Keynote by Alex Saroyan, Co-Founder & CEO of Netris | RAISE Summit 2026

Market Context & Incident History

• Two years ago, a major Neocloud provider experienced a network outage during resource reconfiguration due to a human error in a script.
• Netris technology was subsequently adopted to resolve these issues and prevent recurrence.
• Netris currently manages approximately 35 AI clusters globally, nearing a total of 1 million GPUs.
• The company was recently highlighted in Jensen Huang's presentation at a major NVIDIA event.
• Netris deploys roughly 2 to 3 new AI clusters monthly, accumulating a "learning reservoir" of industry-wide deployment nuances.

AI Cluster Networking Complexity

• AI clusters require a distinct five-layer networking architecture compared to traditional data centers:
  • Edge Networking: Facilitates user communication with hosted GPUs over the internet.
  • North-South Layer: Typically Ethernet-based, connecting servers, GPU servers to storage, and the edge to the frontend.
  • East-West Layer: Connects GPUs across racks, utilizing InfiniBand or Ethernet.
  • Scale-Up Networking: NVIDIA's NVL72 technology connecting GPUs within a single rack.
  • Host/DPU Networking: Handles networking on the GPU host itself.
• Scale challenges are illustrated by a specific GB300 deployment of 18,000 GPUs containing 177,000 cables.
• Network reconfiguration is required for every tenant onboarding, offboarding, or resizing across all network layers.
• Manual configuration via switch commands is deemed insufficient for these scales due to complexity and high risk of total network failure.

Automation Strategy & Operational Shifts

• GPU operators attempting to build in-house automation (using Ansible, Python) face high maintenance burdens as networks evolve.
• Industry trend shows a shift from home-grown solutions to professional commercial software for network automation.
• Netris utilizes "Network Automation Abstraction and Multitenancy" (NAM) as the next generation of networking beyond SDN.
• NAM productizes collective industry learnings from 35+ live deployments to avoid known nuanced pitfalls.

Multi-Tenancy & Security Requirements

• Multi-tenancy is critical for ROI and is required for both AI cloud operators and enterprise AI factories.
• Netris enforces "hard multi-tenancy" where isolation is guaranteed at the hardware network layer, not just via software or Kubernetes namespaces.
• Software-level isolation (e.g., namespaces) is insufficient due to risks of container escape and data access by other tenants.
• Single-Tenant Case Study: Even clusters dedicated to a single "off-taker" require internal multi-tenancy separation.
  • This separation isolates the "service tenant" (for hardware maintenance) from the "off-taker tenant" (customer data).
  • Protocols ensure technicians performing maintenance cannot accidentally access customer data, satisfying strict contractual security requirements.
  • Off-takers can use web interfaces to release specific failed machines from their tenant isolation for maintenance and return them upon repair.

Product Capabilities & Ecosystem

• Abstraction Layer: Provides a unified API for customer-facing portals, allowing requests for compute, elastic IPs, or load balancers without code changes regardless of underlying hardware (GB300, GB200, Vera Rubin, InfiniBand, or Ethernet).
• Developer Efficiency: Reduces time to design, deploy, and simulate clusters while minimizing the risk of accidental network breakage.
• Ecosystem Integration: Netris is designed to integrate with:
  • Storage: VAST, Weka, DDN.
  • Orchestration: Red Hat OpenShift, Dcluster, Rancher, SpectroCloud.
  • NVIDIA Tools: NVIDIA NetOps and BCM.
• Future Outlook: The technology is intended to manage the complexity of next-generation architectures like Vera Rubin as they are introduced to the market.