Samsung's wafer math turns memory scarcity into a packaging boom
Samsung says HBM will claim about 30% of DRAM makers' wafer capacity next year, up from 20% now, squeezing the standard memory used in PCs, phones and ordinary servers. That scarcity is already funding the next layer down: Samsung Electro-Mechanics is spending 6.78 trillion won on substrate capacity that customers are co-financing and guaranteeing volumes for, while Montage's new client clock driver and Hanmi's packaging-tool orders show the same pricing power reaching modules and equipment.
Samsung Electronics has signaled that ordinary memory chips will stay scarce through 2027. EVP Kim Taewoo said Tuesday that high-bandwidth memory will account for about 30% of DRAM makers' total wafer capacity next year, up from roughly 20% now, according to Reuters. HBM and standard DRAM are made on the same production lines and compete for the same wafers, so a 10-point swing across the industry's three suppliers, SK Hynix, Samsung and Micron, means roughly one wafer in ten that would have made the memory in a laptop or phone a year ago is now making HBM for AI servers instead.
The memory shortage is also shifting investment and pricing power downstream, first through Samsung Electro-Mechanics' capacity plan. SEMCO is spending a combined 6.78 trillion won: 4.27 trillion won at its Sejong site on a new plant and equipment for ABF FC-BGA substrates, the largest single-product investment the company has ever made, running through May 2028 with mass production starting that September, plus 2.51 trillion won on substrate production equipment in Vietnam, The Elec reported. FC-BGA packages sit under logic chips including the ones that pair with HBM, and SEMCO will build them at Busan, Sejong and Vietnam. Global big-tech customers are co-funding part of the buildout and guaranteeing medium- to long-term volumes, similar to the way HBM capacity is sold before production begins.
Customer co-funding converts what would normally be cyclical, speculative capex into something closer to a pre-sold order book, the way HBM contracts already lock in volume years ahead. It also means the relief doesn't arrive soon. SEMCO's new FC-BGA capacity is roughly two years out. Substrate tightness is a 2027 story whether or not AI demand cools in the meantime, because nobody can pour a plant into service faster by writing a bigger check today.
Pricing power is also reaching the memory modules used in ordinary PCs. Montage announced mass production on Tuesday of its DDR5 Clock Driver Plus, a chip built to the JEDEC DDR5CKD01 specification for client memory modules such as CUDIMM, CSODIMM, CQDIMM and CAMM, running at speeds up to 9,200 MT/s across three switchable modes, per Montage's release. President Stephen Tai framed it as AI reaching client computing as well as servers. Montage sells the SPD and PMIC chips that already sit on client modules; the clock driver is new content on the same board. A day earlier the company had started mass production of a Gen5 DDR5 register clock driver for servers. Faster client memory needs a dedicated clock-buffering chip that slower modules didn't require, so every AI PC sold adds a small, captured slice of silicon revenue that has nothing to do with whether the PC itself sells at a premium.
Hanmi Semiconductor shows that the same pressure is reaching packaging-equipment suppliers. On September 2 it passed a Taiwan foundry's qualification for 2.5D packaging bonders, flip-chip and thermo-compression tools used on AI system chips, the first time a Korean packaging tool has been deployed inside a Taiwan foundry for that process. The order came in just under the roughly 28.8 billion won disclosure threshold, itself 5% of Hanmi's 2025 revenue of 576.7 billion won, for double-digit units. That followed a 44.2 billion won order in June from SK Hynix for about 15 HBM4 TC Bonder 4.5 Griffin units, deliveries running through early September. Orders from both customers show that tools for bonding chips at AI densities are being bought as quickly as the chips themselves.
Memory is a cyclical business, and this cycle has run hard. AI and chip stocks have sold off in recent weeks, the 10-year Treasury settled at 5.241% Monday, a 19-year high, and a slowdown in AI capex would loosen the assumption that HBM keeps eating DRAM capacity. A slowdown would matter over a multi-year horizon, but today's tightness is already built into production plans. Samsung's 20%-to-30% figure describes wafer allocation already under way across the industry, and reversing it takes a production-line requalification measured in quarters. SEMCO's customers are underwriting the new capacity with their own cash and multi-year volume guarantees before it is built. A capex slowdown would have to unwind commitments already signed, on top of spending that hasn't started yet.
SK Hynix, Samsung and Micron get pricing power on the DRAM that remains for phones, PCs and ordinary servers, on top of whatever HBM itself earns. SEMCO captures higher-value substrate content it can charge over a longer contracted horizon than a typical cycle. Montage picks up a new line item on every AI-capable client module sold. Hanmi adds packaging-tool orders from both a Taiwan foundry and Korea's largest HBM maker. The people footing the bill are the device makers who buy standard DRAM and the FC-BGA substrates underneath their own chips, who now have less capacity to buy from and less leverage to negotiate price.
The next data point comes in late October, when SK Hynix and Samsung report third-quarter results. The desk expects those results to show contract DRAM prices still climbing, evidence that Kim's wafer-share numbers are already showing up in what device makers pay rather than sitting in a forecast. SEMCO's own new capacity doesn't arrive until September 2028, which means the substrate side of this story has almost two full years to run before supply catches the demand its customers are already paying to secure.