AMD-backed AI networking challenger DriveNets expanded its AI fabric portfolio, unveiling platforms designed to interconnect hundreds of thousands of specialized processing units (XPUs).The 2600SL and 2601S are both based on Broadcom's Tomahawk 6 chip and feature 64 ports of 1.6 Tbps (1.6T) capacity to provide a total connectivity capacity of 102.4 Tbps (102T).The air-cooled 2601S and liquid-cooled 2600SL units are capable of supporting deployments spanning scale-up (inside the rack), scale-out (cabinet-to-cabinet), and scale-across (distributed computing). The platforms are set to ship during the current third quarter.DriveNets touted its flexibility across varying network architectures, with the systems able to run across two- or three-level fat tree network designs – the former increasingly being favored by hyperscalers through protocols like multipath reliable connection (MRC) or resilient network graphs (RNG) as they look to reduce congestion and speed-intensive AI inference workloads.The software powering the platforms offers optimization capabilities, allowing users to fine-tune network performance end-to-end. At the hardware level, the startup contends the liquid-cooled configuration differs from rival platforms in that it boasts a “true 100 percent liquid-cooled design, ensuring consistent thermal efficiency across the entire system to maximize power efficiency.”"The most expensive idle asset in the world today is an XPU waiting on the network," DriveNets co-founder and CEO Ido Susan noted. "The new additions to our AI portfolio deliver industry-leading performance at massive cluster scale and let our customers maximize infrastructure utilization and power efficiency on any AI accelerator. These capabilities will only become more critical as the industry moves to heterogeneous AI architectures."DriveNets was founded in 2015 with a focus on software-defined networking, taking a cloud-native approach to routing and connectivity. Among its early wins, the firm provided the core routing system for AT&T.The vendor has since shifted toward AI infrastructure, using its networking expertise and distributed disaggregated chassis approach to improve connections between tens of thousands of chips.In early June, it raised $410 million via a Series D financing round that included AMD, Atreides Management, and Bessemer Venture Partners."As AI systems scale to hundreds of thousands of XPUs, the network fabric has become a critical factor in AI performance, infrastructure utilization, and overall token economics," Charlie Kawwas, president of Broadcom’s semiconductor solutions group, added on the news. "Broadcom's Tomahawk 6 silicon, paired with DriveNets' high-performance AI fabric, delivers unprecedented performance and reliability at massive cluster scale.”The launch sees DriveNets join an ever-growing host of vendors looking to capture hyperscale demand for 1.6T networking, with the generational leap over 800 Gbps (800G) seen as essential to meet bandwidth and latency requirements for clusters scaling beyond 100,000 nodes.Arista Networks was among the more recent names joining a growing cohort of suppliers, joining Edgecore and Aria Networks in offering 1.6T lines built on Broadcom's Tomahawk 6. Broadcom offers its own Tomahawk Ultra line, while Nvidia’s upcoming co-packaged optics (CPO) switches are also designed to support 1.6T.Sample shipments of several 1.6T lines from rival vendors have already commenced, with Dell’Oro Group projecting a ramp-up to kickstart in the second half of this year.