IEEE Nears Ratification of 1.6 Terabit Ethernet Standard Built for AI Data Centers

The networking industry is close to locking in a specification for 1.6 terabit Ethernet, a speed tier shaped largely by the bandwidth appetite of AI workloads. According to dev.to, the IEEE's draft is now in formal review and balloting, with final ratification anticipated later this year, and hardware makers are already demonstrating that 1.6T switches, modules and cabling can operate together outside the lab.

A standard that predates the AI boom

Work on the speed tier began in 2020, before the current wave of AI attention. Dev.to reports that the IEEE formed a study group in 2021 to examine higher speeds, which later became the IEEE P802.3dj Task Force charged with writing the technical specifications. Those drafts are now moving through the IEEE's balloting process. In parallel, industry groups are staging demonstrations meant to show that equipment from different manufacturers interconnects cleanly in realistic conditions, so that enterprise buyers have proof before committing to the next generation of hardware.

Capacity is not the same as 1.6T Ethernet

Some hardware already delivers 1.6 terabits of aggregate capacity by bundling slower links, such as four 400 gigabit ports. Genuine 1.6T Ethernet is one link running at full rate end to end. That distinction matters to engineers designing the fabrics used to train large machine learning models. According to dev.to, manufacturers now trust the stabilizing drafts enough to build early products, focused on short-reach connections inside the data center of up to 2 kilometers over single-mode fiber.

Coherent optics extend the reach

Direct-detect optics struggle to preserve signal integrity over longer distances, so the Optical Internetworking Forum has introduced the 1600ZR implementation agreement, a coherent interface that carries a full 1.6T signal on a single wavelength across 80 to 120 kilometers. Coherent modules are more complex and draw more power than short-reach parts, but they plug into the same routers and switches, letting operators mix short- and long-haul links in one chassis. Dev.to characterizes the arrangement, with the IEEE handling Ethernet itself while the OIF handles optical interconnects, as a deliberate division of labor meant to keep the market from splintering.

Interoperability is the hard part

At these speeds the tolerance for error is minuscule, and defects that hardly mattered at slower rates can wipe out a link entirely. The Ethernet Alliance coordinates testing between competing vendors, covering chip-to-module and module-to-module paths. The new specifications also add intricate forward error correction layers, and validating them across components from different suppliers, such as a switch from one company talking to a storage array from another over a third party's cable, adds up to a substantial workload.

What comes after 1.6T

The roadmap does not end here. The OIF is researching 1600ZR+, which would stretch coherent signals to 1,000 kilometers for links between distant cities, alongside a coherent lite effort aimed at 10-kilometer regional connections. On the Ethernet side, talks about 3.2 terabit speeds are slated to begin in early 2027. Dev.to points to 800 gigabit as the template, gear that began as specialized kit and ended up a data center default, and expects 1.6T to follow the same curve, with 2027 projected as the peak year for volume production and infrastructure upgrades, arriving in step with AI accelerators whose throughput depends on links of this size.

Why it matters

AI clusters are constrained by the network as much as by compute, and training traffic is exactly the kind of sustained, high-volume flow that punishes aggregated links and fragmented ecosystems. A ratified 1.6T standard with demonstrated multi-vendor interoperability gives operators a stable target for their next build-out rather than a bet on proprietary hardware. The timing is the real story: specification, optics and silicon are converging at roughly the same moment as the AI chips that cannot be fully exploited without them, and the planned start of 3.2T discussions in 2027 signals that the bandwidth race is structural rather than temporary.