Nvidia is pushing suppliers to make glass chip bases

What happened

On September 24, 2026, Seoul Economic Daily reported that Nvidia is weighing glass substrates as a way to put more HBM on each AI package. The English piece by Park Si-jin says German equipment maker SCHMID is collaborating on glass-core tools with suppliers in the Intel, Nvidia, and AMD chains, citing a first-half earnings presentation that Investing.com carried. Roland Rettenmeier, SCHMID’s chief strategy officer, said metallization of through-glass vias is still a hurdle and that final customer qualification has not been completed. SCHMID sells the equipment. It is not the company that makes the base.

The same article states that glass raises chip speeds by about 40 percent and cuts power use by roughly half compared with plastic. That is a materials claim in the piece. It is not a measurement from a Nvidia system.

The manufacturing attempts sit with substrate makers, and the public record on them is uneven. NIST’s CHIPS page lists Absolics, a subsidiary of Korea’s SKC, as recipient of up to $75 million in direct funding for a 120,000 square foot plant in Covington, Georgia, with expected capital spending of $343 million, to make glass substrates for advanced packaging. The award page says first customer deliveries were expected in 2025 and production capacity in 2027. Those are the dates printed on the award. They are not a 2026 receipt. TrendForce on September 23, citing ZDNet, still describes Absolics as running evaluations and customer qualification, and says SKC is adding about 92.1 billion won, for a stated total commitment of about 373.1 billion won.

Samsung Electro-Mechanics is the other large Korean substrate house in this lane. It is not Samsung’s memory business. On July 2, 2026, Sumitomo Chemical said its Korean unit Dongwoo Fine-Chem signed a joint venture with Samsung Electro-Mechanics to develop and sell glass-core substrates. The planned company, GlaSSEM, would be owned 66 percent by Samsung Electro-Mechanics and 34 percent by Dongwoo Fine-Chem, based in Pyeongtaek, and is supposed to have a supply system in the second half of fiscal 2027 if approvals land. TrendForce, citing Aju News, says prototypes are already coming off a pilot line in Sejong. Read 2027 as the company’s schedule.

LG Innotek, on its own site in September 2025, said prototype production of glass substrates was underway and that mass production was targeted for 2027 to 2028. TrendForce’s September 23 synthesis, citing Aju News, puts the mass-production aim at 2028 and says the company showed the technology at KPCA Show 2026.

China’s BOE is further from a primary document. The same TrendForce piece, drawing on Aju News, IT Home, and ZDNet, says chairman Chen Yanshun has spoken about investing in a glass-substrate line, that samples are in customer checks with server processors, and that full-scale production is not expected before 2028. That is second-hand trade press, not a filing.

TSMC is not making Nvidia’s glass base. Seoul Economic Daily says Jensen Huang has encouraged TSMC to develop the technology. Encouragement is not a line. TrendForce in June said TSMC is focused on Chip-on-Panel-on-Substrate, or CoPoS, using a 310 by 310 millimeter panel, with 2026 as a validation year, pilot production aimed at 2027, and mass production aimed at the second half of 2028. That note puts commercial glass-core substrates later, likely after 2030. Analyst Ming-Chi Kuo wrote on June 11 that industry checks place glass carriers and glass-core panels inside the CoPoS discussion, with mass production hoped for the second half of 2028 and Nvidia a possible early user. The clocks do not agree, and this story does not pick one of them as fact.

Intel’s research track is older, and it is Intel’s. On September 18, 2023, Intel said it had been studying glass substrates for more than a decade and aimed to offer them in the second half of this decade. The release shows test pieces at the company’s Chandler, Arizona, site. Keep that 2023 roadmap away from anything Nvidia has bought. Beside the Georgia factory, Georgia Tech said on December 12, 2024, that Absolics and Georgia Tech’s 3D Packaging Research Center would receive $100 million under the CHIPS National Advanced Packaging Manufacturing Program for glass-core panel research. That award is research money. It is separate from the $75 million factory award.

The buyer pressure is real and still short of an order. Park reports that Huang met SK Group chairman Chey Tae-won to discuss semiconductor cooperation that included glass substrates. Chey chairs the group that owns SKC, Absolics’ parent. SK hynix, a different company in that group, still makes the HBM. The meeting does not name Absolics as a Nvidia supplier. TrendForce, citing KIPOST, reports that Nvidia has asked substrate makers to move faster. An ask is not a purchase order.

A different substrate problem stays in its own paragraph so it does not get merged into this one. TrendForce reported on September 23 that industry sources put Intel EMIB substrate yields at roughly 45 percent over the past quarter, up from about 30 percent in the second quarter of 2026. EMIB embeds silicon bridges in the package substrate. It is not a glass core. TrendForce still expects CoWoS-L to remain the mainstream AI packaging technology through 2028. In that same EMIB report, Intel chief executive Lip-Bu Tan called substrates a bottleneck and said Intel has to prepay to hold suppliers’ attention.

Why it matters

The useful mistake is treating a tool announcement, a foundry panel plan, and a research sample as the same object. The companies that would manufacture a glass core are on prototypes, pilot lines, and their own calendars. TSMC’s near-term packaging step, in TrendForce’s June account, is a larger panel. Intel’s Chandler pieces are a 2023 research track with its own clock. Nvidia can lean on all of them because it buys the packages. The EMIB yield gain is a separate substrate headache, and it should stay in that chapter.

What to watch is a named Nvidia product tied to a named substrate maker, a public customer result from Absolics, Samsung Electro-Mechanics, LG Innotek, or BOE, whether TSMC’s 2027 CoPoS pilot is described as a glass core or as a panel on a different core, and whether Intel’s language about the second half of the decade becomes a product with a customer.

Korean HBM supply beside these packages is Samsung and SK hynix racing capacity. The packaging path still called mainstream is TSMC’s CoWoS allocation.