Best Miami News connects businesses and publishers

collapse
Home / Daily News Analysis / Nvidia-backed Upscale takes on AI networking silos with Token Fabric

Nvidia-backed Upscale takes on AI networking silos with Token Fabric

Oct 22, 2026  Twila Rosenbaum 60 views
Nvidia-backed Upscale takes on AI networking silos with Token Fabric

Upscale Targets the Divide Between Two AI Networks

AI clusters rely on two distinct networking domains. Scale-up networking connects accelerators inside a rack, while scale-out networking links racks across a data center. A single AI training or inference job often spans both domains, but many organizations manage them as separate networks with separate teams, tools and operating models. That split creates operational complexity and can limit performance when workloads expand from a handful of accelerators to large clusters.

Upscale, a Santa Clara, California-based networking startup backed by Nvidia, is trying to close that gap. The company introduced Token Fabric, an architecture that combines its own SkyFabriX scale-up silicon, scale-out systems built on Nvidia Spectrum-X Ethernet and a shared software layer for operations. Upscale plans general availability in early 2027. Early-access and joint-validation programs are already running.

The announcement lands as AI infrastructure spending shifts toward larger clusters and more demanding workloads. Networking has become a critical bottleneck because accelerators can sit idle when data cannot move quickly enough between them. Upscale CEO Barun Kar said compute is ramping up at a rapid scale while networking has fallen behind. Token Fabric is the company’s answer to that imbalance.

Company Background and Funding

Upscale emerged from stealth in September 2025 with a $100 million seed round. In June, it raised an additional $190 million and outlined plans for SkyHammer, a custom scale-up switch chip. SkyFabriX is based on the SkyHammer architecture. Nvidia, which supplies the Spectrum-X silicon used in Upscale’s scale-out systems, joined the June round. Upscale has now raised about $500 million and has more than 300 employees.

That funding profile matters in AI networking, where hardware development cycles are long and customers demand validated systems rather than individual components. Building silicon, systems and software requires deep capital and engineering resources. Upscale is positioning itself as a vertically integrated supplier that can deliver pieces of the stack or the full architecture depending on the customer.

Token Fabric Brings Silicon, Systems and Software Under One Architecture

Token Fabric is not a single switch or a single software product. It is a combination of hardware and software spanning both scale-up and scale-out networking. The architecture is intended to give operators one operational view across a cluster even when the underlying traffic types and protocols differ.

SkyFabriX Scale-Up Silicon

SkyFabriX is the scale-up switch silicon at the center of Upscale’s design. It supports ESUN, or Ethernet for Scale-Up Networking, as well as standard Ethernet/IP. It is also designed for evolving protocols such as UALink, the Ultra Accelerator Link standard. The switch capacity is 115.2 Tbps, with a roadmap to multiple petabits per second.

Scale-up domains are where accelerators communicate tightly, often requiring low latency and high bandwidth. Pooling GPUs and other accelerators into larger domains can improve utilization for AI workloads, but it also raises the bar for switch silicon. Upscale’s claim is that custom silicon can support both current Ethernet-based approaches and newer scale-up standards as they mature.

Scale-Up Switch Trays

Upscale offers scale-up switch trays in custom and standard rack form factors. These trays pool GPUs and other accelerators into large scale-up domains. The form-factor flexibility is important because AI infrastructure is not standardized across cloud providers, enterprises and system vendors. Some customers want custom racks; others want standard building blocks that fit existing data-center designs.

Scale-Out Systems

Upscale builds its scale-out systems around Nvidia Spectrum-X silicon at speeds from 400G and 800G up to 1.6T. With SkyOS, the systems integrate into Nvidia and other accelerator clusters. Scale-out networking carries traffic between racks and across the data center, so Ethernet and RoCE have become common choices for AI fabrics that need to interoperate with broader infrastructure.

SkyOS and SkyCMD

SkyOS is the network operating system for both fabrics. It is based on SONiC, the open-source Software for Open Networking in the Cloud, and optimized for AI protocols. SkyCMD is the orchestration and observability layer. It provides a single orchestration plane across scale-up and scale-out, giving operators a common way to manage what have traditionally been separate network domains.

