HSINCHU, Taiwan — The glass substrate was supposed to be the ideal foundation for the next generation of chip packaging, and TSMC was supposed to be the company that made it work. Instead, the world’s largest contract chipmaker appears to be choosing the practical path first. TSMC’s first-generation co-packaged optics, or CoPoS, packaging scheme may skip glass substrates entirely and use an improved organic substrate for its initial production, according to CommonWealth Magazine, citing people familiar with TSMC’s plans.
CoPoS is TSMC’s flagship packaging technology for the era after CoWoS, the chip-on-wafer-on-substrate method that has become indispensable to AI accelerators. The new approach integrates optical components directly with computing chips, replacing the electrical wires that carry data between chips with beams of light. As AI models grow larger, the connections between chips have become a bottleneck: electrical signals lose integrity over distance and generate heat, while optical links can move far more data with far less power.
The choice of substrate is where the technology gets difficult. Glass offers the best signal integrity, allowing optical signals to travel with minimal loss, and its flatness and stability make it theoretically ideal for the precision alignment that co-packaged optics demand. But glass substrates remain immature as a manufacturing material: the supply chain is thin, yields are low, and the equipment to process glass at scale barely exists. Organic substrates, by contrast, are made by a mature industry that has supplied chip packaging for decades, and improved versions can approximate much of glass’s performance.
TSMC’s reported decision is a trade between the theoretical and the practical. By shipping its first CoPoS generation on organic substrates, the company can bring the technology to market faster, prove the optical architecture in real products, and buy time for the glass supply chain to mature. Analysts said the choice reflects a broader shift in TSMC’s thinking: with AI customers demanding more capacity every quarter, speed to market has become more valuable than the last percentage point of technical performance.
The stakes are commercial as well as technical. CoPoS is expected to appear in AI accelerator cards and high-speed networking chips beginning around 2027, and the production cadence of those products depends on when TSMC can deliver the packaging in volume. Customers including Nvidia and the networking-chip makers that serve data centers are watching closely, because the packaging choice determines how fast their next-generation products can ship.
The packaging industry has been watching glass for years. Intel has invested in glass-substrate research with partners, Samsung has outlined its own plans, and Amkor has announced a dedicated glass-substrate factory in the United States. The technology is widely expected to arrive eventually; the question has always been when. TSMC’s reported decision suggests the answer is: later, after the organic path proves the market.
The organic route has its own risks. The improved organic substrates TSMC would use are more expensive than conventional ones, and their performance, while improved, still falls short of glass in the signal-integrity measurements that matter most at the highest speeds. If the first CoPoS generation cannot deliver the bandwidth that the most demanding AI systems need, TSMC may face pressure to accelerate the glass transition sooner than planned.
The case for co-packaged optics rests on a simple limit: electrical connections cannot keep up. As AI accelerators grow, the links between chips and the networks around them are becoming the constraint on system performance, and optical connections can carry data across those links with a fraction of the power and heat. The technology is expected to appear first in the networking switches and accelerator cards that move data inside the largest AI clusters, where bandwidth demand is growing fastest.
Glass has been the industry’s favored candidate for the substrate underneath these optical systems, and the investment has been building for years. Intel has worked with partners on glass-substrate research, Samsung has outlined its own plans, and Amkor has announced a dedicated glass-substrate factory in the United States — a sign that the industry expects the technology to arrive, but also that it is not ready at scale today. The supply chain for glass substrates, from the material makers to the equipment that processes it, remains a work in progress.
TSMC’s reported choice would be a pragmatic pause on that timeline. By qualifying organic substrates first, the company can ship its first CoPoS generation on equipment and materials that exist today, and it can push the glass transition to a later generation when the economics improve. The risk is competitive: if a rival — Intel’s foundry business or Samsung’s packaging arm — brings glass-based co-packaged optics to market first with better performance, TSMC’s early organic products could face a performance gap in the most demanding systems.
For TSMC, the decision fits the pattern of its recent packaging strategy. The company has expanded CoWoS capacity aggressively over the past two years, building new packaging plants in Taiwan to meet demand that has repeatedly exceeded its projections, and it has shown a consistent willingness to choose volume over perfection. CoPoS on organic substrates extends that logic into the optical era. The first generation will be a test of whether the market accepts the compromise — and of how much time it buys for glass to catch up.
