NVIDIA’s 2025 photonic switch revolution: powering the AI future
Article By : Martin Vallo
In 2025, NVIDIA will launch photonic switches with co-packaged optics (CPO), a bold move driven by technological necessity and surging market demand. This innovation, set to debut with the Quantum-X InfiniBand switches in late 2025 and Spectrum-X Ethernet switches in 2026, targets the explosive growth of AI-driven data centers.
This analysis is based on Yole Group’s collection of reports, including Co-Packaged Optics 2025, Silicon Photonics 2025 (2024 edition still available), Semiconductor Laser Industry 2024, Optical Computing 2024, Optical Transceivers for Datacom and Telecom 2024.
In this collection of photonics reports, including a focus on market & technology and technology & cost, Yole Group explores the adoption of photonic technology, especially co-packaged optics solutions, and its transformative impact on the industry. Yole Group’s expert, Martin Vallo, shares some insights into the current state of the technology and its implications for the ecosystem and supply chain.
Dive deep into the future of photonics for AI and stay tuned for more exciting developments in 2025 as Yole Group hits the road again!
Tackling power efficiency bottlenecks
AI workloads, often spanning millions of GPUs, are pushing traditional data center networking to its limits. Conventional pluggable optical transceivers, which convert electrical signals to optical ones for high-speed data transmission, account for up to 10% of a data center’s energy consumption. As GPU clusters scale, this power drain siphons energy away from critical compute tasks. NVIDIA’s CPO solution integrates silicon photonics directly into the switch ASIC, eliminating standalone transceivers. The result? A claimed 3.5x improvement in power efficiency. This frees up energy for computation, slashes operational costs, and addresses a pressing challenge for increasingly power-hungry AI factories.
Bandwidth for the AI era
The rapid evolution of AI models demands colossal network bandwidth to interconnect vast GPU arrays. Traditional copper interconnects falter under these conditions, suffering signal degradation over distance and limited bandwidth density. Even pluggable optics struggle to keep up. NVIDIA’s photonic switches, like the Spectrum-X and Quantum-X series, deliver up to 1.6 terabits per second per port and aggregate bandwidths of 400 terabits per second.
Martin Vallo, Senior Technology & Market Analyst, Photonics, at Yole Group, comments: “By co-packaging optics with switch silicon, NVIDIA boosts signal integrity and bandwidth density, enabling seamless scaling of AI clusters to handle next-generation workloads.”
Beyond copper’s limits
AI factories represent a new breed of data centers requiring a network overhaul. Copper-based electrical interconnects hit physical ceilings in speed, distance, and heat management as bandwidth needs soar. Silicon photonics, leveraging light instead of electricity, sidesteps these constraints.
Martin Vallo from Yole Group asserts: “With reach extending to 100 meters and reduced latency, NVIDIA’s CPO switches mark a strategic pivot to optical networking. This aligns with an industry-wide shift toward photonics as copper reaches its breaking point in high-performance computing.”
A mature ecosystem
NVIDIA isn’t alone in its leap into CPO—it’s backed by a powerhouse coalition. Partnerships with TSMC, Coherent, Corning, Foxconn, and Lumentum have forged a robust supply chain. TSMC’s advanced packaging expertise, like CoWoS and SoIC, pairs with NVIDIA’s innovations in switch ASICs and micro-ring modulators. This technological convergence has reached fruition just in time for mass production, aligning perfectly with NVIDIA’s roadmap and the GTC 2025 announcement.
Staying ahead in the AI race
The AI boom has sparked fierce competition among tech giants like NVIDIA, Broadcom, Marvell, and Intel to dominate infrastructure innovation. NVIDIA’s CPO push isn’t just about keeping pace—it’s about leading. By integrating photonic switches into its GPU-centric ecosystem, NVIDIA positions itself as a comprehensive solution provider for AI factories. Rivals still tethered to traditional networking might get left behind as NVIDIA pushes ahead with this cutting-edge tech.
To end this exciting week…
NVIDIA’s release of photonic switches with co-packaged optics in 2025 is a response to the urgent need for power-efficient, high-bandwidth, and resilient networking to support massive AI clusters. It’s a calculated move to solve current data center pain points, leverage maturing photonic technology, and maintain leadership in the AI infrastructure race. While the full impact depends on adoption rates and cost-effectiveness at scale, the timing reflects both a technological tipping point and a strategic vision for the future of AI computing.
Stay tuned with Yole Group’s investigations on www.yolegroup.com and LinkedIn!
About the author
Martin Vallo, PhD, is Senior Technology & Market Analyst, Photonics, at Yole Group.
Martin specializes in optical communication and semiconductor lasers within the Photonics & Sensing activities at Yole Group. With 12 years’ experience in semiconductor technology, Martin is currently involved in the development of technology & market products as well as the production of custom consulting projects.
Prior to his mission at Yole Group, Martin worked at CEA (Grenoble, France), focusing on the epitaxial growth of InGaN/GaN core-shell nanowire LEDs by MOCVD and their characterization for highly flexible photonic devices.
Martin graduated from The Academy of Sciences, Institute of Electrical Engineering (Slovakia), with an engineering degree in III-nitride semiconductors.
Source: www.yolegroup.com