TechnologyUnited States

Upscale Unveils Token Fabric with 115.2 Tbps SkyFabriX and NVIDIA Spectrum-X

Credited to Converge Digest · convergedigest.com7h ago

Useful

Upscale has introduced Token Fabric, a full-stack AI networking architecture combining its new 115.2 Tbps SkyFabriX scale-up switch silicon, NVIDIA Spectrum-X Ethernet-based scale-out systems, and a unified software and orchestration platform. The architecture is designed to connect heterogeneous GPUs and specialized AI accelerators within racks, across pods, and throughout large-scale AI data centers. The Santa Clara-based company is targeting hyperscalers, neoclouds, XPU developers, and enterprises seeking standards-based alternatives for building AI infrastructure.

At the center of the announcement is SkyFabriX, a new generation of scale-up switching silicon derived from Upscale’s previously announced SkyHammer architecture. The 115.2 Tbps platform supports Ethernet for Scale-Up Networking (ESUN), UALink over Ethernet (UALoE), SUE-T, and standard Ethernet/IP, with a roadmap encompassing UALink and optical interconnect technologies. Upscale plans to offer switch-tray systems in custom, Open Rack V3 (ORV3), and conventional rack form factors. For scale-out networking, Upscale is developing systems based on NVIDIA Spectrum-X switch silicon, supporting 400G and 800G Ethernet with a roadmap extending to 1.6T.

The software layer consists of SkyOS, an AI-native network operating system spanning both scale-up and scale-out fabrics, and SkyCMD, an orchestration and observability platform intended to provide centralized management across thousands of network elements. Customers will be able to purchase individual silicon components, integrated switching systems, or a complete networking stack. Upscale says early-access and joint-validation programs are underway, with general availability planned for early 2027.

TOKEN FABRIC | AI Networking Architecture

● 115.2 Tbps SkyFabriX: Purpose-built scale-up switch silicon based on the SkyHammer architecture. ● Open interconnects: ESUN, UALoE, SUE-T and Ethernet/IP, with evolving UALink support. ● NVIDIA Spectrum-X: Scale-out switching systems spanning 400G and 800G, with a 1.6T roadmap. ● SkyOS: Unified network operating system across scale-up and scale-out domains. ● SkyCMD: Agent-native orchestration, observability and network management. ● Optical roadmap: Pluggable optics, near-package optics (NPO) and co-packaged optics (CPO). ● Availability: Early access and validation underway; general availability targeted for early 2027.

Analysis: Upscale Moves from Scale-Up Silicon to a Complete AI Networking Stack

Token Fabric significantly expands Upscale’s strategy beyond the SkyHammer scale-up architecture introduced in 2025. The company is now positioning itself as a full-stack AI networking supplier, combining 115.2 Tbps SkyFabriX silicon for scale-up, NVIDIA Spectrum-X for Ethernet scale-out, and SkyOS/SkyCMD for common network operations and orchestration. The architecture is particularly aimed at heterogeneous AI systems in which GPUs and specialized XPUs need to operate within the same infrastructure. Support for ESUN, UALoE, SUE-T and Ethernet/IP gives Upscale several standards-based paths into this emerging market, while d-Matrix’s participation provides an early example of how a non-GPU accelerator could connect into the architecture.

Scale Up — SkyFabriX — 115.2 Tbps ESUN • UALoE • SUE-T • Ethernet/IP

Scale Out — NVIDIA Spectrum-X 400G • 800G • 1.6T roadmap

Software — SkyOS + SkyCMD Unified operations, observability and orchestration

The architecture also lets Upscale concentrate its proprietary silicon investment on the performance-sensitive scale-up domain while leveraging NVIDIA’s established Spectrum-X platform for scale-out Ethernet. TSMC’s involvement adds leading-edge manufacturing and advanced packaging expertise, while Upscale’s optical roadmap points toward pluggable, near-package and eventually co-packaged optics as bandwidth density increases. The key milestones now shift from architecture announcements to execution: SkyFabriX silicon validation, interoperability across heterogeneous accelerators, measurable latency and power results, and production deployments. Upscale says early-access and joint-validation programs are underway, with general availability targeted for early 2027.

