Nvidia-backed Upscale unveils fabric that lets engineers break vendor lock-in and mix rival XPUs

Nvidia-backed Upscale AI unveiled its Token Fabric, a full-stack networking portfolio that connects heterogeneous XPUs from different suppliers both inside the rack (scale-up) and across cabinets (scale-out), the company said. The offering combines Upscale's SkyFabriX switch silicon, which supports ESUN, UALink and SUE-T standards, with Nvidia Spectrum-X silicon for scale-out, tied together by the SkyOS network operating system and SkyCMD orchestration plane. Upscale CEO Barun Kar said the fabric "brings heterogeneous accelerators, scale-up, scale-out and software together as one coherent AI system, built on open standards," positioning it as a way for AI infrastructure engineers to avoid vendor lock-in.

Upscale AI, the Nvidia-backed startup developing network switch silicon, has revealed a full-stack portfolio allowing partners to build heterogeneous compute stacks. The Upscale Token Fabric for AI factories let users connect hardware from different suppliers both inside the rack scale-up and across cabinets scale-out . Arguing that rival networking vendors only offer components customers must integrate themselves, the Upscale team claims to have built a system that intertwines heterogeneous XPUs across networks specifically for AI. “The industry has spent decades connecting computers. AI demands something different: one computer, built at data-center scale,” said Upscale CEO Barun Kar. “Upscale Token Fabric brings heterogeneous accelerators, scale-up, scale-out and software together as one coherent AI system, built on open standards, so the network never gets in the way. Operators focus on delivering the best AI compute for their workloads; the integrated stack delivers the performance and time to market.” On the scale-up side, the SkyFabriX switch silicon is based on the startup’s initial SkyHammer https://www.sdxcentral.com/news/upscale-ai-bets-on-pureplay-openstandards-switches-for-ai-fabrics/ architecture. In a bid to be as open as possible, Upscale opted for SkyFabriX to support as many standards as possible, including Ethernet for Scale-Up Networking ESUN https://www.sdxcentral.com/news/microsoft-and-meta-drive-new-ethernet-standards-for-ai-era/ , Ultra Accelerator Link UALink https://www.sdxcentral.com/news/ualink-consortium-takes-another-swing-at-nvidias-nvlink-supremacy-with-specification-20/ , and Scale-up Ethernet Transport SUE-T https://www.opencompute.org/projects/networking/ . Deepti Chandra, Upscale’s VP for product and marketing, explained to SDxCentral https://www.sdxcentral.com/news/upscale-ai-bets-on-pureplay-openstandards-switches-for-ai-fabrics/ earlier this year that its decision to be fully open stems from its heterogeneous aspirations, wanting to ensure support across multiple accelerator vendors rather than locking into a single ecosystem. By going as open as possible and supporting a multitude of evolving standards, Upscale argues it would help allow SkyFabriX to provide low-latency connectivity between any GPU or XPU over one fabric. On the scale-out side, Upscale leans on its partnership with Nvidia, building its offering atop Spectrum-X silicon. The special sauce ensuring interconnectivity between disparate designs is software. Specifically, the network operating system SkyOS, which provides a control plane between clusters – in this case, providing seamless connections for Nvidia and XPU clusters. SkyCMD, meanwhile, acts as an orchestration plane across both scale-up and scale-out domains to monitor network operations. “The network is the backbone of the AI factory, keeping accelerators connected and productive as systems scale,” said Nvidia’s networking chief, Gilad Shainer. “By using Spectrum-X Ethernet for Token Fabric’s scale-out systems, Upscale is helping customers build high-performance, open and heterogeneous AI infrastructure.” Upscale’s Token Fabric builds on something Nvidia has been working on for some time in the form of NVLink Fusion https://www.sdxcentral.com/news/nvidia-opens-nvlink-fabric-to-hyperscaler-custom-silicon/ , which finally allows rival XPUs and CPUs to be used in Nvidia rack-scale platforms. The startup’s platform goes one step further, giving AI infrastructure engineers the means to build what they want to, for their workload needs without vendor lock-in, and network it all together. “As AI infrastructure becomes more complex and expands beyond the hyperscalers, there’s a growing need for integrated, full-stack architectures that deliver the required performance and time to market,” Sameh Boujelbene, VP at analyst firm Dell’Oro Group. “Upscale Token Fabric is addressing this demand, bringing scale-up and scale-out networking together in a unified architecture for heterogeneous AI infrastructure.” General availability for Token Fabric is planned for early 2027, with early-access and joint-validation programs already underway.