Upscale AI: How to Scale Open, Heterogeneous AI Data Centres

As companies navigate the demands of intensive AI workloads, Upscale AI unveiled plans to enhance open scaling across AI clusters.

Upscale AI and NVIDIA are providing Ethernet-based networking infrastructure designed for scalable, heterogeneous AI cluster deployments.

The initiative centres on Ethernet systems built using NVIDIA Spectrum X Ethernet switch silicon, combined with Upscale AI's SONiC software, optimised specifically for AI applications.

The deployment targets scalable, low-latency fabrics designed to handle today's demanding AI workloads across heterogeneous architectures spanning compute, accelerators, memory and storage.

"NVIDIA Spectrum-X Ethernet switch silicon is setting a new standard for Ethernet-based AI performance," says Barun Kar, CEO of Upscale AI.

"Pairing this technology with our purpose-built systems and AI-optimised SONiC software allows us to deliver the best of both worlds: an open, highly scalable and interoperable architecture with operational simplicity."

Upscale AI is joining the NVIDIA Partner Network, strengthening its position as what it describes as a "pure-play provider of NVIDIA AI-native networking infrastructure".

This membership offers Upscale AI opportunities to work within the NVIDIA ecosystem, where reference architectures and validated designs support the company's ambition to scale AI data centre networks.

Addressing heterogeneous AI clusters

To reduce complexity when heterogeneous AI clusters operate together, Upscale AI supports open and interoperable Ethernet networking.

The company's AI fabric has been designed with consideration for the diversity of compute environments, enabling deployment using multi-vendor, production-grade AI infrastructure whilst maintaining seamless operation.

AI-optimised scale-out systems, built on NVIDIA hardware, integrate ASIC-native telemetry, where chips directly generate and export real-time performance data, potentially delivering lossless Ethernet behaviour.

Additional advantages include predictable performance, operational simplicity and reliability at scale.

The heterogeneous approach allows organisations to mix different hardware vendors and computing architectures within a single AI cluster without sacrificing performance or operational consistency.

This flexibility proves particularly valuable for enterprises that have made previous infrastructure investments or require specific hardware capabilities for different workload types.

By supporting open standards, Upscale AI's solution enables data centres to avoid vendor lock-in while maintaining the performance characteristics required for modern AI training and inference workloads.

Industry collaboration and market outlook

"To lead in the trillion-parameter model era, scalability and efficiency are paramount," notes Gilad Shainer, SVP at NVIDIA.

"We look forward to collaborating closely with Upscale AI as the team leverages the NVIDIA Spectrum-X Ethernet platform to help companies build the world's most advanced open AI infrastructure."

Upscale AI's approach offers a full stack solution that facilitates high-speed data movement, workload isolation and scalable orchestration in environments containing different types of compute architectures.

By combining hardware, software and lifecycle services, the Spectrum X Ethernet scale-out systems preserve the flexibility of open-source management.

The collaboration reflects broader industry momentum towards standardised, Ethernet-based networking for AI workloads.

As AI models continue to grow in size and complexity, the networking infrastructure connecting compute resources becomes increasingly critical to overall system performance.

Market analysts suggest that organisations deploying large-scale AI infrastructure are prioritising solutions that balance performance with operational flexibility.

Practical deployment for enterprises

"As AI infrastructure evolves toward increasingly heterogeneous architectures, scalable and operationally sound Ethernet fabrics are becoming essential," explains Alan Weckel, Co-Founder and Technology Analyst at 650 Group.

"By combining NVIDIA Spectrum-X Ethernet switch silicon with a full-stack approach that integrates systems, software and support, Upscale AI is addressing a key gap in making open, scale-out AI networking practical for large scale AI deployments."

According to Upscale AI, these scale-out systems will become available later in 2026 for AI data centres building diverse, multi-vendor infrastructure. The timing could prove significant for organisations seeking to expand their AI capabilities whilst maintaining operational flexibility.

For organisations currently operating AI workloads on disparate infrastructure, Upscale AI's solution offers a migration path that preserves existing investments whilst enabling future expansion.

The company's full-stack approach aims to reduce the operational burden typically associated with managing complex, multi-vendor AI networking environments.