BIP NYC

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

Upscale introduces Token Fabric for converged AI networking

AI infrastructure has grown around two distinct network domains: scale-up networking that connects accelerators inside a rack and scale-out networking that links racks across a data center. A single AI training or inference job can span both domains, yet many organizations operate them as separate fabrics with separate tools, teams, and operational models. Upscale, a Santa Clara, Calif.-based networking startup with Nvidia backing, is trying to remove that division with Token Fabric, an architecture announced today that combines scale-up silicon, scale-out Ethernet systems, and a shared software layer.

Token Fabric is not a single switch or a single operating system. It is a vertically integrated portfolio that includes Upscale’s SkyFabriX scale-up switch silicon, scale-out systems built on Nvidia Spectrum-X Ethernet, switch trays, the SkyOS network operating system, and SkyCMD orchestration and observability software. The company plans general availability in early 2027 and says early-access and joint-validation programs are already running.

Why AI networking has been split in two

Scale-up networking is the high-bandwidth, low-latency domain that pools accelerators into a larger compute unit. It is often tied to proprietary or emerging accelerator-link technologies. Scale-out networking is the fabric that connects those pooled systems across racks and rows, using Ethernet or InfiniBand. The two domains have different traffic patterns, latency targets, and failure modes. Scale-up traffic is dominated by tightly coupled accelerator-to-accelerator communication. Scale-out traffic must move data across longer distances and among many more endpoints.

That division made sense when AI clusters were smaller. It becomes harder to defend as models grow and as clusters expand to hundreds of thousands of accelerators. A single job may consume capacity across both fabrics, so a bottleneck or management gap in one domain can limit the entire system. Separate operational stacks can also slow troubleshooting when congestion, optics failures, or cable problems appear. Upscale’s pitch is that the two networks should be designed, operated, and observed as one system.

“If you look at compute, compute is ramping up on a rapid scale, and networking has fallen way behind,” said Barun Kar, CEO of Upscale. The implication is that networking can no longer be treated as a collection of discrete boxes and domains. It must be integrated with the compute it serves.

What is inside Token Fabric

Token Fabric spans hardware and software for both scale-up and scale-out. Upscale is offering the pieces individually or as a full stack, depending on the customer’s engineering resources and operating model.

  • SkyFabriX silicon. SkyFabriX is the scale-up switch silicon. It supports ESUN, or Ethernet for Scale-Up Networking, and standard Ethernet/IP. It is also designed for evolving protocols such as UALink, the Ultra Accelerator Link. The switch capacity is 115.2 Tbps, with a roadmap to multiple petabits per second.
  • Scale-up switch trays. These trays come in custom and standard rack form factors. They pool GPUs and other accelerators into larger scale-up domains, giving customers a way to build high-bandwidth compute blocks without designing every element from scratch.
  • Scale-out systems. Upscale builds 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-based clusters and other accelerator environments.
  • 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.
  • 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-size 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. “Everything has to be integrated, so that’s where we come in,” Kar said. “We are vertically stacked.”

Funding, silicon roadmap, and Nvidia backing

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 the SkyHammer architecture. Nvidia, which supplies the Spectrum-X silicon used in Upscale’s scale-out systems, joined that June round. Upscale has now raised about $500 million and has more than 300 employees.

The funding history matters because AI networking startups face long development cycles, complex silicon design, and demanding customer validation. Building switch silicon, a network operating system, orchestration software, and full systems requires a broad engineering effort. Upscale’s strategy is to own enough of the stack to deliver an integrated product while still offering components to customers that want to build their own platforms.

The SkyHammer and SkyFabriX roadmap also points to the scale of the problem. AI clusters are pushing beyond the capacity of conventional switch designs, and scale-up domains are growing larger. Upscale says SkyFabriX starts at 115.2 Tbps and has a path to multiple petabits per second. That trajectory reflects the bandwidth demands of tightly coupled accelerator systems, where every additional accelerator increases the pressure on the switching fabric.

