Aria Networks has introduced its Deep Networking platform, aiming to optimize AI supply chains, backed by a $125 million funding round. Founded in January 2025 by Mansour Karam, the company has shifted focus from traditional switch-centric models to a path-centric, telemetry-based approach. This innovation positions the network as an active participant in performance rather than just passive infrastructure.
The Deep Networking platform integrates specialized switching hardware, a modified version of SONiC, and real-time telemetry from various components across the network. This granular data collection at the microsecond level enables adaptive tuning of load balancing, congestion notifications, and failover strategies without relying on threshold breaches or manual intervention.
Karam emphasizes that effective AI requires purpose-built architecture, a claim that stems from his expertise at Apstra. The capability to collect telemetry directly at the ASIC level distinguishes Aria from traditional network monitoring systems, allowing for immediate reaction to network events.
At its core, Aria’s platform addresses two primary metrics: Model FLOPS Utilization (MFU) and token efficiency—essentially focusing on how effectively the hardware is being utilized. Karam explains that the network’s performance significantly impacts both metrics, demonstrating the need for an optimized infrastructure in AI environments.
Aria’s switch portfolio comprises various models built on Broadcom ASICs, all compliant with standardized operating systems. The company is also pioneering a different operational model by embedding “forward deployed engineers” within customer environments, ensuring that real-world data continuously informs product improvements and iteration.
This fresh approach aims at not only advancing network performance but also ensuring consistent uptime and integration with existing systems for better operational efficiency in AI-driven workloads.
