Marvell Teralynx T100 debuts 102.4 Tbps AI data center switch

Marvell’s Teralynx T100 is an infrastructure story with direct implications for how far and how fast AI clusters can scale. At 102.4 Tbps and sub‑1000W typical power, it targets one of the hard constraints on frontier model training: network bandwidth and power budgets inside GPU- and accelerator-heavy racks. As racks push toward 120 kW, shaving 25% off switch power while increasing radix to 512 ports lets operators pack more compute into the same power envelope.

For the AGI race, this kind of plumbing matters because once model architectures mature, system-level bottlenecks become the limiting factor. Faster, lower-latency fabrics improve GPU utilization, reduce tail latency and enable larger, more tightly coupled clusters, which in turn support bigger models and more ambitious experiments. Marvell is staking out a leadership position in AI Ethernet fabrics, competing against Nvidia’s own networking stack and other silicon vendors. If Teralynx T100 gains traction with hyperscalers, it will diversify the networking layer of AI supercomputers and potentially lower the cost per unit of effective compute, nudging the frontier-model capability curve forward.