Upscale, a Santa Clara, California-based networking startup with backing from Nvidia, has introduced Token Fabric, an architecture intended to bring together the two halves of AI cluster networking: scale-up and scale-out. The company says the design combines its own SkyFabriX scale-up silicon, scale-out systems built on Nvidia Spectrum-X Ethernet, and a shared software layer for operations. General availability is planned for early 2027, with early-access and joint-validation programs already under way.

The launch addresses a structural problem in AI infrastructure. Scale-up networking connects accelerators inside a rack, while scale-out networking connects racks across a data center. A single AI training or inference job can span both domains, but many organizations manage them as separate networks with separate tools and teams. That split becomes more costly as clusters grow and as AI workloads place stricter demands on latency, bandwidth and utilization.

Key facts at a glance

• Upscale has introduced Token Fabric, which combines scale-up and scale-out networking for AI clusters.
• The architecture includes SkyFabriX scale-up switch silicon, scale-out systems based on Nvidia Spectrum-X Ethernet, and the SkyOS network operating system with SkyCMD orchestration.
• SkyFabriX supports ESUN and standard Ethernet/IP, is designed for evolving protocols such as UALink, and offers 115.2 Tbps of switch capacity with a roadmap to multiple petabits per second.
• Scale-out systems are available at 400G, 800G and 1.6T speeds and are designed to integrate with Nvidia and other accelerator clusters.
• Upscale plans general availability for early 2027 and is running early-access and joint-validation programs now.
• The company emerged from stealth in September 2025 with a $100 million seed round, raised an additional $190 million in June, and has now raised about $500 million. It has more than 300 employees.
• Nvidia joined the June funding round and supplies the Spectrum-X silicon used in Upscale's scale-out systems.

Upscale CEO Barun Kar framed the problem in terms of pace. Compute is ramping up on a rapid scale, he said, while networking has fallen well behind. The imbalance matters because AI clusters are increasingly judged by how quickly and efficiently they can produce tokens, not simply by how many accelerators they contain.

A single stack for two fabrics

Token Fabric is built from hardware and software for both scale-up and scale-out. At the silicon level, SkyFabriX is the company's scale-up switch silicon. 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. The switch has a capacity of 115.2 Tbps, and Upscale has outlined a roadmap to multiple petabits per second.

The company offers scale-up switch trays in custom and standard rack form factors. Those trays pool GPUs and other accelerators into larger scale-up domains, allowing more compute resources to communicate as a single high-performance system. For scale-out, Upscale builds systems around Nvidia Spectrum-X silicon at speeds from 400G and 800G up to 1.6T. With SkyOS, those systems are designed to integrate into Nvidia and other accelerator clusters.

Software is the connective tissue. 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 is optimized for AI protocols. SkyCMD provides orchestration and observability, exposing a single orchestration plane across scale-up and scale-out. That means multiple network elements can sit behind one layer while compute platforms in a heterogeneous cluster use it to control the network.

Kar described the operating system and orchestration layer as essential to performance. An operating system has to be lean and efficient to be fast, flexible and secure, he said, and an orchestration layer on top is needed to support capabilities such as heterogeneous compute.

Consumption models for different customer types

Upscale is not forcing customers to take the entire stack. The company says buyers can adopt silicon, systems, software or the full combination. The full-stack option targets neoclouds and enterprises that do not have hyperscaler-sized network engineering teams. An integrated systems option targets XPU vendors, meaning makers of non-GPU accelerators, that want validated hardware without building a complete network stack themselves. 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's argument is that integration is the differentiator. Everything has to be integrated, he said, and that is where the company comes in. Upscale describes itself as vertically stacked, meaning it controls the silicon, systems and software layers rather than relying on a patchwork of third-party components.

Extending protocols rather than replacing them

Token Fabric is built from the protocol layer up, but Upscale says it 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 intended to preserve compatibility with established networking practices while adding the capabilities AI workloads need.

Upscale is also contributing to open projects underneath the stack. Those efforts include SONiC, extensions to the Switch Abstraction Interface for ESUN, the Ultra Ethernet Consortium and UALink. The company's participation matters because AI networking standards are still evolving. Customers want to avoid locking themselves into proprietary protocols, but they also need silicon and software that can support new accelerator interconnects as they mature.

The strategy places Upscale in a crowded and fast-moving part of the infrastructure market. AI clusters are growing larger, and the network is increasingly seen as a determinant of overall system efficiency. A cluster can have thousands of expensive accelerators, but if the network cannot keep them fed with data or allow them to communicate with low latency, utilization falls and costs rise. That is why networking startups and established chip vendors are both investing heavily in scale-up and scale-out technologies.

Are tokens the right metric for network performance?

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 indeed 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 relevant measures. Optimizing those measures requires networking, especially in clusters that scale to hundreds of thousands of accelerators.

There is still a gap between traditional network operations and token-oriented optimization. According to Kar, the main components of that gap are latency, bandwidth and scale-up features in the silicon. On the software side, predictive analytics and telemetry help keep machines running. The goal is to move from reactive troubleshooting to a model in which the network is continuously observed, tuned and protected against failures that would otherwise reduce token throughput.

The other side of the challenge is optimizing networking 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, whether optics will fail or whether there is congestion, and the agents look for ways around those problems. That kind of automation becomes more important as clusters grow and as human operators struggle to keep pace with the volume of telemetry and the speed of AI workloads.

Funding, backing and the road to availability

Token Fabric builds on two earlier milestones for Upscale. The company 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, its custom scale-up switch chip. SkyFabriX is based on that SkyHammer architecture. Nvidia, which supplies the Spectrum-X silicon, joined the June round. Upscale has now raised about $500 million and has more than 300 employees.

The backing from Nvidia gives Upscale a connection to one of the most important suppliers in AI infrastructure, but the startup is not limiting itself to Nvidia accelerators. Its scale-up silicon supports standard Ethernet/IP and is designed for UALink, while its software layer is intended to work across heterogeneous compute environments. That positioning could help Upscale appeal to neoclouds, enterprises, XPU vendors and hyperscalers that want different levels of integration.

The competitive landscape is likely to intensify before Token Fabric reaches general availability. AI networking has become a focal point for chipmakers, switch vendors, cloud providers and startups, all of which are trying to reduce bottlenecks between accelerators and across data center fabrics. Upscale's bet is that customers will prefer a vertically integrated stack that spans both scale-up and scale-out, rather than assembling separate solutions for each domain.

For now, the company is focused on early-access and joint-validation programs. General availability is planned for early 2027. The success of Token Fabric will depend on whether Upscale can deliver silicon, systems and software that work together at scale, and whether customers are willing to adopt a newer networking stack as AI clusters continue to expand.

Source:Network World News