Lumentum, Qualcomm Technologies and Corning are demonstrating a high-density optical die-to-die interconnect architecture designed for next-generation AI scale-up systems. The ECOC 2026 showcase combines Qualcomm’s D2D interface technology with Lumentum’s 1060nm VCSEL optical engine and Corning’s multimode fiber connectivity, targeting bandwidth-dense links that extend beyond the practical reach of electrical interconnects.
AI accelerators are becoming increasingly powerful, but connecting those compute resources is creating a parallel engineering challenge. As AI systems scale across packages, boards and racks, electrical interconnects face increasing limitations around reach, power consumption, signal integrity and physical bandwidth density.
Lumentum, Qualcomm Technologies and Corning are demonstrating one potential solution at ECOC 2026: extending die-to-die (D2D) connectivity over optical links.
The board-level proof of concept combines Qualcomm’s D2D interface IP subsystem with Lumentum’s VCSEL-based optical interconnect platform and Corning’s multimode fiber connectivity. The demonstration is being shown at Lumentum’s stand 2050 and Qualcomm Technologies’ stand 1050/MB28 in Málaga, Spain.
The architecture is designed for future co-packaged optics (CPO) and near-packaged optics (NPO), while supporting emerging D2D standards and interfaces such as UCIe. Instead of keeping high-speed electrical connections between compute components, the approach converts the connection to optical transmission across distances measured in tens of meters.
That distinction matters for AI scale-up architectures. Scale-up networks connect processors, accelerators and memory resources within increasingly large compute systems, making bandwidth density and latency critical to overall system performance.
The demonstration uses optical channels operating at 32Gbps NRZ. Lumentum and Qualcomm say the massively parallel architecture is designed to provide approximately 1Tbps/mm of shoreline bandwidth density today, with a path toward approximately 4Tbps/mm through higher lane rates and greater channel density.
These figures are company-stated targets associated with the demonstration rather than measurements of a production AI system.
At the optical layer, Lumentum is using a 1060nm VCSEL platform with two-dimensional VCSEL and photodetector arrays and backside lens technology. The company says the optical engine provides approximately 10Tbps of aggregate transmit and receive capacity.
Lumentum’s use of VCSEL technology is also connected to its existing manufacturing infrastructure. The company says the platform builds on its high-volume 3D-sensing manufacturing base, potentially providing a manufacturing foundation for scaling optical interconnect components.
The company’s earlier 2026 demonstration at OFC used a high-density 1060nm VCSEL array co-packaged with a host ASIC and targeted scale-up protocols including UCIe and PCIe. (lumentum.com)
Qualcomm’s contribution focuses on the interface between compute silicon and the optical system. The company says its D2D technology is designed to optimize power and silicon area while supporting direct-drive optical connectivity.
That could become increasingly relevant as AI architectures move beyond conventional processor-to-processor connections. UCIe, for example, is designed to provide standardized chiplet connectivity inside advanced packages, while optical technologies can potentially extend high-bandwidth connections beyond those packages.
Corning supplies another critical component: the physical fiber connection. Its multimode fiber glass ferrules provide the dense optical interface used in the demonstration, highlighting an often-overlooked constraint in AI infrastructure—fiber density.
As optics move closer to GPUs, CPUs, switches and accelerators, the fiber itself has to fit into increasingly constrained packaging and board environments. Connector size, alignment, signal integrity and manufacturability therefore become part of the system-level design problem.
The broader market provides a strong incentive to solve those problems. IDC says global AI infrastructure spending totaled $318 billion in 2025 and forecasts it will reach $487 billion in 2026, with spending expected to surpass $1 trillion by 2029. (idc.com)
Networking is already absorbing part of that investment. IDC reports that data-center Ethernet revenue grew 64.5% year over year to $12.3 billion in the second quarter of 2026, while 800GbE represented 41.2% of data-center Ethernet revenue. (idc.com)
The Lumentum-Qualcomm-Corning demonstration addresses the layer below those switches: the physical interconnects that ultimately connect compute resources to the networking fabric.
The industry is consequently exploring several approaches at once, including higher-speed electrical SerDes, silicon photonics, VCSELs, CPO, NPO and optical D2D links. Companies such as NVIDIA, Broadcom and Marvell are also developing technologies aimed at scaling AI fabrics and reducing the power cost of moving data.
The latest demonstration does not represent a production deployment. Instead, it shows how optical D2D architectures could evolve from board-level demonstrations toward future AI systems in which optics are integrated much closer to compute.
For AI infrastructure designers, that shift could be significant. If processor and accelerator bandwidth continues growing faster than practical electrical interconnects can support, optical D2D may become another layer in the architecture used to keep compute resources connected without allowing interconnect power and distance to become the next scaling bottleneck.
Market Landscape
AI infrastructure is driving demand for higher-bandwidth networking at every level, from chip-to-chip connectivity to data-center fabrics. IDC’s Q2 2026 Ethernet data shows 800GbE already accounting for 41.2% of data-center Ethernet revenue, while overall data-center switching revenue grew 64.5% year over year.
At the infrastructure level, IDC forecasts AI infrastructure spending will reach $487 billion in 2026 and exceed $1 trillion by 2029.
This creates a technology stack in which compute, memory, SerDes, optical engines, fiber connectivity, switches and packaging increasingly have to be designed together.
Top Insights
- Lumentum, Qualcomm and Corning are demonstrating optical D2D connectivity aimed at future AI scale-up architectures.
- The architecture combines Qualcomm D2D technology, Lumentum 1060nm VCSEL optics and Corning multimode fiber connectivity.
- The demonstration targets approximately 1Tbps/mm shoreline bandwidth density today, with a company-stated path toward 4Tbps/mm.
- Optical D2D could extend high-bandwidth connections beyond electrical reach limits while targeting lower latency and power consumption.
- The collaboration illustrates how photonics and advanced packaging are becoming increasingly important to AI compute scaling.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
