Upscale Introduces Token Fabric to Unify AI Scale-Up and Scale-Out Networking
Networking startup Upscale announced Token Fabric, a unified hardware and software architecture designed to bridge the gap between intra-rack and cross-rack AI cluster connectivity.

Key points
- Upscale launched Token Fabric to integrate scale-up and scale-out networking into a single operational layer.
- The architecture combines custom SkyFabriX silicon with Nvidia Spectrum-X Ethernet systems.
- General availability is scheduled for early 2027, with early-access programs currently active.
Networking startup Upscale introduced Token Fabric, a new architecture designed to unify scale-up and scale-out networking for artificial intelligence clusters. The announcement addresses the operational complexity organizations face when managing separate networks for intra-rack accelerator connections and cross-rack data center communication. According to Network World, the solution aims to treat these distinct layers as a cohesive unit rather than isolated silos.
Token Fabric combines Upscale’s proprietary SkyFabriX scale-up silicon with scale-out systems built on Nvidia Spectrum-X Ethernet. A common software layer manages operations across both domains. General availability for the full stack is planned for early 2027. Early-access and joint-validation programs are currently running for prospective customers. The company intends to allow adoption of individual components or the complete integrated stack.
Technical Architecture and Components
The hardware foundation relies on SkyFabriX, a custom scale-up switch chip based on the previously announced SkyHammer architecture. This silicon supports ESUN and standard Ethernet/IP protocols while accommodating evolving standards like UALink. SkyFabriX offers a switch capacity of 115.2 Tbps, with a roadmap targeting multiple petabits per second. Switch trays are available in both custom and standard rack form factors to pool GPUs and other accelerators into large scale-up domains.
Scale-out systems utilize Nvidia Spectrum-X silicon to handle traffic from 400G and 800G up to 1.6T speeds. These systems integrate into Nvidia and other accelerator clusters using SkyOS, the network operating system. SkyOS is based on SONiC and optimized for AI protocols. SkyCMD serves as the orchestration and observability layer, providing a single plane to manage both scale-up and scale-out fabrics simultaneously.
Company Background and Funding
Upscale, based in Santa Clara, California, emerged from stealth in September 2025 with a $100 million seed round. In June, the company raised an additional $190 million and outlined plans for its custom switch chip. Nvidia joined the June funding round and supplies the Spectrum-X silicon used in the scale-out systems. Upscale has now raised approximately $500 million and employs more than 300 people.
Barun Kar, CEO of Upscale, stated that compute is ramping up rapidly while networking has fallen behind. He emphasized the need to align networking capabilities with the speed of compute expansion. The Token Fabric announcement builds on these earlier milestones, positioning the company to address the growing disconnect between hardware performance and network infrastructure.
What should you do?
Organizations planning large-scale AI deployments should evaluate whether their current network architecture can handle unified scale-up and scale-out traffic. Review vendor roadmaps to determine if early-access programs align with your deployment timelines. Assess whether your team has the expertise to manage separate networks or if a unified stack reduces operational overhead. Monitor announcements for general availability dates to plan procurement accordingly.
What to do and how to stay safe: Upscale
- Audit current AI cluster networking to identify bottlenecks between intra-rack and cross-rack communications.
- Evaluate whether a unified software layer could simplify operations for your specific infrastructure size.
- Monitor vendor early-access programs to test integration with existing Nvidia or other accelerator clusters.
- Plan budget and procurement cycles around the early 2027 general availability target for full stack adoption.
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is Token Fabric?
Token Fabric is a hardware and software architecture by Upscale that unifies scale-up and scale-out networking for AI clusters using custom silicon and Nvidia Spectrum-X systems.
When will Token Fabric be generally available?
General availability for Token Fabric is planned for early 2027, with early-access and joint-validation programs currently running.
What components make up the Token Fabric architecture?
The architecture includes SkyFabriX scale-up silicon, Nvidia Spectrum-X based scale-out systems, SkyOS for network operations, and SkyCMD for orchestration and observability.


