TSMC, the world's largest contract chipmaker (foundry), has for the first time disclosed validation data for glass substrates aimed at next-generation AI-chip packaging — a move that lands squarely on the first-mover thesis Korean materials suppliers have been building for years.
According to a supply-chain briefing reported on June 16 by DigiTimes, a Taiwanese tech-industry outlet, and relayed by Korea's Chosun Biz, TSMC ran joint simulations with Ibiden, a Japanese ABF-substrate maker, and Innolux, a Taiwanese panel maker, on applying a glass core substrate to its CoPoS (chip-on-panel-on-substrate) advanced-packaging platform — the next-generation panel-level successor to CoWoS.
What TSMC actually showed
The headline question for any chip-supply-chain investor is simple: how close to production-grade is this? The disclosed numbers are the answer. Using a 0.8mm glass core substrate tested at a 5x-reticle chip-on-wafer specification and an 85×110mm package size — dimensions consistent with a large AI GPU package — TSMC reported that the glass substrate improved package coplanarity (a measure of how flat the package surface stays, i.e. resistance to warpage) by 16%, cut the effective coefficient of thermal expansion by 19%, and raised effective elastic modulus by 31%. On the power-integrity side, resistance fell 27% and inductance fell 42%. Those figures are confirmed across DigiTimes' reporting and a verbatim breakdown carried by TradingKey.
TSMC also said its test samples showed none of the micro-cracking or delamination that has long been glass's signature weakness in large packages — a result Chosun Biz's industry sources read as a meaningful step toward commercial viability.
Why does the metallurgy matter for the AI roadmap? Because the package itself is becoming the bottleneck. Substrate sizes have grown from roughly 90×90mm in 2020 to 120×120mm in 2023–2025, with a 120×150mm class expected around 2028, per figures cited by Chosun Biz. Nvidia's Blackwell GB200 used an 81.5×74.8mm substrate; the Vera Rubin VR200 now nearing production moves to 100×91mm. As packages enlarge, conventional organic substrates warp under heat and struggle with power delivery — the exact failure modes glass is meant to fix.
The pressure on Korea’s contenders
Until now, the glass-substrate front-runners were widely seen as Intel (which has pursued the technology for more than a decade) and a cluster of Korean players. Samsung Electro-Mechanics, the components affiliate of the Samsung group, is running a pilot line at its Sejong site and is targeting mass production in the second half of 2027 while in quality evaluations with potential customers including Broadcom, according to TrendForce. SKC (011790.KS), a Korean chemicals-and-materials group, has built a U.S. base through its subsidiary Absolics in Covington, Georgia, and TrendForce describes it as the most aggressive entrant, already supplying prototype substrates to a U.S. telecom-chip customer.
LG Innotek (011070.KS), Korea's largest camera-module and substrate maker and the primary company in this news cluster, operates a glass-substrate pilot line at its Gumi plant and had targeted commercialization around 2027–2028, per Chosun Biz; TrendForce notes it is partnering with precision-glass firm UTI to strengthen mechanical durability but gives no fixed mass-production date.
The strategic worry is ecosystem control rather than the data sheet alone. A Korean packaging executive quoted by Chosun Biz argued that TSMC validating glass with Ibiden and Innolux signals an attempt to set the next AI-packaging standard inside a Taiwan–Japan supply chain first — the same playbook that made TSMC's CoWoS the default for Nvidia and left the industry capacity-constrained as AI demand surged. If qualification and large-customer certification coalesce around TSMC's partners, Korean suppliers could hold the technology yet still struggle to win designs at Nvidia, AMD or Broadcom.
What the disclosure does not settle
Mass production is not imminent. TSMC chairman C.C. Wei said the company has stood up a pilot line and expects large-scale production within two to three years, and Chosun Biz reports TSMC told suppliers it still needs further work on glass thickness and multi-die layout. Industry analysts cited in the same report say the through-glass-via (TGV) challenge — drilling and copper-filling tens of thousands of micro-holes to carry signal and current vertically through an insulating glass core — is not fully solved. SEMI, the global semiconductor industry association, and research firm Global Net Corp. projected in a May 27, 2026 report that glass core substrates would reach only limited-volume production around 2028, before a 67.2% compound annual growth rate from 2028 to 2040.
Government backing remains a tailwind for the Korean side: on June 17, Koo Yoon-cheol, Korea's Deputy Prime Minister and finance minister, toured LG Innotek's Gumi Plant 4 and other AI-industry sites, pledging budget, tax and financing support for advanced components, robotics and "physical AI," according to Newsis. That visit, however, centered on sensors and actuators rather than glass substrates specifically.
The data point to watch
The thesis that Korea keeps its lead — or loses it — will be tested by customer qualifications, not lab simulations. The near-term confirming signal is whether Samsung Electro-Mechanics hits its stated second-half-2027 mass-production target and whether SKC's Absolics converts prototype shipments into a named design win before TSMC's own pilot line scales. For LG Innotek specifically, the open question is when it commits to a firm commercialization date, which it has not yet publicly fixed.
This article is for informational purposes only and does not constitute investment advice. Figures are attributed to the cited sources as of publication; readers should verify current data before making any decision.
Sources:
