News

Startups

Artificial Intelligence

Americas

Aria Networks Raises $125M to Launch the World’s First AI-Native Deep Networking Platform

Aria Networks emerges from stealth with $125 million in funding to solve AI infrastructure bottlenecks through token-optimized networking.

Aria Networks emerges from stealth with $125 million in funding to solve AI infrastructure bottlenecks through token-optimized networking.

NewDecoded

Published Apr 8, 2026

Apr 8, 2026

5 min read

Palo Alto-based startup Aria Networks emerged from stealth today, announcing the general availability of its Deep Networking solution alongside a $125 million Series A funding round. Lead investor Sutter Hill Ventures is joined by Atreides Management, Valor Equity Partners, and Eclipse Ventures in supporting the mission to build the first network that thinks. The platform is specifically engineered to maximize token efficiency, which has become the defining metric for profitability in the AI era.

The company is addressing a critical bottleneck where expensive GPU clusters often sit idle due to suboptimal network performance. Aria argues that while the network comprises only a small fraction of total cluster costs, a mere one percent improvement in hardware utilization can recoup the entire networking investment. Their approach turns the network from a traditional constraint into a strategic multiplier for intelligence production.

Aria's technology stack is built on five core pillars including AI-optimized hardware and hardened SONiC software. The platform provides microsecond-level telemetry with up to 10,000 times higher resolution than legacy tools. This granular visibility allows for real-time insights across switches, transceivers, and hosts in a single unified view, ensuring that no data flow is left unmonitored.

At the heart of the offering are intelligent agents that perform autonomous root-cause analysis and dynamic load balancing. These agents leverage deep domain expertise to understand network states and can even be queried by operators using natural language. This partnership between human and machine ensures that clusters maintain peak performance without the manual, error-prone workflows typical of older systems.

The hardware launch includes 800GbE and 1.6T switches designed for the massive power densities of modern AI factories. These units are available in both air-cooled and liquid-cooled form factors to support the next generation of server infrastructure. Industry leaders like Broadcom and AMD have already validated the platform for use with their latest silicon and high-performance NICs.

Founded by industry veterans Mansour Karam and Subhachandra Chandra, Aria brings together top engineering talent from Google, Meta, and Arista. Gavin Baker, Managing Partner and CIO at Atreides Management, has joined the company board of directors. This leadership team combines decades of networking expertise with a focus on the unique demands of large-scale AI training and inference.

The new capital will be used to expand the company's global footprint from four to 15 data center locations by the end of the year. Aria intends to double its engineering headcount and aggressively scale its enterprise sales operations in international markets. This expansion is timed to meet the rapid industry transition to 1.6T networking speeds projected over the next two years.

With customer orders already being fulfilled, Aria Networks is positioned to redefine how AI infrastructure is optimized at scale. The company remains committed to an open, vendor-agnostic ecosystem that prevents proprietary lock-in. By focusing on the production of cheap, efficient intelligence, Aria aims to lower the cost of innovation for the entire AI industry.


Decoded Take

Decoded Take

Decoded Take

This launch signals a fundamental shift in data center economics where the token replaces the packet as the primary unit of value. While industry incumbents focus on general-purpose bandwidth, Aria Networks is betting that specialized, agent-driven orchestration is required to keep pace with explosive GPU demand. By prioritizing Model Flop Utilization (MFU), the company is moving networking from a utility cost to a revenue-generating asset for AI factories. This development reflects a broader industry trend toward open Ethernet standards as proprietary fabrics struggle to scale within the increasingly diverse hardware ecosystems of modern cloud providers.

Share this article

Related Articles