Korea's hidden AI race is made of glass

Samsung, LG and SK are racing to develop glass substrates for AI chip packaging, but mass production remains elusive due to manufacturing challenges. SKC's US subsidiary Absolics aims to be first to mass-produce glass substrates, but commercial output has been delayed from 2024 to 2025 and beyond. The shift from plastic to glass substrates could ease a key bottleneck in AI hardware, but meaningful volume is unlikely before 2028.

Samsung, LG and SK see glass substrates as next upgrade in AI chip packaging, but mass production remains elusive LG Innotek CEO Moon Hyuk-soo recently offered a neat summary of Korea’s next AI hardware bet. Glass substrates, he said, have largely finished development. But “a few homework problems” remain before anyone can mass-produce them. Solutions to those problems are now worth billions of dollars. Korea’s most surprising AI trade this year was not a chip. It was the firms that make the surrounding parts. Shares of Samsung Electro-Mechanics and LG Innotek have multiplied in 2026, as shortages in high-end package substrates, and in Samsung Electro-Mechanics’ case, the tiny capacitors that steady power around a chip, gave both companies pricing power investors had not expected. Investors, having learned that AI money keeps running downhill, are now betting on what comes next. It is a stranger thing than a capacitor: a sheet of engineered glass meant to sit beneath the chip itself. The boldest version of that bet does not belong to either of this year's winners, Samsung or LG. It belongs to SKC, a mid-sized chemicals group that has spent the past few years selling almost everything else it owns to chase glass. Its US subsidiary, Absolics, built a plant in Georgia for one purpose: to be the first company in the world to mass-produce a glass substrate. So far the world is still waiting. The floor the chip stands on The term hides a simple idea. Every advanced chip rests on a small platform that connects its hair-thin wiring to the coarser circuitry of the board below. Today that platform is built from layers of plastic resin. It works. But as AI accelerators grow larger and run hotter, the plastic tends to warp, the way a wooden table bends in a damp room, distorting the signals it is meant to carry. Glass does not bend like that. It stays flatter, loses less signal and can help move power and data through a denser package more efficiently. Replace the plastic base with glass, the argument runs, and one of the next chokepoints in AI hardware loosens. The trouble is in the factory. Cha Yong-ho, an analyst at LS Securities who toured the Absolics line in 2024, explained glass substrates borrow from three industries at once. “It is about 50 percent substrate, 30 percent display and 20 percent semiconductor,” he said. By his count, the process runs to some 190 steps before inspection, all performed on a material that can shatter. A microscopic crack formed early can split the part later, during testing or packaging. That is why Cha does not see glass as a separate theme from this year’s AI parts boom. “Glass substrates are the next chapter of FC-BGA, not a separate theme,” he said. Why the date keeps moving The hard part is not only drilling microscopic holes through glass. It is plating them, stacking layers, cutting panels and checking the finished product without breaking it. Clark Tseng, senior director of market intelligence at SEMI, the global chip industry group, still says the market is in an “exploratory phase.” He describes glass core substrates as “one possible solution” for future high-end packages, not a settled replacement for today’s materials. "Meaningful volume is unlikely before around 2028," he added. China is also entering the race to glass substrates through the display and PCB route. BOE has said it completed a large-area glass substrate pilot line. Visionox and AKM Meadville are also moving into the field. One Korean materials industry official said Chinese firms are trying to build the process “all at once,” which could speed up market entry. SKC says Absolics is closer than the market thinks. A sample has reportedly gone to a US chip customer for reliability testing, and the company has signaled that a pass could allow it to start preparing for production this year. The problem is that SKC has sounded close before. Commercial output was first expected in 2024, then pushed back to 2025, then to customer approval in early 2026. In its most recent filing, the company conceded it could no longer specify a mass production date. That filing also pointed to the deeper technical challenge. Absolics’ flagship design is the hard, valuable one that Im So-jung of Eugene Investment & Securities calls “subtraction and hiding”: folding capacitors and other small parts inside the glass to shrink the package and free space on top for the chips that do the computing. By the company’s own account, the approach could cut package height from 3.9 millimeters to 1.8 millimeters. It is also the version that keeps slipping. A bet SKC cannot afford to lose Whether glass ever reaches volume is not a question anyone can answer yet. What can be seen now is who has staked what. Samsung Electro-Mechanics and LG Innotek are playing a careful hand. They already hold the substrate customers from this year's boom, so they can present glass as a next step rather than a survival plan. Samsung moved late last year to form a glass-core venture with Japan’s Sumitomo Chemical, aiming for volume after 2027. LG Innotek has pointed to 2028 as its commercialization target. SKC has no such fallback, and over the past few years it sold its chemicals unit, its film unit and other businesses, one after another, reshaping the whole company around glass. This spring it raised a further 1.17 trillion won, about $840 million, from shareholders, roughly half of it bound for Absolics. Whatever happens to glass now happens to SKC. mjh@heraldcorp.com