ⓘ Lightmatter
Optical connections could replace some of the copper inside AI servers. Lightmatter and Flex have just demonstrated a compact 204.8 Tbps switch that can take this closer to reality.Optical connections could replace some of the copper inside AI servers. Lightmatter and Flex have just demonstrated a compact 204.8 Tbps switch that can take this closer to reality.
Aurelian Nedelcu, Published
Lightmatter has unveiled a 204.8 Tbps optical switch developed with electronics manufacturer Flex, squeezing an enormous amount of potential networking bandwidth into a single rack unit. The system was presented at the 2026 Open Compute Project Global Summit, currently taking place in California, USA, alongside a separate optical computing prototype.
The technology tackles a growing problem in AI data centers: moving information between increasingly powerful processors. Conventional copper connections become harder to manage as systems expand, while optical links can carry data over greater distances with less signal loss.
Lightmatter's solution uses its Passage optical engines and Guide DR laser modules to transfer data using light. The switch combines eight bidirectional optical engines inside a liquid-cooled chassis measuring just one rack unit in height, approximately 4.45 cm.
According to Lightmatter, the design provides two to four times the front-panel connection density of conventional approaches using external laser modules. That could allow more bandwidth without requiring larger networking equipment. The company also demonstrated a separate computing tray containing four AI accelerators and two CPUs, connected through optical links offering a combined 102.4 Tbps of bandwidth. Both designs follow Open Compute Project hardware specifications, which could make integration into existing data centers easier.
However, the switch is currently a production-validation platform, while the computing tray remains a proof of concept. Neither demonstrates widespread commercial deployment, and the claimed density advantages have yet to be independently confirmed.
Lightmatter's approach shows how future AI servers might communicate faster without requiring substantially more rack space.