Glass Substrates Could Slash Power Consumption in Next-Generation AI Chips

Why Glass Is Different

Glass substrates offer several properties that address organic substrates' limitations. Glass expands and contracts much less with temperature changes than organic materials, reducing the thermal stress that causes solder joint failures over time. Glass can be produced with much higher dimensional precision—flatter, more uniform—than organic substrates, enabling tighter tolerances in chip mounting and reducing the number of defective packages.

Most importantly for AI applications, glass enables much higher density interconnection than organic alternatives. Smaller holes—called through-glass vias—can be created in glass than the through-silicon or through-organic vias used in current packaging, allowing more connections to be packed into less space. More connections mean faster data transfer between the processor and its memory, which is currently one of the primary bottlenecks limiting AI chip performance.

The Energy Efficiency Angle

AI data centers consume enormous amounts of electricity—a single large language model training run can consume as much power as thousands of homes over weeks. A significant fraction of that energy is consumed not in computation itself but in moving data between processors and memory. The fundamental limitation is that data transfer through electronic connections dissipates energy as heat proportional to the distance traveled and the number of connections traversed.

Tighter interconnection enabled by glass substrates reduces the distance data must travel and enables lower-voltage signaling, both of which reduce energy per bit transferred. If glass packaging can reduce interconnect power consumption by even 20-30 percent, the aggregate impact across millions of data center chips would be substantial—both in energy savings and in the reduced cooling infrastructure required to remove waste heat.

The Manufacturing Challenge

Glass substrates are not without manufacturing challenges. Glass is brittle and requires different processing techniques than organic materials. Creating the precise through-glass vias needed for high-density interconnection requires laser drilling and chemical etching processes that are more complex than analogous steps in organic substrate manufacturing. And building the supply chain—glass production, via formation, metallization, and assembly—from scratch is a multi-year industrial effort.

Absolics's entry into commercial production represents the first time glass substrates have been available at industrial scale, not just as research demonstrations. Intel's investments in the technology, which the company has publicly discussed as part of its roadmap to regain semiconductor leadership, provide both validation and a major potential customer for the emerging supply chain.

The Roadmap

Industry analysts expect glass substrates to appear first in the highest-performance AI accelerator chips, where the performance and efficiency benefits justify the premium manufacturing cost. As the supply chain matures and volumes increase, costs should decline, eventually making glass substrates competitive with organic alternatives in a broader range of applications.

If the technology proves out at scale, it could become a standard element of the chip packaging stack used in AI accelerators, high-performance CPUs, and eventually consumer electronics over the next five to ten years—another example of a niche industrial material becoming ubiquitous through the demands of high-performance computing.

This article is based on reporting by MIT Technology Review. Read the original article.