80 percent, up 6.3 percentage points (abe9b2fa).
The paradox: record profit without AI products
Nanya produces no HBM (High Bandwidth Memory), the memory type stacked on top of GPUs in AI datacenters, and the company's president Lee has, according to the source, explicitly ruled out entering that product line. Nanya's business is conventional DRAM — DDR4 and DDR5 — for PCs, servers and consumer electronics.
Yet it is precisely this position that makes the company the clearest single illustration of what the AI buildout is doing to the memory market: HBM production is so demanding on silicon that it devours fab capacity that would otherwise supply ordinary DRAM. The supply of DDR4 and DDR5 shrinks, prices rise, and a small manufacturer like Nanya — which still has its customer base intact — harvests the gains without touching the AI market directly.
The mechanism: three times the wafer area per gigabyte
At the core of the story is a technical calculation: one gigabyte of HBM3E requires roughly three times as much silicon wafer area as one gigabyte of standard DDR5. The reason is that HBM is stacked vertically through so-called through-silicon vias (TSV), a construction that requires extra silicon (db833994).
The consequence is that every gigabyte of HBM built by Samsung, Micron and SK Hynix "costs" roughly three gigabytes of conventional memory in lost wafer capacity. Samsung executive vice president Kim Taewoo has estimated that HBM will occupy close to 30 percent of wafer capacity in 2027 (db833994).
The same point underlies the analysis of Micron by DA Davidson analyst Gil Luria, who set a price target of $3,000: HBM requires roughly three times the wafer capacity of conventional DRAM, he writes, stretching the silicon supply for all memory types (f81c9d75). He also points out that prices for conventional DRAM and NAND are rising faster than HBM prices, precisely because the three big memory manufacturers are allocating most of their resources to HBM.
The market confirms the picture — DDR4 rising faster than DDR5
If AI's consumption of wafer capacity works as the theory says, one would expect DDR4 prices to rise as hard — or harder — than DDR5. That is exactly what Barclays' price analysis shows, as cited by XenoSpectrum: contract prices for 8Gb DDR4 chips rose 4 percent month over month and 24 percent over the past three months, versus 3 and 20 percent for comparable DDR5 product. TrendForce data puts the average contract price of DDR4 8GB SO-DIMM modules — small memory modules typically used in laptops — at $142 in late August 2026, versus $133 for DDR5. The older technology was thus more expensive than its successor (94ee9304).
It is worth noting that Barclays' method of price aggregation cannot be verified from public sources, as XenoSpectrum itself points out. The comparison between DDR4 and DDR5 prices is therefore indicative rather than definitive. But the direction — conventional memory rising faster than HBM-related prices — is the same one Luria describes (f81c9d75).
The same week also brought the heaviest corroborating signal: Samsung Electronics expected, according to Quartz/Yahoo Finance, citing Reuters, third-quarter operating profit of roughly 107.4 trillion won (roughly $80.17 billion) — around nine times the year before, and, according to Reuters, the first time a technology company has passed 100 trillion won in quarterly operating profit. The memory division is the main driver (270bc9c5). This is guidance released ahead of Samsung's full accounts on October 29 — the divisional figures are not confirmed results yet.
How long will it last?
The short answer, according to available analyses: longer than most have expected.
Goldman Sachs writes, with reference to Nanya's management, that "until the fourth quarter of 2027, the chance of new DRAM supply is extremely low, and the market will remain tight" (abe9b2fa). Goldman expects the tightness to last through at least all of 2027, and both Samsung and Micron are, according to the Reuters-based reporting, counting on the gap between supply and demand lasting through 2028 (270bc9c5).
For Nanya, that means bit shipments will stay roughly flat until the company's new fab comes online. The ramp is planned in two steps: 30,000 wafers per month by the end of 2028 and 50,000 wafers per month by the end of 2030 (abe9b2fa). Until then, Nanya's earnings are in practice a price story, not a volume story.
The management also appears to be locking in prices. According to the Goldman note, long-term agreements (LTAs) now cover 50–70 percent of production, and three-year deals make up over 20 percent of volume. AI-related revenue is, according to the same source, already over 20 percent, with a target of 40–50 percent by 2029–2030 (abe9b2fa) — remarkable for a company that produces no HBM, and suggesting that Nanya is selling conventional memory into AI-adjacent systems, such as server RAM.
Who pays the bill?
The transmission mechanism ends with the consumer. When HBM steals wafer area from DDR4 and DDR5, prices rise for the memory modules in every ordinary PC, every laptop and every phone. According to the TrendForce data, DDR4 modules are now in fact more expensive per module than DDR5 (94ee9304). For PC manufacturers and consumers looking to upgrade their memory, the squeeze in practice means higher prices on older technology.
This link — AI datacenters devouring wafer capacity, ordinary customers paying — is the whole point of the Nanya numbers: they are not an AI company's success story, but a cost bill for everyone else.
How solid are the numbers?
Several caveats are worth highlighting:
- No primary source. Nanya's own financial filings are not among the available source material; all figures come from secondary reporting and a Goldman Sachs note as reported via Futu News. They should be read as reported, not independently verified.
- Different margin measures. The source for Nanya's quarterly results gives an 85.7 percent gross margin, while Goldman Sachs puts the operating margin at 80 percent (up 6.3 points). These are different measures, but the exact reconciliation between them is not documented in the sources.
- Translation uncertainty. The Goldman note, rendered in English from Chinese, alternates between referring to Nanya Technology and "South Asia Technology" as the source of management statements — probably a translation error. Statements attributed to management should therefore be treated with caution.
- Indicative price data. Barclays' DDR4/DDR5 comparison rests on a methodology that cannot be publicly verified.
- Guidance, not results. Samsung's 107.4 trillion won is a forecast. Complete, broken-out figures come on October 29.
Open questions
It remains uncertain how quickly demand for HBM itself could flatten out — and thus how hard the price waves on conventional memory will rebound. The sources document that the three big memory manufacturers expect tightness through 2027 or 2028, but they say nothing about what happens if the AI capacity buildout slows, or whether new fabs, including Nanya's own, could be built faster than planned. Nor is it documented how much of the price increase actually shows up in retail prices for PCs and phones, versus how much is absorbed higher up the value chain.
What is very clear after this week, though: a company with a 1.6 percent market share and zero HBM volume could earn nearly three billion dollars in a single quarter. That speaks plainly about how expensive AI's consumption of silicon has become — even for everyone who just wants more RAM in their laptop.