On October 13, 2026, at SEMICON West 2026 in San Francisco, NVIDIA presented a technology that places optical engines on the same interposer as compute chips, along with a technology that parallelizes communication using light of different wavelengths. According to THE ELEC, a Korean semiconductor trade outlet that covered the event on site, NVIDIA calls the former "OOI (Optics on Interposer)" and has already built a prototype. OOI shortens the electrical wiring needed before a signal is converted to light, while DWDM (dense wavelength division multiplexing) increases the amount of data a single optical fiber can carry. Both are research efforts aimed at expanding communication bandwidth in AI systems, but they improve different parts of the link. The 256Gb/s transmission speed reported here is also a research-stage result, and it has not been decided that it will be adopted as-is in a future product.
OOI places the optical engine on the interposer to shorten electrical wiring
The defining feature of OOI is that it moves the optical engine, previously placed on the organic package substrate, onto the interposer that carries the compute chip. An interposer is an intermediate component that connects multiple semiconductor chips with fine wiring. According to THE ELEC's on-site report, electrical signals previously traveled from the compute chip through the interposer and the organic substrate to the optical engine. With OOI, the connection is made directly through short, high-density wiring on the interposer. This removes the section of the electrical path that passes through the organic substrate, with the aim of raising connection density and cutting power consumption. However, this does not mean the organic substrate disappears from the package as a whole; what changes is the route the electrical signals take to the optical engine.
An optical engine is a group of circuits that convert between electrical and optical signals. Conventional CPO (Co-Packaged Optics) places a semiconductor chip such as a switch ASIC and an optical engine in the same package, shortening the electrical wiring needed before conversion to light. In a technical explainer published in August 2025, NVIDIA described soldering the optical engines of Spectrum-X Ethernet Photonics directly onto the module substrate and using detachable connectors for the fiber connections. OOI amounts to a design that brings the optical engine of such CPO even closer to the compute chip.
The benefit of moving the optical engine closer is not limited to shorter electrical wiring. Fine wiring on the interposer allows many electrical signals to be transmitted in parallel at high density. In the study "Beyond CPO", published in the Journal of Lightwave Technology by NVIDIA's Ben Lee, Nikola Nedovic and colleagues, the authors explain that the interposer's high-density wiring can improve the efficiency of an approach that uses many parallel connections while keeping each signal's speed low. Rather than pushing a small number of wires to ever-higher speeds, the idea is to spread relatively slow signals across many wires.
Even if the fiber side's transmission capacity is raised, the connection becomes a bottleneck if the compute chip cannot send enough data to the optical engine. OOI aims to resolve exactly this constraint on the electrical side. It resembles configurations that place a GPU and HBM on the same interposer, but the optical engine does not replace HBM's function. It is a technology for increasing the density of the connection between the compute chip and the optical communication circuits.
256Gb/s over eight wavelengths: DWDM parallelizes optical communication
While OOI improves the connection between the compute chip and the optical engine, DWDM raises the transmission capacity of the optical fiber itself. DWDM (Dense Wavelength Division Multiplexing) places independent data on light of different wavelengths and sends it simultaneously over a single optical fiber. Because the receiving side separates and reads the signals by wavelength, multiple data paths can be consolidated into one fiber. According to THE ELEC, NVIDIA transmitted data at 32Gb/s on each of eight wavelengths, achieving a total bandwidth of 256Gb/s.
Helpful for understanding this configuration is research presented by NVIDIA's Nandish Mehta and colleagues at the Optical Fiber Communication Conference (OFC) in March 2026. The conference paper "A 256 Gb/s DWDM Optical I/O in a 3D-stacked EIC/PIC Silicon Photonics Platform" reports an "8+1" configuration consisting of eight data lanes and one lane dedicated to clock transmission. Each data lane runs at 32Gb/s, so the total data bandwidth is 8 × 32 = 256Gb/s. The dedicated clock lane conveys timing to the receiving side and is not counted in the data bandwidth.
The OFC research used a silicon photonics platform in which an electronic integrated circuit (EIC) built on a 7nm process and a photonic integrated circuit (PIC) built on a 65nm process were stacked in 3D. However, it cannot be confirmed whether this experiment and the OOI prototype shown at SEMICON West use the same circuit configuration. Nor do the experimental results in the OFC abstract automatically become the specifications of a future commercial product. It is also worth noting that related demonstration results for the DWDM technology had already been reported as of March 2026.
