According to industry sources on the 20th, Samsung's HBM4 yield was below 60% in the early stages of mass production but has recently climbed to around 80%. HBM involves a complex process in which a defect in just one of multiple stacked DRAM chips can render the entire finished product unusable. The higher the yield, the more sellable products can be produced from the same equipment and wafer input. Samsung's push to expand high-value HBM production centered on HBM4 is seen as backed by confidence in its yields.
The next challenge after stabilizing yields is supply capacity. As the computing performance of artificial intelligence (AI) accelerators rises, demand for high-capacity, high-speed HBM is growing rapidly. If a supplier cannot deliver sufficient volume when customers want it, technological competitiveness will not translate into earnings growth.
Samsung is accordingly accelerating its HBM4 supply expansion. The company projected that third-quarter HBM4 revenue would more than triple from the previous quarter. It expects HBM4 to account for more than 60% of total HBM revenue in the second half. Three months after announcing the start of HBM4 mass-production shipments in February, related revenue surpassed $1 billion (about 1.4 trillion won). Some observers expect it to exceed $10 billion (about 14 trillion won) by year-end.
Raising the share of high-value products matters for profitability. HBM4 costs about 2 million won per chip, more than double the price of HBM3E (fifth generation). The next-generation HBM4E (seventh generation) is also expected to command a higher price than HBM4. With yields stabilized, increased output means more sellable products from the same wafer input, so capacity expansion can feed directly into revenue and profit growth.
That is directly tied to Samsung's earnings improvement, as roughly 97% of its operating profit comes from the memory business. According to financial data provider FnGuide, the average estimate for Samsung's operating profit this year stands at 375.7 trillion won as of the past month, 8.6 times last year's 43.6 trillion won. Next year's forecast is 561.2 trillion won, about 190 trillion won more than this year.
Overall HBM output is also expected to jump a notch. Samsung's HBM production scale, measured by average monthly wafer input, is projected to rise nearly 40% from about 180,000 sheets this year to about 250,000 next year. With the HBM4 family taking a larger share, the product mix will shift further toward high-value items.
Signs of the output increase are also emerging in back-end processes. Outsourced cleaning volume for glass carriers used in high-stack HBM production is expected to rise 2.5-fold, from 20,000 sheets a month this year to 50,000 next year. A glass carrier is a glass support that prevents warping or breakage when DRAM wafers for HBM are thinned. As stacking rises from eight layers to 12 and 16, DRAM must be made thinner, increasing the carrier's importance. While cleaning volume cannot be mapped one-to-one to HBM output, there is no dispute that it signals the direction of expanding high-stack HBM production.
The HBM output increase could also affect the commodity DRAM market. As more wafers go into HBM production, supply of commodity DRAM used in servers and PCs may be constrained. That is why some expect Samsung's HBM volume expansion to strengthen its bargaining power on server and PC DRAM prices.
HBM4E is another variable that will determine the pace of future capacity expansion. Samsung supplied 12-layer HBM4E samples to major customers in May and began product validation. Its yield at the reliability evaluation stage is said to have risen above 70%. The actual timing of mass production, however, could vary depending on customer certification progress and market demand. The key is how quickly the process optimization and yield improvement experience accumulated with HBM4 can be applied to HBM4E.
Market share competition is another reason for the rush to expand volume. According to Counterpoint Research, Samsung's HBM market share by revenue jumped from 21% in the first quarter of this year to 33% in the second. Over the same period, SK Hynix's share fell from 58% to 50%. The gap between the two narrowed from 37 percentage points in the first quarter to 17 percentage points in the second.
Samsung's relatively quick stabilization of HBM4 yields is attributed to a business structure that combines memory, foundry and advanced packaging. HBM4 uses 10-nanometer-class sixth-generation (1c) DRAM and a base die built on a 4-nanometer process. Because DRAM, the base die and packaging can be linked within a single company, process optimization proceeded in parallel from the development stage. As Samsung's HBM supply grows, stabilization and improved profitability in its foundry business, along with market share gains, are expected to follow.
