On September 30, 2026, Taiwan's Commercial Times, citing Korean media reports, said NVIDIA is urging substrate makers in Taiwan, South Korea and Japan to accelerate glass substrate development, and may adopt the technology before 2028 at the earliest.
Joint development efforts involving TSMC and others have shown that glass could be used as a substrate material for semiconductor packages. However, we have yet to see any announcement that reliability evaluation has been completed for a mass-production product at the size needed for large AI chips from 2027 onward.
The 2028 timeline is not a guarantee that supply will be ready. It should be read as a target date that companies are aiming for in order to make the reported early adoption happen.
A verification package for a glass-core substrate, co-developed by TSMC and others, has an outline area equal to about 52% of the substrate needed for the 2027-generation CoWoS (TSMC's advanced packaging technology). But this compares outline dimensions across different generations and does not indicate how mature the technology is.
Even Samsung Electro-Mechanics and SKC, which are ahead in development, face open questions over mass-production timing and customer reliability evaluation.
To assess the reports of NVIDIA's adoption plans, it is necessary to look beyond technical progress and consider three separate challenges: scaling up to larger sizes, manufacturing yield, and customer qualification.
Where Did NVIDIA's Reported "Within Two Years" Development Goal Come From?
The Commercial Times article presents a series of information about NVIDIA as Korean media reports. It also said NVIDIA is considering adopting glass substrates and working with companies including German equipment maker SCHMID.
However, it is not clear who at NVIDIA said this or in what setting. The article itself says the adoption timing needs confirmation from the companies involved.
In South Korea, KIPOST reported on September 22 that NVIDIA, which had previously been cautious about glass substrates, is now making concrete plans, and pointed out that adoption could come sooner than the industry expected.
Seoul Economic Daily reported on September 24, citing remarks at SCHMID's earnings briefing, that NVIDIA is studying adoption with multiple companies.
The SCHMID remarks behind these reports go back to the first-half earnings briefing held on August 25.
Roland Rettenmeier, SCHMID's Chief Sales Officer (CSO), said in the Q&A that many companies in the Intel, NVIDIA and AMD supply chains are showing interest in glass-core substrates.
He said the company is also working with major supply chain companies to support the commercialization of glass-core substrates.
However, what Rettenmeier's remarks directly show is interest in glass substrates across the supply chains of chipmakers. They do not confirm that NVIDIA itself has decided to adopt them or has set a formal development deadline for substrate makers.
The same Q&A also covered the challenges of commercialization.
Rettenmeier said the biggest technical obstacle is the metallization of TGVs (through-glass vias). TGVs are structures that carry electrical signals through tiny holes drilled in the glass substrate, and a stable conductive layer must be formed on the inner walls and elsewhere.
He also said qualification by end customers is another item to watch.
Being technically able to make a substrate is a separate stage from a customer approving its use in a mass-production product.
Investment in related technologies is also moving forward in Taiwan.
At its July 30 earnings briefing, AUO said its NT$8.6 billion capital spending plan includes investment in pilot lines for TGV, RDL (redistribution layer) and glass core.
AUO Chairman Paul Peng said it will take another two to three years for glass-core substrate technology to mature.
Innolux and TPK-KY have also acknowledged that it will take time to reach mass production.
In other words, while there are reports that NVIDIA is aiming for early adoption, comments from equipment makers and substrate-related companies suggest that technical development and customer qualification still need time.
TSMC's Verification Glass Package Is About Half the Area Needed for the 2027 Generation
AI accelerators achieve high computing performance and memory bandwidth by combining GPUs and HBM (high-bandwidth memory) in a single package.
As generations advance, the number of chips and HBM stacks increases, and packages are growing larger.
At its 2026 North America Technology Symposium, TSMC said it is already mass-producing 5.5-reticle-size CoWoS, with yields above 98%.
The reticle is a measure of the maximum pattern area an exposure tool can form in one exposure; a standard 1-reticle field is 26 × 33 mm.
TSMC's roadmap plans to expand CoWoS size to 9.5 reticles in 2027, 14 reticles in 2028, and more than 14 reticles in 2029.
For the 9.5-reticle generation in 2027, the package substrate is expected to measure 120 × 150 mm.
Note that the yield of currently mass-produced 5.5-reticle CoWoS refers to existing packaging technology, not the mass-production yield of glass-core substrates.
A problem that arises as packages get larger is substrate warpage caused by heat.
Semiconductor packages are built by stacking multiple materials with different coefficients of thermal expansion (CTE), such as silicon chips, interposers and organic substrates.
When temperatures change, each material expands and contracts at a different rate. This creates stress between the materials and causes the whole package to warp.
As packages grow, the difference in expansion and contraction between materials also grows, making warpage control increasingly important.
One reason glass-core substrates are attracting attention is that they may suppress this kind of thermal deformation.
