A port speed is the maximum rate one physical Ethernet port moves data: 400G means 400 gigabits per second, 800G means 800 gigabits, and 1.6T means 1.6 terabits. The data center has now climbed that ladder in under a decade. We published in early 2018 that 200G, 400G and 800G would all ship within five years, and the projection held. 400G entered volume in the cloud upgrade cycle, 800G became the AI-era volume speed in 2024 and 2025, and early 1.6 Tbps port shipments occurred in 2025. Each step doubles the ceiling for the switches, servers, and interconnects built around the port, which is why every speed transition reshuffles the vendor landscape. We track these transitions quarterly, and our primer on the wider stack lives in What Is Data Center Networking?
Port speeds double because the electrical lane underneath them doubles. 400G emerged off 50 Gbps-per-lane SERDES technology announced in 2017, and the same doubling carried the market from the 28G-to-56G lane transition that brought 100G ports to 400G ports into today’s 224G PAM4 lanes, which is the math behind 3.2T-class parts. Switch capacity shows the same stair step: Ethernet switch silicon moved from 25.6 Tbps parts in 2022 to 51.2 Tbps parts in 2024 and 2025, with 102.4 Tbps platforms emerging in 2H 2026, and per-port speeds ramped from 400G to 800G with 1.6T on the horizon. The next electrical lane, 448 Gbps SERDES, is already in standards work at the OIF’s 448 Gbps workshop, the first SERDES speed developed after the AI inflection. History rhymes here: 100 Gbps revenue passed 10 Gbps revenue for the first time in 2Q18, and each later transition has compressed the window between speeds.
Data center switching. The 800G upgrade cycle is the heart of the AI factory build-out. We project the 800G market for Ethernet switching and NICs will exceed $50B across 2026 and 2027, with nearly 200M ports shipped in that window. AI scale-out Ethernet revenue is set to surge past $100B by 2030, and pluggable optics at 800G and 1.6T plus emerging co-packaged optics could exceed 50% of annual switch revenue by 2030. The pace is without precedent: the industry now ships more networking capacity in a single quarter than it shipped in the entire prior year. Merchant silicon points the same direction, with Tomahawk 6-class 102.4T designs aimed at clusters beyond 100,000 accelerators and co-packaged optics moving from differentiator toward requirement.
Servers, NICs and DPUs. The same ladder runs inside the server. PCIe moved from Gen 4 at 16 Gbps per lane toward Gen 6 at 64 Gbps per lane, which shrank the usable copper trace from about eight inches to under four and pulled retimers into every AI rack. Server-to-top-of-rack links climbed from 25G to 50G and now 100G lanes, and server speeds are rising faster over the next five years than they did over the prior ten, as multiple speed transitions drive demand for more DSPs. NICs stepped up in step: vendor slates at the 2026 OCP Global Summit put 400G (Thor 2) and 800G (Thor Ultra) NICs in the same rack designs as 102.4T switches. On copper, Marvell’s second-generation 1.6T AEC DSP pushes reach past six meters, and its 3.2T retimer carries 16 lanes of 224G PAM4 for XPU-to-XPU scale-up links. Scale-up domains are growing with the speeds, from NVIDIA’s 8-GPU and 72-GPU systems toward 576 and 1,152 GPU domains. Our SmartNIC and DPU Vendor Guide covers the device landscape in detail.
Routers, DCI and transport. Between buildings, the ladder runs through coherent optics rather than direct-connect modules. One Coherent Lite wavelength carries 800G up to 10 km without amplification, and we project the Coherent Lite market will surpass 10 million units within three to five years as data center capital spending doubles in 2026 and potentially again in 2027. The modules land in Ethernet switches and routers alike. On longer reaches, ZR is entering its third generation at 800 Gbps with ZR+ pushing to 1000 km, and the ZR market is set to exceed one million DSPs annually. The substitution is already visible in transport: the metro portion of the optical transport market declined in 1Q25 while long-haul grew, because coherent pluggables replace metro systems. All of this feeds a wider WAN build-out as AI networking approaches $100B a year by the end of the decade.
Optics and power. 800G optical transceivers led the AI networking surge: in June 2024 we published that data center AI networking would grow to nearly $20B in 2025, led by Ethernet, InfiniBand and 800G optical transceivers, with record levels of 800 Gbps ports and optics across the following 18 months. Power is the constraint that shapes the next steps. Optical connectivity already nears 10% of a total AI factory’s power, and co-packaged optics cuts optical network power by nearly 5x. That is why 1.6T arrives with new form factors: Credo’s 1.6T active light cable targets up to 50 meters at 12W per cable end, shipping in 2027, and row-scale optics sit alongside co-packaged designs in the 1.6T cycle.
What comes next. The 1.6T ramp builds through 2026 and 2027 alongside 102.4T switch silicon, and the industry is already staging 3.2T: 224G lanes multiplied by 16 is the 3.2T math, and the first parts sample today. We think the telling datapoint is cadence, not any single speed: the industry now ships more networking capacity in a quarter than it did in the entire prior year, and more than 200 ports a minute will barely keep pace with AI demand. Our quarterly report family tracks port shipments from 25G through 800G, 1.6T and 3.2 Tbps.
How was this calculated? The figures in this Explainer come from 650 Group’s published quarterly research on data center switching, AI networking, and interconnect markets, which track vendor revenue and port shipments by speed each quarter. We published the speed-transition history in the 200G/400G forecast (2018) and the market sizes in the reports linked throughout, and the full catalog is on our research reports page, including the published view that AI networking approaches and potentially exceeds $1T annually in the early 2030s.
650 Group researches several of the technologies we covered in this Explainer, including Data Center Ethernet Switching, SmartNICs, and more.