Palo Alto-based Aria Networks has officially launched its groundbreaking platform, “Networks that Think”, marking the arrival of the world’s first AI-native network designed specifically to maximize token efficiency in AI workloads. Built from the ground up, the network leverages a novel approach called Deep Networking, which redefines how AI infrastructure is optimized, monitored, and scaled.
Token efficiency has emerged as a critical metric in the AI factory era. It measures Model Flop Utilization (MFU) and cost per token, providing a direct indication of whether an AI cluster delivers on its investment. When networks underperform, they hinder the entire stack, slowing training gradients, disrupting inference processes, and preventing effective scheduling across thousands of accelerators. Conversely, optimized networks multiply the performance of all other components. Aria Networks estimates that even though the network represents only 10 to 15 percent of the total cluster cost, a mere one percent improvement in MFU can recover the network’s entire expense.
“AI factories require solutions that produce tokens efficiently and cost-effectively,” an Aria spokesperson said. “Deep Networking transforms the network from a constraint into a competitive advantage, allowing operators to maximize their infrastructure investment and scale without losing performance.”
Unlike legacy networking solutions that rely on static configurations and treat telemetry as an afterthought, Aria Networks integrates five core principles into its architecture. The system operates on AI-optimized hardware with hardened SONiC, delivering 800GbE and 1.6T switching in both air-cooled and liquid-cooled form factors. It continuously collects high-resolution telemetry data across switches, transceivers, and hosts, offering operators a single, unified view of network performance. Intelligent agents at every layer analyze signals, extract insights, and take autonomous action, while embedded networking expertise ensures decisions are grounded in deep domain knowledge. The platform evolves seamlessly over time, incorporating updates that optimize performance for every new workload, creating a self-improving flywheel where efficiency grows with usage.
From deployment, operators experience a network that continuously fine-tunes itself, adapts in real time to workload variations, and aligns configurations with operator intent. The network operates collaboratively with human operators, providing telemetry insights, alerting on issues as they arise, and enabling natural-language interaction for problem-solving. Embedded Field Deployment Engineers, integrated directly within customer teams, manage architecture, lifecycle performance tuning, and AI infrastructure integration, ensuring the network delivers measurable value from day one.
The launch coincides with notable strategic developments at Aria Networks, the 2025 founded startup. Gavin Baker, Managing Partner and CIO at Atreides Management LP, has joined the company’s board of directors, reinforcing the firm’s technical and strategic vision. Alongside existing board members, including Mansour Karam, Subhachandra Chandra, and Stefan Dyckerhoff, Baker brings decades of expertise in networking and AI infrastructure. Investors including Atreides Management, Valor Equity Partners, Eclipse Ventures, and Sutter Hill Ventures have also backed the company.
As Ethernet continues to dominate AI back-end deployments due to its openness and multi-vendor scalability, liquid cooling adoption is expected to reach 76 percent of AI servers in 2026. With rack densities approaching one megawatt, demand for higher-capacity 1.6T switches is accelerating faster than the 800G transition, and Aria’s AI-native SONiC switches provide leading-edge performance with no vendor lock-in.