According to Tom's Hardware, the CTE of a glass core can be tuned to a range of about 3 to 10 ppm/°C, bringing it closer to silicon's roughly 2.6 ppm/°C. Compared with conventional cores made of organic materials, it could reduce warpage by about half.
Glass substrates are not just a new material. As AI chip packages grow larger, they are a candidate to solve manufacturing problems that conventional materials struggle to handle.
So how far has actual technical development come?
TECH+ reported on joint development results with Ibiden and Innolux, based on analysis by Taiwanese semiconductor industry sources of TSMC's presentation materials at JPCA Show 2026 on June 11.
According to the report, the joint development verified a configuration using a 0.8 mm-thick glass-core substrate carrying chips equivalent to 5 reticles. The package's outline dimensions were 85 × 110 mm.
The joint development and simulation-based verification reportedly showed a 16% improvement in COP (coplanarity), a measure of package warpage, and a 19% reduction in effective coefficient of thermal expansion.
No serious warpage or delamination was reported to have been observed.
However, these are reported joint development and analysis results. They do not mean that a mass-production product using a large glass-core substrate has passed customers' final reliability qualification.
Let's compare the package dimensions.
The verification package measures 85 × 110 mm, giving an outline area of 9,350 mm².
By contrast, the 9.5-reticle generation TSMC plans for 2027 is said to require a package substrate of 120 × 150 mm, or 18,000 mm².
The ratio of outline areas is as follows:
9,350 ÷ 18,000 ≒ 0.519
In other words, the outline area of this verification package is equivalent to about 52% of the substrate expected for the 2027 generation.
The ratio of 5 reticles to 9.5 reticles is also about 52.6%.
However, reticle count is based on exposure area and is not the same measure as the outline area of the whole package. It cannot be concluded that development is about halfway done merely because the numbers are close.
85 × 110 mm is the outline size of the verification package, while 120 × 150 mm is the substrate size needed for a future generation. What can be read from this comparison is that further verification may be needed to scale up, not how far yield or reliability targets have been met.
Even if glass substrates are adopted in 2028, they will not necessarily be used from the start in the largest 14-reticle-generation configurations. The required substrate size varies with the product and package specifications.
The verification adopted a structure in which the glass core is sandwiched between upper and lower layers of ABF (Ajinomoto Build-up Film).
This does not replace all the materials in a conventional package with glass; it changes the material in the central part to glass.
Ibiden, a Japanese semiconductor package substrate maker, also took part in this joint development.
Samsung Electro-Mechanics and SKC Still Face Uncertainty Over Mass-Production Plans
In glass substrate development, Samsung Electro-Mechanics and SKC's subsidiary Absolics are watched as frontrunners.
But a chronological look at the two companies' mass-production plans and related reports shows that, even as development continues, customer qualification and production facility preparation are taking time.
| Date | Company | Announcement / Report | Source |
|---|---|---|---|
| Feb. 5, 2026 | SKC | CFO said commercialization is somewhat later than market expectations | Greened |
| Early July 2026 | Samsung Electro-Mechanics | Set a target of starting operations in the second half of 2027 when establishing the GlaSSEM joint venture | The Elec |
| July 2026 | SKC | Reported that mass production would slip to 2027, with final reliability testing planned to finish in 2026 | Business Post |
| Aug. 12, 2026 | Samsung Electro-Mechanics | Reported that GlaSSEM's equipment orders were repeatedly postponed and mass production could come in 2028 or later | The Elec |
| Aug. 19, 2026 | Samsung Electro-Mechanics | Denied reports of failing customer qualification; said glass-core materials target the second half of 2027 and finished substrates target 2028 | Company explanation as presented by TrendForce |
| Aug. 25, 2026 | SCHMID | Described TGV metallization and end-customer qualification as key challenges for commercializing glass substrates | Earnings briefing |
| Sept. 22, 2026 | SKC | Announced additional investment in Absolics; continuing customer qualification and technical demonstration | SKC official announcement |
For Samsung Electro-Mechanics, reports and the company's own explanation need to be considered separately.
In August, The Elec reported that orders for GlaSSEM's manufacturing equipment were repeatedly postponed and that mass production could come in 2028 or later.
It also cited industry sources as saying the delay was caused by a failure to pass customers' reliability evaluation.
However, on August 19, Samsung Electro-Mechanics denied the reports of failed customer qualification and said evaluation is still under way.
The company's stated targets are the second half of 2027 for glass-core materials and 2028 for finished glass substrates.
It would therefore not be accurate to treat this as a simple postponement from 2027 to 2028. The difference in the products concerned, and the fact that the company denies the report, should be kept in mind.
For SKC, meanwhile, it was reported in July that the start of mass production would slip to 2027.
However, all SKC had officially acknowledged was that commercialization was somewhat later than market expectations.
As of August 27, Tom's Hardware also reported that Absolics was aiming for mass production at the end of 2026, so target dates differ between reports.
SKC then announced additional investment in Absolics on September 22.
