Upscale AI

What Upscale AI does

Upscale AI delivers scalable, standards-based open AI backend networks designed to meet the speed, scale, and control that AI workloads require. The company provides full-stack AI networking solutions for various infrastructure challenges, from rack-scale training clusters to cloud-scale deployments. Their solutions are purpose-built for modern AI infrastructure, addressing the limitations of general-purpose networks that were not designed for the complexity of AI. Upscale AI focuses on open standards, flexibility for heterogeneous compute, and economic efficiency for AI workloads, spanning silicon, systems, and software.

Products

• SkyHammer™A scale-up networking solution designed to deliver deterministic, low-latency performance for synchronized AI workloads. Built from the ASIC up, it reduces communication bottlenecks across accelerators, memory, and storage, enabling large-scale training and inference to run faster, more efficiently, and with more predictable performance. It unifies distributed GPU clusters into a single, memory-coherent system, delivering a lossless, deterministic fabric with bounded tail latency using open standards.
• Open Ethernet FabricsScale-out solutions providing open, interoperable Ethernet systems built on NVIDIA Spectrum-X switch silicon and a SONiC-based network operating system. Designed for heterogeneous AI clusters, these systems connect accelerators, memory, and storage into a high-performance fabric for distributed training and large-scale inference. They are engineered to improve effective bandwidth, reduce network bottlenecks, and bring greater predictability to large GPU clusters.
• Unified SONIC SubstrateProvides consistent control, telemetry, and operations across scale-up and scale-out AI networking environments.

Key capabilities

• Scalable, standards-based open AI backend networks
• Designed for speed, scale, and control of AI workloads
• Full-stack platform spanning silicon, systems, and software
• Purpose-built economics for AI workloads
• Reduces network bottlenecks and improves efficiency
• Supports heterogeneous compute and evolving infrastructure requirements
• Deterministic, low-latency scale-up networking (SkyHammer)
• Reduces communication bottlenecks across accelerators, memory, and storage
• Open, interoperable Ethernet systems (Open Ethernet Fabrics)
• Built on NVIDIA Spectrum-X switch silicon
• SONiC-based network operating system
• Connects accelerators, memory, and storage into high-performance fabrics
• Consistent, low-latency throughput for AI inference
• Sub-microsecond latency and zero packet loss for AI training
• Maximizes accelerator utilization for large-scale, synchronized model training
• Hardware-level efficiency and reliability for production rack-scale deployments
• Visibility and control for large clusters at scale

Use cases

• Rack-scale training clusters
• Cloud-scale deployments
• AI Inference at production scale
• Large-scale AI training
• Synchronized AI workloads
• Distributed AI clusters
• Token serving at production scale

AI approach

Upscale AI designs and builds networking infrastructure specifically for AI workloads. Their products, SkyHammer and Open Ethernet Fabrics, are purpose-built from the ASIC up to handle the speed, scale, and complexity of modern AI, reducing bottlenecks and improving efficiency for large-scale training and inference. They focus on open standards and full-stack solutions spanning silicon, systems, and software.

Tech named: NVIDIA Spectrum-X switch silicon, SONiC-based network operating system, UltraEthernet, OCP SAI, OCP ESUN, UALink

What it says sets it apart

• Built from the ground up for AI networking, not patching legacy architecture
• Focus on open standards (UltraEthernet, OCP SAI, OCP ESUN, SONIC, UALink)
• Full-stack platform spanning silicon, systems, and software
• Purpose-built economics for AI workloads
• Flexibility to support heterogeneous compute and adapt across evolving infrastructure requirements
• Deterministic, low-latency scale-up networking (SkyHammer)
• Open, interoperable Ethernet systems built on NVIDIA Spectrum-X switch silicon and SONiC
• No legacy architecture to work around, no ecosystem to protect
• Actively shaping open standards for AI networking

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.