What happened
Upscale introduced Fabric, a full‑stack networking solution designed for AI factories. The platform combines a new scale‑up switch silicon called SkyFabriX with custom switch‑tray systems, and scale‑out switches powered by NVIDIA Spectrum‑X. A single software stack—SkyOS and the orchestration tool SkyCMD—provides unified management across GPUs and other XPUs. Upscale says the architecture can connect any accelerator, scale accelerators inside a rack pod, and extend scaling across an entire data center. Customers may adopt the solution as silicon only, as pre‑built systems, as software‑only, or as a full stack, with flexible consumption models for hyperscalers, XPU partners, neoclouds, and enterprises. Early‑access and joint‑validation programs are already running, and the company will showcase the technology at the OCP 2026 conference. General availability is planned for early 2027.
Upscale’s Fabric merges two previously separate networking approaches. The scale‑up side relies on SkyFabriX, a silicon switch built on the earlier SkyHammer architecture, supporting protocols such as ESUN, UALoE, and SUE‑T, and offering deterministic ultra‑low latency and memory‑semantic communication. The scale‑out side uses NVIDIA Spectrum‑X‑based switches ranging from 400 Gb/s to 1.6 Tb/s, with future Spectrum‑X generations planned.
The software layer, SkyOS, is described as an AI‑native network operating system that provides a single control plane across both scale‑up and scale‑out fabrics. SkyCMD adds agent‑native orchestration, allowing thousands of network elements to be observed, reasoned about, and acted upon as a unified system.
Upscale emphasizes flexible consumption: customers can purchase just the silicon, the pre‑built switch trays, the software stack, or a fully integrated rack solution. Early‑access programs are already in place, and the company will present a live demo at the Open Project (OCP) 2026 exhibition.
Why it matters
The announcement targets a market that analysts forecast will reach nearly $1 trillion in AI back‑end switch spending by 2030, with AI networking alone projected to exceed $200 billion. Current networking offerings are either discrete components that require complex integration or legacy systems not optimized for AI workloads. By delivering an open, standards‑based fabric that unifies scale‑up and scale‑out networking, Upscale aims to simplify deployment, improve utilization, and lower the cost per of AI . If adopted widely, the platform could accelerate the rollout of large‑scale AI clusters, reduce engineering overhead for hyperscalers and enterprises, and set a new baseline for heterogeneous accelerator connectivity.
Analyst forecasts from Dell’Oro and 650 Group suggest AI networking spend will dominate data‑center budgets in the next five years. A unified fabric could capture a sizable share of that spend by reducing the need for multiple vendor solutions and simplifying operations.
Heterogeneous AI workloads—mixing GPUs, TPUs, and emerging XPUs—require low‑latency, high‑bandwidth interconnects that current Ethernet‑centric designs struggle to provide. Fabric’s memory‑semantic communication promises to meet these demands.
By offering an open, standards‑based stack, Upscale positions itself to attract a broader ecosystem of hardware partners, potentially accelerating innovation beyond the company’s own silicon.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').A route planner searches possible journeys using explicit rules. What does this illustrate about AI?
What to watch next
Key indicators will include the uptake of early‑access programs, the breadth of XPU partner integrations, and customer announcements from hyperscalers or large cloud providers. Competition from established networking vendors and emerging open‑fabric initiatives will shape market dynamics. The performance of SkyFabriX silicon and its roadmap to multi‑petabit‑per‑second bandwidth will be closely watched, as will the evolution of SkyOS and SkyCMD in handling increasingly complex AI workloads.
Adoption rates among hyperscalers and large enterprises, especially any public statements about deploying Fabric in production.
Performance benchmarks released by Upscale or third‑party labs that validate the claimed multi‑petabit‑per‑second roadmap.
Competitive responses from networking giants such as Cisco, Arista, or Mellanox, and any standard‑body activity around AI‑specific interconnect protocols.