News

It’s not TSMC, it’s not ASML! Corning relies on a piece of glass to block Nvidia's fate

2 min read
Kuai Technology reported on July 13 that when people discuss the bottleneck of Nvidia’s AI computing power, the first things that come to mind are always TSMC’s advanced processes and ASML’s lithography machines. But what really stuck in Nvidia's life was Corning, an old American company that makes glass. A recent closed-door exchange between Citigroup and Nvidia revealed a key signal: CPO co-packaged optics has officially crossed the mass production threshold. This technology packages optical modules and switching chips together, using light instead of electricity to transmit data. Corning's core weapon is glass bridge technology, which is a three-dimensional optical waveguide structure engraved on a glass substrate. It reduces the optical signal insertion loss of each channel from 1.5dB to less than 0.5dB, and the transmission quality and distance are greatly improved. This packaging process is extremely difficult. Previously, about 60% of the total packaging cost came from the key process of active alignment. After Corning successfully suppressed costs, its production capacity is said to be able to meet about 70% of Nvidia's demand in 2026. Corning has achieved self-sufficiency in its glass substrates, widening the cost gap to more than 30% in one fell swoop. Industry insiders bluntly said that if you want to do a CPO, you can't avoid Corning. The barriers it has built on the material side are something that no one can really break through in the short term. CPO was first installed on NVIDIA's Spectrum series switches, serving NVL72 complete cabinets under architectures such as Blackwell Ultra and Rubin. TSMC's CoWoS advanced packaging also participated in the collaborative promotion. Domestic optical module manufacturers such as InnoLight are accelerating to follow up, but the core glass materials are still subject to others. CPO mass production means that AI interconnection competition is quietly shifting from chip manufacturing to materials and packaging processes.