According to the company, it is continuing customer reliability evaluation and technical demonstration to advance commercialization of its glass substrate business.
Absolics plans to commercialize both types, those with chips and other components embedded inside and those without. Because the stage of technical verification and customer qualification differs by product, it is difficult for a single mass-production start date to represent the state of all development.
The important point is that customer qualification is not a mere formality.
At its February earnings briefing, SKC's CFO explained that because the usage environment differs by end product, such as data centers and AI accelerators, reliability testing suited to each is required.
If unexpected verification items are added, the development period also extends.
Manufacturing challenges remain as well.
Glass can chip at the edges or develop microcracks due to shock and stress during processing.
According to MIT Technology Review, in Absolics' early mass-production trials, hundreds of glass panels sometimes broke every few days.
TGV metallization, which SCHMID pointed to as a technical challenge, is also a process that affects mass-production yield and product reliability.
Commercializing glass substrates requires not only improving material properties but also ensuring fine-processing precision, stable manufacturing yield, and reliability when built into the final product.
Even if NVIDIA's Demands Speed Up Development, Customer Qualification Takes Time
On the other hand, if a major customer like NVIDIA gives a specific adoption timeline, capital investment by substrate makers and equipment makers could accelerate.
Companies are already investing with mass production in mind.
Absolics has received up to $75 million in direct support from the U.S. Department of Commerce and produced samples intended for mass production in the first quarter of 2026.
Japan's Dai Nippon Printing (DNP) has also set up a pilot line for glass-core substrates at its Kuki plant in Saitama Prefecture and announced plans to begin sample shipments in early 2026.
DNP is aiming to build a system for starting mass production in fiscal 2028.
Because fiscal 2028 in Japan runs from April 2028 to March 2029, it cannot be treated as equivalent to starting mass production before 2028.
The joint development results from TSMC and others also show the potential of using glass as the core material of packages.
These moves show that glass substrate development has moved beyond basic research into a stage that anticipates actual manufacturing processes and mass production.
However, past delays in development plans have not necessarily been due to insufficient demand.
An SKC representative said in February that the development period had lengthened because reliability evaluation items that were not originally expected were added, and that the exact timing of mass-production shipments could not be determined.
For Samsung Electro-Mechanics' joint venture GlaSSEM as well, The Elec reported that equipment orders were postponed while customers continue evaluating prototypes.
There were reports on September 16 that moves toward placing equipment orders had resumed, but KIPOST said the first order would come in October at the earliest.
Even if a major customer asks for development to be hurried, the process of confirming product reliability cannot itself be skipped.
For TSMC as well, it has been reported that mass production requires continued research on glass thickness and large-package design.
A researcher at Korea's Kyobo Securities also pointed out that companies need to go beyond stating a target year for starting mass production and show how customer qualification progress and manufacturing yield are improving.
These challenges are especially important when adopting glass-core substrates in large AI chip packages at the CoWoS class.
It should not be assumed that everything, including small products and glass substrates of different structures, will reach mass production at the same time.
2027 Becomes a Key Verification Period Ahead of 2028 Adoption
What to watch going forward is how far major glass substrate makers can obtain customer qualification, and at what size, through 2027.
If reports that NVIDIA is considering adoption before 2028 are true, suppliers will need to meet customer requirements for materials, manufacturing processes and reliability before then.
For large AI packages in particular, whether suppliers can handle products at the 9.5-reticle level that TSMC plans for 2027 will be one important benchmark.
However, because NVIDIA has not disclosed which product it will first use glass substrates for, it cannot be asserted that every supply candidate must produce 9.5-reticle-class products from the start.
Equipment delivery lead times are also important in considering adoption timing.
The Elec reported that even if Samsung Electro-Mechanics can place equipment orders in 2026, building the production line would take until the second half of 2027 at the earliest.
To realize plans to supply mass-production products in 2028, companies must establish a qualification outlook and proceed with introducing the necessary equipment before then.
Whether any of SKC's Absolics, Samsung Electro-Mechanics, DNP or others announces concrete results on customer qualification or mass-production shipments of glass-core substrates for large AI packages by around the end of June 2027 could serve as one gauge of the likelihood of early adoption.
That said, end-June 2027 is an observational benchmark derived from reported equipment lead times and companies' mass-production targets, not a qualification deadline formally set by NVIDIA or substrate makers.
Before that, at its July 27 earnings briefing, SKC indicated that results of package-level reliability evaluation being conducted in Taiwan could come in 2026.
These evaluation results and the customer qualification progress that companies announce going forward will be important for considering glass substrate adoption around 2028.
Reports that NVIDIA is eager to adopt glass substrates further raise market expectations. But asking for a shorter development period is not the same as solving the technical challenges required for mass production.
Going forward, what will indicate progress toward practical use is not only the planned year of adoption but also how far companies can show the package sizes they can handle, the results of customer reliability evaluation, and stable manufacturing yield.