Care is also needed in what the speed figures refer to. For example, the "200Gbps per lane" cited for existing CPO technology and the "256Gb/s per optical fiber" from bundling eight wavelengths with DWDM measure different things, so a performance improvement cannot be calculated from a simple comparison. The official SEMICON West session overview positions two directions as complementary: using DWDM to raise bandwidth density while simplifying the electrical circuitry of each lane, and easing the connection constraints with the compute chip through high-density integration of optical engines. In other words, a design that combines the two technologies is under consideration: OOI to raise the density of the electrical-side connection, and DWDM to extend the fiber's transmission capacity.
From a 2023 concept to 2026 demonstrations: how the research has progressed
The idea of placing optical engines on an interposer is not new to this event. In a 2023 paper, NVIDIA had already proposed optical integration on the interposer and examined its advantages and implementation challenges. It also presented related experimental results on 256Gb/s DWDM at OFC in March 2026. The SEMICON West talk can be seen as positioning this research as a way to further extend AI-system communication capability and pointing toward combining the two.
Arranging NVIDIA's research page, the OFC abstract and the SEMICON West session overview chronologically shows what stage each piece of research is at.
| Release date / venue | Main content | Stage of research / development |
|---|---|---|
| February 15, 2023: study in Journal of Lightwave Technology | 2.5D optical integration on the interposer and low-speed, parallel electrical connections | Verification of a thermal control loop and simulations comparing designs |
| March 15–19, 2026: OFC, paper M4B.2 | 256Gb/s DWDM using eight data lanes and a dedicated clock lane | Experimental results published using 3D-stacked electronic and photonic integrated circuits |
| October 13, 2026: SEMICON West talk | Future AI connectivity combining DWDM and optical integration on the interposer | Presented implementation challenges and the direction of technology development for next-generation systems |
This table organizes research results and talk content published at different times; it does not show the progress of continuously evaluating the same prototype. The 2023 date is the publication date listed on NVIDIA's research page, not the time the research itself began. Nor is there information indicating that the OFC DWDM experiment and the OOI prototype reported in October were part of a single series of validations using the same equipment.
What stands out this time is that technologies proposed and demonstrated in the past were presented as a concrete design direction for further raising AI-system connectivity. While reporting that an OOI prototype has been developed, THE ELEC also says that both OOI and DWDM remain at the research stage and that technical challenges are being evaluated with a view to future products. The announcement should therefore not be taken to mean adoption in a specific GPU or the start of mass production.
Thermal management, packaging and serviceability will decide practicality
In a configuration like OOI that brings the optical engine closer to the compute chip, the electrical wiring can be shortened, but heat management becomes important. Heat generated by compute chips such as GPUs may affect the operation of nearby optical circuits. The official SEMICON West session overview also lists thermal control and packaging as major challenges, in addition to design trade-offs between devices and circuits. NVIDIA's 2023 research also verified a temperature control loop to cope with harsh thermal environments, showing that dense integration of optical engines requires thermal design to be considered together with electrical wiring.
NVIDIA's existing CPO products show implementation approaches to these issues. According to the technical explainer the company published in 2025, Quantum-X Photonics and Spectrum-X Ethernet Photonics separate the laser sources from the switch body and place them in a dedicated temperature-controlled environment. This design suppresses wavelength drift from temperature changes and early laser degradation. Because the external laser source modules can be replaced in the field, a failure can be handled without replacing the switch itself. Bringing the optical engine closer to the compute chip is not necessarily the same as integrating the laser source in the same place.
How optical fibers are connected also affects ease of mass production and maintenance. NVIDIA explains that Spectrum-X Ethernet Photonics uses detachable optical connectors, which help improve assembly yield and automate manufacturing. This is an implementation example from an existing CPO product, and it has not been decided that a future OOI will use the same structure. Still, the more densely optical engines are integrated, the more important it becomes to address precise fiber connections, inspection during manufacturing, and part replacement after failures. Shortening electrical wiring alone does not produce a product that can be mass-produced and easily serviced.
At present, it remains unclear how much power consumption OOI can actually cut, which GPUs or switches will adopt it, or when mass production will begin. In future evaluations, what will matter is whether stable communication is possible in an environment where the compute chip generates heat, and how far the power consumption of the entire connection, including electrical and optical circuits, can be reduced. Whether the benefits of dense integration can be maintained while accounting for manufacturing yield, cooling and replacement after failures will also be tested. If these challenges can be solved, the combination of OOI and DWDM could become a strong option for further expanding the communication bandwidth of AI systems.