Aria Networks closed a $125 million funding round, as the AI-native networking firm launched a solution designed for token efficiency.

Aria’s Series A round was backed by Sutter Hill Ventures, Atreides Management, Valor Equity Partners, and Eclipse Ventures. The firm was founded last year by Mansour Karam, who’d previously founded intent-based networking vendor Apstra that was acquired by Juniper Networks in 2020.

Aria’s raison d'etre is summed up by Deep Networking, the platform it launched to coincide with the big-bucks funding. Deep Networking is based on a suite of Ethernet-based switches using the software for open networking in the cloud (SONiC) set of standards that has underpinned architecture from the likes of Hewlett Packard Enterprise (HPE), specifically its AMD Helios-based AI rack-scale system.

The Aria solution runs on Broadcom’s Tomahawk 6 chipset and touts 800-gigabit Ethernet (GbE) and 1.6 Tb/s (1.6T) switching in liquid-cooled and air-cooled form factors. Deep Networking also includes end-to-end telemetry, and “intelligent agents at every layer,” spanning the switching ASIC through to cloud orchestration.

Token titan?

Aria noted Deep Network maximizes token efficiency, which it defines as “the defining metric of the AI factory era and the single best proxy for whether an AI cluster is delivering on its investment.”

“While networking represents just 10–15% of an AI cluster’s total cost, any inefficiency in the network has an outsized impact on accelerator utilization, training time, and ultimately, revenue,” Karam wrote in a product launch blog. “At the center of this shift is model flop utilization (MFU), [which] directly determines token efficiency and cost per token. Improvements in MFU translate directly into improvements in revenue. The network is the highest-leverage component in that equation: when it underperforms, it drags down everything else; when it’s optimized, it lifts the entire stack.

“This is the problem we set out to solve. Deep Networking is our answer,” Karam continued.

The firm reported its release depends on replacing manual, error-prone workflows with seamless, automatic fine-tuning across routing, load balancing, and failover. Through intent-based configuration, operators can define their objectives while the system executes the technical complexity, utilizing real-time, adaptive optimization to potentially resolve bottlenecks before they impact performance.

Aria called this an “agentic partnership” as opposed to a black-box solution, providing users with deep telemetry and natural language interaction to costrategize with AI agents. In addition, the solution integrates embedded field deployment engineers (FDEs) within the customer’s team, keeping humans in the loop amongst said AI agents.

Karam’s launch blog saw him thank Aria’s investors, as well as partners Broadcom and Amphenol, plus the broader ecosystem for its work in open, standards-based networking. Presumably this refers to the SONiC Foundation, whose members include Nokia, Arista Networks, Cisco, Dell Technologies, Google, Intel, Marvell, Nvidia, and Microsoft. While not a member, Amphenol supports SONiC-based deployments.

“We’re building for the next decade of AI scale, and we’re doing it with the urgency and conviction that this moment demands," Karam wrote. "In a market where high-end networking silicon is in extremely high demand and allocations are tightening, early engagement matters. The operators who move first will have a structural advantage in data center economics that compounds over time."