A growing number of networking startups are tackling the unique challenges posed by AI-driven infrastructure. Aria Networks aims to address these issues with a distinctive approach that leverages comprehensive network telemetry to deliver practical solutions for network operators.

Figure 1. Aria Networks Launches to Supercharge AI Networking, Says Apstra Founder.

While SDN and intent-driven networking were considered groundbreaking in their era, AI is now driving a fundamental transformation in network design and deployment. AI networking demands more than incremental improvements: backend Ethernet networks connecting GPU clusters have significantly different requirements compared with traditional cloud infrastructure. Figure 1 shows Aria Networks Launches to Supercharge AI Networking, Says Apstra Founder.

From Switch-Centric to Path-Centric Architecture

He highlighted that overall data center networking growth has hovered in the single digits over the past decade or two. AI networking, however, is dramatically altering that trajectory. “When you see growth this explosive, customer needs often feel underserved, and that opens the door for new entrants,” Karam said.

Aria’s technical strategy sets it apart from traditional vendors by focusing on end-to-end path optimization rather than individual switch performance. Karam notes that conventional networking companies primarily see themselves as switch providers, with software efforts concentrated on switch operating systems instead of cluster-wide operations.

“It’s no longer just about the switch itself—it’s about the complete end-to-end path,” Karam explained. “For job scheduling, what really matters are the paths traffic takes across the network from start to finish.”

Telemetry at Microsecond Resolution

Aria targets the backend Ethernet networks connecting GPUs in AI clusters, building on Broadcom merchant silicon and the open-source SONiC network operating system.

Deterministic vs. Probabilistic Network Optimization

Aria is not only creating networking hardware for AI systems but is also leveraging AI to enhance network performance.

Karam distinguishes between traditional rule-based, deterministic methods and AI-driven probabilistic approaches to network optimization. “When I built my previous company, Apstra, we didn’t have AI, so everything was handled deterministically—everything was rule-based,” he explained.

While deterministic approaches work well in controlled environments, they have inherent limitations. AI-driven probabilistic methods, on the other hand, offer advantages in detecting performance issues intuitively and reacting dynamically in complex scenarios. “If you want to be more intuitive in identifying performance problems or responding in real time, probabilistic AI methods provide a unique advantage,” Karam said.

However, he cautions that simply layering AI onto existing architectures is insufficient. He criticized vendors that rely on AI chatbots as a superficial enhancement to legacy systems.

“A lot of vendors claim they ‘bring AI,’ but often what they do is just add an AI chatbot on top of an old architecture,” Karam noted. “Across every domain, we’ve seen that AI only becomes truly effective when the system is purpose-built for it—meaning the architecture is designed from the ground up with AI in mind.”

Peeling Back the Layers

Aria is taking a staged approach to sharing details about its platform as it continues to develop. While the company has disclosed foundational elements such as SONiC and microsecond telemetry, it is gradually revealing more as the technology matures.

“The power of AI is enormous,” Karam said. “Combining AI with high-quality data to tackle problems and optimize performance presents an immense opportunity.”

Source: NETWORK WORLD

Cite this article:

Priyadharshini S (2025), Apstra Founder Unveils Aria to Boost AI Network Performance, AnaTechMaz, pp.249