Company Profile: Upscale

Leadership — Barun Kar, CEO and Co-Founder Rajiv Khemani, Executive Chairman and Co-Founder

Products — Token Fabric • SkyFabriX • SkyHammer • SkyOS • SkyCMD

Markets — AI Infrastructure • Hyperscalers • Neoclouds • Semiconductor Platforms • Enterprise AI

🌐 We’re tracking the latest developments in AI infrastructure and data center networking. Follow our ongoing coverage at https://convergedigest.com/category/ai-infra/

Original · Converge Digest

Front Method

Chips & Supply · United States

CIQ’s Fuzzball 4.3 Runs Validated Open-Weight Models on a Customer’s GPUs with One Command

Credited to HPCwire · hpcwire.com · 11h ago

Open models deploy from a catalog with validated presets, and in-cluster agents discover them automatically, across heterogeneous GPU fleets on premises and in major clouds.

RENO, Nev., Oct. 8, 2026 — CIQ, the enterprise software company behind Rocky Linux, today announced the general availability of Fuzzball 4.3, which brings ready-to-run AI models and agents to its sovereign AI orchestration platform. The release helps enterprises operationalize open-weight models and agentic AI on infrastructure they already run, with the control and governance required for AI in production. Fuzzball 4.3 deploys validated models from a catalog, connects agents to them automatically, and scales inference across heterogeneous GPU fleets, on premises and in the cloud.

Enterprises are shifting from AI pilots to production inference and repatriating workloads to private infrastructure for cost savings, security, and sovereignty. Meeting these goals often means standing up a new infrastructure stack before the first model serves a token, then integrating models, gateways, and agents by hand. Fuzzball 4.3 gives teams a faster path to productivity without rearchitecting their environment.

“Enterprises have already made their infrastructure decisions, and private AI should fit into them. With Fuzzball 4.3, a team starts a model, starts an agent, and the two find each other, on the GPUs, clusters, and clouds the organization already runs. That is what turnkey should mean,” said Gregory Kurtzer, CEO of CIQ and founder of Rocky Linux.

Models start with one command from the self-service model catalog. Eleven presets cover open-weight models from the GPT-OSS, Llama 4, Gemma 4, Qwen3-Coder, Mistral, and Nemotron 3 families, and a vLLM-based inference entry serves any other compatible Hugging Face model.

Models autoscale behind a single, stable endpoint and can scale down to zero running instances when idle, returning GPU capacity to other workloads, ready to spin back up for additional requests. And large models can scale across multiple nodes for additional resources.

Fuzzball 4.3 connects agents to models automatically. OpenCode and hermes-agent, two coding agents in the workflow catalog, run inside the cluster and discover running models through gateways, with no manual integration. Each model ships behind a built-in LiteLLM gateway with an OpenAI-compatible endpoint. Or, run one central gateway that fronts every model a user is authorized to reach. Another entry in the catalog provides a retrieval service to connect your private agents to enterprise documents.

Governance is built into the platform. Idle pools wake only for authenticated requests. Administrators control which organizations and groups can use which resources, and every workload runs unprivileged and rootless. Fuzzball 4.3 supports heterogeneous GPU fleets across NVIDIA and AMD for serving and health-aware scheduling, all under one operating model.

Fuzzball 4.3 is generally available now. Fuzzball deploys to AWS, Google Cloud, Oracle Cloud, CoreWeave, and Azure, as well as on-premises clusters built with Warewulf, VMware, or bare metal. To get started, contact CIQ at ciq.com/products/fuzzball.

CIQ is the founding support and services partner for Rocky Linux and a leading provider of Enterprise Linux infrastructure. CIQ delivers secure, performant software infrastructure for the full range of modern workloads. Global enterprises, government agencies, research institutions and supercomputing centers trust CIQ to run their most advanced workloads in production, from mission-critical environments to the systems behind major scientific and AI breakthroughs. CIQ’s products include the Rocky Linux from CIQ (RLC Pro) family of operating systems, Ascender Pro for IT automation, Fuzzball job-based container orchestration, Warewulf Pro cluster provisioning, and Apptainer, the container system built for high-performance computing, with Apptainer Pro coming soon. Learn more at ciq.com.