Three Consumption Models for Different AI Network Teams

Upscale is offering Token Fabric through three consumption models. Customers can adopt silicon, systems, software or the full stack. The full-stack option targets neoclouds and enterprises that do not have hyperscaler-sized network teams. An integrated systems option targets XPU vendors, meaning makers of non-GPU accelerators that want validated hardware without building a complete network stack. A third option provides silicon or custom systems with a software development kit for hyperscalers that run their own network operating system.

All three options draw on the same portfolio of silicon, systems and software. Kar said everything has to be integrated, which is where Upscale comes in. The company describes itself as vertically stacked. That model lets Upscale sell to customers with different levels of networking expertise while keeping a common architecture underneath.

Protocol Strategy: Extend, Don’t Replace

Token Fabric is built from the protocol layer up. Upscale is extending existing protocols rather than replacing them. Scale-up traffic uses ESUN and standard Ethernet/IP. Scale-out traffic uses standard Ethernet with RoCE, which carries remote direct memory access over Ethernet. The approach is designed to preserve compatibility with existing operational practices and with the broader Ethernet ecosystem.

Upscale is also contributing to open projects underneath the stack. Those include SONiC, extensions of the Switch Abstraction Interface for ESUN, the Ultra Ethernet Consortium and UALink. Open-source and standards work is increasingly important in AI networking because customers do not want a proprietary island that cannot interoperate with accelerators, switches and management tools from multiple vendors.

The protocol choices reflect a broader industry debate. InfiniBand has been a long-standing option for high-performance AI clusters, while Ethernet vendors have pushed improvements for AI workloads. Open standards such as UALink and Ultra Ethernet aim to create more choice at the accelerator and network layers. Upscale’s strategy is to support those efforts while providing integrated products that can be deployed before every standard is fully settled.

Software as the Unifying Layer

Software ties the two fabrics together. SkyOS abstracts the underlying hardware and provides a control plane across the cluster. SkyCMD exposes that control plane through one layer. Multiple network elements sit behind it, and any compute platform in a heterogeneous cluster can use it to control the network.

Kar said the operating system has to be lean and mean to be fast, flexible and secure. The orchestration layer on top enables capabilities such as heterogeneous compute. That matters because AI clusters increasingly mix different accelerators, CPUs and storage systems. A network that only manages one vendor’s hardware or one traffic type would recreate the silos Token Fabric is meant to remove.

The common software layer also has implications for operations. Network teams can use the same tooling and telemetry across scale-up and scale-out domains instead of stitching together separate management systems. That can reduce configuration drift, simplify troubleshooting and make it easier to automate changes across a large cluster.

Tokens as the New Networking Metric?

Network teams have traditionally tracked packets, throughput, loss and jitter. The question for an AI network is whether tokens are the right unit of optimization. Kar argued that tokens are the right measure. Users pay based on tokens today, and he named time to first token, tokens per second, tokens per dollar and tokens per watt as the key measures.

Optimizing those measures requires networking, especially in clusters that scale to hundreds of thousands of accelerators. A network that adds latency or drops traffic can delay the delivery of tokens and reduce the effective utilization of expensive compute. Tokens per dollar and tokens per watt connect network performance directly to business and energy costs, which is a different framing from traditional packet-level metrics.

There is a gap between that goal and many networks today. According to Kar, the main components of the gap are latency, bandwidth and scale-up features in the silicon. On the software side, predictive analytics and telemetry keep machines running. Token Fabric is designed to address those areas together rather than treating them as separate purchasing decisions.

Agentic AI Moves Into Network Operations

The other side of the challenge is optimizing the network with agentic AI. Upscale has built agent-based operations directly into the platform. Kar said an agent is always looking at the network, trying to figure out whether cables will fail, optics will fail or congestion will appear, and the agents look for ways around those problems.

Agentic operations could be especially useful in large AI clusters where manual troubleshooting is slow and failures are costly. If a link degrades or congestion builds, an agent could identify the issue and suggest or execute a remediation path. The effectiveness of that approach will depend on telemetry quality, guardrails and how much autonomy customers are willing to grant. Upscale’s bet is that the same integrated software layer that unifies scale-up and scale-out can also support automated, ongoing network optimization.


Source:Network World News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy