Famous memory analyst: A market that didn’t exist in the past is forming, and Japan’s photomask manufacturers are the biggest winners

As ongoing AI compute demand continues to drive the importance of high-bandwidth memory (HBM), the Asian semiconductor industry is seeing a wave of structural adjustments. According to reports from Korean media, Samsung Electronics and SK Hynix are accelerating the shift of internal resources toward HBM4 development while also adjusting their supply-chain strategies. In response, Citrini Research Korea analyst Jukan recently published an industry viewpoint, noting that a “new market that did not exist before” is taking shape: the HBM4 giants will outsource photomask work, and Japanese photomask manufacturers will become potential beneficiaries.

HBM4 first—photomask outsourcing frees up a new market

According to a report by the Seoul Economic Daily, Samsung and SK hynix have recently reassigned some veteran engineers previously responsible for photomask (photomask) manufacturing and inspection to HBM4-related departments, aiming to accelerate logic-chip integration and improve yields. This adjustment has led some photomask processes that were previously completed in-house to shift toward external outsourcing.

However, the two memory makers are not outsourcing everything. The most advanced EUV (extreme ultraviolet) process (2–5 nanometers) photomasks still remain produced internally, and the outsourcing scope is mainly concentrated in more mature process nodes.

Jukan’s analysis indicates that this means a “new market that did not exist before” is forming. Japanese photomask manufacturers will become potential beneficiaries, especially companies with technical accumulation in the DRAM sector, such as Dai Nippon Printing (DNP) and TOPPAN Holdings’ photomask business. He believes that even if the unit price (ASP) of these non-EUV photomasks is not high, the emergence of demand “from zero to one” in itself carries strategic significance.

SK Hynix reduces its HBM4 shipment target—yet actual output increases

On the other hand, SK Hynix is adjusting its HBM4 shipment strategy. According to ZDNet Korea, the company plans to cut this year’s HBM4 shipment volume to NVIDIA by 20% to 30% and use HBM3E and server DRAM as substitutes.

Jukan estimates that the original HBM4 shipment target of roughly 6 billion Gb could fall to about 4 billion Gb after the reduction. Using that as a reverse reference, the number of Rubin GPU units in the market this year is likely to be around 1.6 million units. Still, he emphasizes that the key is not the number of GPUs, but the change in memory supply structure.

From a process perspective, HBM4 uses a larger die area and tighter TSV (silicon via) spacing, which reduces the number of usable good dies per wafer. By comparison, HBM3E yields have already entered a mature range of above 80%. When capacity shifts from HBM4 back to HBM3E or DDR5, actual bit output increases instead.

HBM price pressure emerges, but demand could offset it

This kind of shift will translate into pressure on pricing. Jukan points out that current market expectations for higher pricing of HBM3E may be restrained by the increase in supply. However, he also offers another possibility: if sufficient HBM3E can meet AI inference-server demand more quickly, thereby expanding the overall inference TAM, then demand growth could, in turn, support prices.

In other words, the HBM market is entering a tug-of-war phase between “increased supply vs. a demand surge.”

Profit structure flips: DDR5 becomes the biggest winner

At the profit level, this round of adjustment is actually favorable for memory makers. Jukan’s analysis is that the gross margin of server DDR5 is currently significantly higher than that of HBM4, and it is even expected to exceed 90% in the second quarter—more than 20 percentage points higher than HBM3E. This means that allocating capacity to DDR5 is a more attractive choice for manufacturers.

This also leads to a key conclusion: HBM is no longer the only high-gross-margin story, as traditional DRAM has regained pricing power due to renewed demand from AI servers.

Pressure mounts for Samsung—HBM4 timeline may be delayed

By contrast, Samsung’s situation looks relatively unfavorable. Jukan notes that even if SK Hynix delays its HBM4 shipments, NVIDIA still cannot rely solely on Samsung’s capacity to push Rubin GPUs into large-scale mass production, which could ultimately lead to delays in the overall product timeline. That, in turn, gives SK Hynix time to catch up and narrow the gap between the two sides. At the same time, for Samsung, investing capacity into HBM4 also means giving up the opportunity cost of high-margin DDR5, further compressing strategic flexibility.

However, for upstream suppliers, the outlook is not optimistic. Jukan specifically called out that inspection-equipment makers that were expected to benefit as HBM4 progresses—such as DI and Unitest—may face risks of delayed revenue recognition. Because the HBM4 adoption timeline is slowing down, related equipment demand will also be postponed, putting short-term financial performance under pressure.

This article Well-known memory analyst: A market that did not exist before is forming—Japanese photomask makers become the biggest winners First appeared on Chain News ABMedia.