Upscale launches Token Fabric networking platform for AI data centers

Upscale says Token Fabric combines scale-up and scale-out networking with common software and operations.[1] Its stated goal is to connect any accelerator within a pod and across a data center, rather than limit the platform to one chip supplier.[1] Upscale says the network is intended to keep accelerators productive instead of waiting for data.[2][1]

Upscale says its SkyFabriX scale-up switch silicon delivers 115.2 terabits per second, with a roadmap to multiple petabits per second.[1] The company says SkyFabriX supports open standards including OCP ESUN and UALoE.[1] For scale-out, Upscale says its systems use NVIDIA Spectrum-X and are offered at 400G, 800G and 1.6T speeds.[1] Upscale describes SkyOS and SkyCMD as the platform’s software foundation and orchestration plane; it says SkyCMD provides orchestration and observability across both fabric types.[1]

Upscale says Token Fabric is intended for hyperscalers, XPU partners, neoclouds and enterprises, with customers able to choose its silicon, systems, software or full stack.[1] It positions the platform for neoclouds and enterprises without hyperscaler-size network teams, XPU partners seeking validated hardware, and hyperscalers designing their own systems and operations.[1] Reuters reported that the platform targets neoclouds and large cloud providers building data centers and seeking chips from multiple suppliers.[2]

Reuters reported that Upscale plans to release the first part of Token Fabric in the fourth quarter and roll out the full platform in stages through 2027.[2] Upscale separately says general availability is planned for early 2027, and that early-access and joint-validation programs are underway.[1] CEO Barun Kar told Reuters he expects Token Fabric revenue next year to reach tens of millions of dollars, potentially the low hundreds of millions.[2] Reuters also reported that Upscale was valued at $2 billion in June, when an extension led by Premji Invest brought its total funding to $500 million.[2]