Broadcom Debuts 51.2-Tbps ‘Tomahawk Ultra’ Switch Chip to Take on Nvidia AI Networks
Updated on
Published on
Broadcom has begun shipping its Tomahawk Ultra networking chip, a 51.2-Tbps Ethernet switch ASIC built for hyperscale AI workloads. Announced Wednesday, the device doubles the bandwidth of last year’s Tomahawk 6 and supports 800G ports through 224-Gbps SerDes, giving cloud providers enough throughput to wire 256-GPU training pods without resorting to Nvidia’s pricier InfiniBand fabric.
The chip’s headline specs include sub-250-nanosecond end-to-end latency, in-network collective acceleration for AI operations and a lossless congestion-control scheme Broadcom claims matches InfiniBand’s performance while preserving Ethernet’s open standards. Partners like Quanta Cloud Technology and Arista said switches based on Tomahawk Ultra will sample this quarter, with volume production slated for early 2026.
Analysts see the launch as Broadcom’s boldest strike yet at Nvidia’s Spectrum-X and NVLink ecosystems. With AI data-center power budgets soaring, operators are eager for alternatives that integrate into existing Ethernet gear. Broadcom shares rose more than 1 % on the news, while Nvidia slipped in early trading as investors weighed fresh competition for its networking dominance.