Protocol strategy: extend rather than replace

Token Fabric is built from the protocol layer up. 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. That choice matters because customers do not want to throw away operational knowledge or interoperability with existing Ethernet environments.

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. The involvement in open standards is a signal that the company is trying to position Token Fabric as part of a broader Ethernet ecosystem rather than as a proprietary island.

The networking industry has spent years improving Ethernet for AI and cloud workloads. RoCE helped Ethernet carry remote direct memory access traffic with lower overhead. SONiC created an open network operating system that could run across multiple switch platforms. The Ultra Ethernet Consortium has been working on Ethernet enhancements for AI and high-performance networking. UALink and ESUN are aimed at the scale-up domain, where accelerator-to-accelerator links have historically been specialized. Token Fabric sits at the intersection of those efforts.

Software ties the two fabrics together

Hardware alone does not solve the silo problem. 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.

That matters for mixed environments. AI clusters are not always homogeneous. They may combine GPUs from one vendor with accelerators from another, or they may mix generations of hardware as capacity is added. A common orchestration layer can reduce the need for separate management tools for each fabric and each compute platform. It can also give operators a single view of network state across scale-up and scale-out domains.

“Your operating system has to be lean, mean in order to be fast, flexible, and secure,” Kar said. “And then you have to have the orchestration layer on top to enable things like heterogeneous compute.” The message is that the network operating system must be efficient enough for demanding AI traffic while the orchestration layer provides the policy, automation, and visibility needed for mixed infrastructure.

Are tokens the new 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. Token metrics connect network performance to the economics of AI services. Customers pay for tokens, and operators care about how quickly they can produce them, how many they can produce per second, and how much they cost in dollars and watts.

“I would say that’s the right thing,” Kar argued. He named time to first token, tokens per second, tokens per dollar, and tokens per watt as the measures that matter. 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 reduce token throughput and raise cost per token.

There is a gap between token-level goals and many networks today. According to Kar, the main components of that gap are latency, bandwidth, and scale-up features in the silicon. On the software side, he said predictive analytics and telemetry keep machines running. Token optimization is not just about raw bandwidth. It is about delivering the right data to the right accelerator at the right time, with predictable latency and minimal disruption.

Agentic AI in network operations

Upscale has built agent-based operations into the platform. The idea is to use AI agents to monitor the network continuously and identify problems before they cause job failures. That includes watching for cable failures, optics failures, and congestion, then looking for ways around those issues.

“An agent is always looking at your network, trying to figure out whether your cables will fail or your optics will fail or if there’s congestion, and the agents are looking for ways around it,” Kar said. For large AI clusters, that kind of continuous monitoring is important because a single failing link can affect a large job. Agents can help operators move from reactive troubleshooting to proactive management.

The combination of token metrics and agentic operations reflects a broader shift in AI networking. The network is no longer just plumbing between servers. It is part of the performance and economics of AI compute. As clusters grow, the ability to observe, optimize, and remediate network behavior becomes as important as the switching capacity itself.

Consumption models and target customers

Upscale’s consumption models are designed to match different levels of customer sophistication. Neoclouds and enterprises without hyperscaler-size network teams can choose the full stack. That option gives them integrated hardware and software without requiring them to assemble a complete AI network from separate components. XPU vendors can choose integrated systems that have been validated with Upscale’s networking, which can shorten development and validation cycles. Hyperscalers that run their own network operating system can choose silicon or custom systems with a software development kit.

The common thread is integration. Upscale is not trying to force every customer into one delivery model. It is trying to make the same underlying silicon, systems, and software available in different forms. That flexibility could help the company address customers that want a turnkey solution as well as those that want to build their own network stack.

Token Fabric’s planned general availability in early 2027 will be a key test. Early-access and joint-validation programs are already running, which suggests the company is working with customers before broad commercial release. The architecture’s success will depend on whether Upscale can deliver integrated operations across both fabrics without sacrificing the openness, interoperability, and performance that AI operators require.


Source: Network World News


Share:

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