NVIDIA's CPO Technology Enhances AI Performance and Bandwidth Efficiency

NVIDIA has unveiled prototypes of its next-generation Co-Packaged Optics (CPO) and Optical Fiber technologies, which are poised to significantly enhance the performance of future AI systems. This development, showcased during SEMICON West 2026, marks a critical step in the ongoing evolution of optical technologies in computing, particularly as demand for advanced AI capabilities continues to surge. The shift from traditional architectures to more integrated solutions is essential for meeting the increasing requirements for bandwidth and power efficiency in data centers.

NVIDIA's research, led by Principal Research Scientist Nikola Nedovic, emphasizes the transition of optical engines from organic substrates to silicon interposers. This change is expected to yield substantial benefits, including higher bandwidth density and reduced power consumption. The new technology, referred to as Optics-on-Interposer (OOI), facilitates the movement of electrical signals within the interposer, utilizing shorter and finer interconnects. This architecture mirrors the design principles that have successfully brought GPUs and High Bandwidth Memory (HBM) closer together, thereby optimizing performance.

In addition to OOI, NVIDIA is advancing its dense wavelength-division multiplexing (DWDM) technology, which is currently in the research phase. This technology allows multiple data streams to be transmitted simultaneously over a single optical fiber, significantly increasing data throughput. NVIDIA's demonstrations indicate that this system can achieve speeds of up to 32 Gb/s across eight wavelengths, culminating in a total bandwidth of 256 Gb/s per fiber path. Such advancements are critical as they will enable the scalability needed for future AI applications.

NVIDIA's leadership in CPO technology is further underscored by its ongoing development of a comprehensive range of chips tailored for AI platforms, including the BlueField-5 and NVLink 8 CPO. The company is also preparing for the release of its Rosa Feynman solutions, anticipated in 2028, which are expected to surpass the capabilities of the Rubin Ultra in the CPO domain. This strategic positioning indicates NVIDIA's commitment to maintaining its competitive edge in the rapidly evolving AI landscape.

Market dynamics are also shifting in favor of CPO technologies, as evidenced by recent earnings reports from suppliers like Lumentum and Coherent. These reports dispel previous concerns regarding delays in ultra-high-power CPO laser demand, revealing a robust growth trajectory with scale-out revenues expected to commence in the second half of 2026 and scale-up volumes following in 2027. Customer deployments are projected for 2028, aligning with NVIDIA's timelines and investments.

The implications of these developments are profound. As NVIDIA accelerates its investments in capacity and research and development, the broader market for AI and optical technologies is likely to experience significant transformation. The integration of CPO technologies will not only enhance the performance of AI systems but also drive the next generation of AI factories, setting a new standard for efficiency and capability. Competitors will need to adapt quickly, as the advancements in optical integration signal a pivotal shift in how data centers will operate in the near future. This evolution will redefine the competitive landscape, compelling firms to innovate or risk obsolescence in a market increasingly driven by the demands of AI.