Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Upscale’s Token Fabric is a launch-stage networking platform designed to combine two parts of an AI cluster that operators often manage separately: scale-up connections within a rack and scale-out connections between racks. It brings together Upscale’s SkyFabriX scale-up switch silicon, scale-out systems based on NVIDIA Spectrum-X Ethernet, and software intended to operate and orchestrate both fabrics from a common view. The design is a product announcement, not yet proof of better performance or broad compatibility across chip vendors.
What problem is Token Fabric meant to solve?
A large AI job can rely on both fast communication among accelerators within a rack and network links connecting racks across a cluster. Those scale-up and scale-out fabrics have often been built and managed as separate systems. Upscale’s thesis is that operating them together could make it easier to manage the network supporting an AI workload end to end.
CEO Barun Kar told Network World, “If you look at compute, compute is ramping up on a rapid scale, and networking has fallen way behind.” The company describes its intended optimization around AI tokens, but the launch coverage does not establish tokens as a validated or industry-wide network-performance metric.
What does Token Fabric include?
Token Fabric is a combined hardware and software stack, rather than a standalone dashboard. Network World’s October 8, 2026 report describes these components:
#1 Best Overall
- Supports optical fiber cable to span longer distances and provided high data transmission rates between servers and network components
- 100 Gigabit Ethernet provides high bandwidth performance, ease of use and reliability for your network backbone
- Supports layer 3 switching for enhanced performance and usability
- Power Supply power source for dependable power supply
- Rail mounting is ideal for factory areas, in vehicles and marine applications
| Component | Role |
|---|---|
| SkyFabriX | Upscale’s scale-up switch silicon, described as supporting ESUN and standard Ethernet/IP and designed with evolving protocols such as UALink in mind. Network World reports switch capacity of 115.2 Tbps and a roadmap to multiple petabits per second; these are reported specifications and plans, not independent benchmark results. |
| Scale-up trays | Custom and standard rack-form-factor switch trays intended to pool GPUs and other accelerators into larger scale-up domains. |
| Scale-out systems | Systems based on NVIDIA Spectrum-X Ethernet. The report cites 400G, 800G, and 1.6T speeds. |
| SkyOS | A network operating system for both fabrics, described as based on SONiC. |
| SkyCMD | An orchestration and observability layer intended to provide a common control plane across the fabrics. |
In the reported design, scale-up traffic uses ESUN and standard Ethernet/IP; scale-out uses standard Ethernet with RoCE. Upscale also contributes to SONiC, SAI extensions for ESUN, the Ultra Ethernet Consortium, and UALink, according to Network World. Support for standards or open-source projects should not be taken as evidence that every combination of accelerators, switches, and software has been validated.
How is it intended to be deployed?
Upscale describes three adoption models, as reported by Network World:
Rank #2
- Flexible 10/25/40/100GbE Interface Speeds, Support Stacking; Achieve Reliable Links With 25G NVIDIA Mellanox NIC; CE, Reach, RoHS, FCC, WEEE Certified;
- Full stack: For neoclouds and enterprises that do not have hyperscaler-sized networking teams.
- Integrated systems: For XPU vendors seeking validated hardware without building their own complete networking stack.
- Silicon or custom systems with an SDK: For hyperscalers that want hardware options while running their own network operating system.
These are company-described options, not evidence of signed customers or completed deployments. Reuters reports that Upscale is targeting neoclouds and large cloud providers looking to use chips from multiple suppliers, but the cited launch coverage does not show that broad multi-vendor interoperability has already been demonstrated.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
When is Token Fabric expected to be available?
Availability remains a plan, not a completed general release. Network World reported on October 8, 2026 that early-access and joint-validation programs were underway and general availability was planned for early 2027. Reuters reported an initial piece planned for Q4 2026 and a staged full rollout through 2027. These timelines describe reported plans and may change.
Rank #3
- Wired connection speed up to 1000 Mbps
- 5 Gigabit Ethernet auto sensing ports
- Easy plug & play connection
- Quality of Service (QoS) traffic prioritization
- Meets EEE (Energy Efficient Ethernet) 802.3az standard
Reuters also reported that CEO Barun Kar expected Token Fabric to generate tens of millions of dollars in revenue in 2027, potentially reaching the low hundreds of millions. That is a forecast, not achieved revenue. In June 2026, Upscale raised a reported $190 million financing extension, bringing total funding to $500 million at a reported $2 billion valuation, according to Reuters via Investing.com.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What remains unproven?
The launch reports describe architecture and specifications, but do not document an independent benchmark, completed customer deployment, comparative test, or broad validation across vendors. Before treating Token Fabric as a proven fit for a cluster, infrastructure teams would need evidence that matches their own requirements, including:
Rank #4
- Full duplex IEEE 802.3x flow control and half duplex backpressure with intelligent port based congestion detection and broadcast rate control
- All ports support auto MDI/MDI X cable detection
- Non head of line blocking architecture
- QoS: Prioritize video and audio applications for better streaming (IEEE 802.1p)
- Power Save: Advanced power saving mode detects unused ports
- Whether the scale-up and scale-out layers cover the intended workloads and cluster topology.
- Which accelerator, switch, and software combinations have actually been validated.
- Which protocols and standards are supported in the specific configuration being considered.
- Whether SkyOS and SkyCMD provide the required operational control and observability across both fabrics.
- Which deployment model is available, and whether it fits the organization’s existing network operations.
- Availability in the relevant timeframe and independently measured performance under comparable conditions.
Kar described an agent-based operating concept to Network World: “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.” This is the CEO’s description of the platform’s intended behavior; the launch coverage does not verify those capabilities in production.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
