AI News · New apps ·
Upscale AI launches Token Fabric to connect chips from different suppliers
Nvidia-backed Upscale AI launched Token Fabric, a networking platform for data centers using AI chips from multiple suppliers. It combines hardware and software to connect processors without requiring customers to integrate separate networking systems. Upscale says the platform uses open standards and supports graphics chips, custom AI chips and network cards from any vendor.
Key points
- Token Fabric combines hardware and software to connect AI chips from different suppliers.
- Upscale says the platform uses open standards and supports chips and network cards from any vendor.
- The platform combines Upscale’s networking chip, Nvidia Spectrum-X technology and network monitoring software.
- Initial delivery is planned for the fourth quarter of 2026, with staged rollout continuing through 2027.
- Pricing and independent performance results were not reported.
What happened: Upscale AI, a Santa Clara networking startup backed by Nvidia, launched Token Fabric on October 8, 2026, Reuters reported. The platform combines hardware and software to connect AI processors from different suppliers across a data center, without requiring customers to integrate separate networking systems. Upscale says it uses open standards and supports graphics chips, custom AI chips and network cards from any vendor. That could give infrastructure buyers more flexibility when choosing processors, although compatibility and performance with specific hardware combinations were not reported.
The details: Token Fabric combines SkyFabriX, Upscale’s own networking switch chip, with switches built on Nvidia’s Spectrum-X Ethernet technology, The Next Web reported. Two software layers, SkyOS and SkyCMD, run and monitor the network. According to the company, the software identifies bottlenecks and flags equipment that may need replacement. The platform covers connections within a server rack, known as scale-up networking, and connections between racks across a data center, known as scale-out networking. Its aim is to keep expensive AI chips working rather than waiting for data.
Background: Large AI systems spread work across thousands of chips that must continually exchange data, The Next Web explained. Slow connections can leave processors idle, making networking an important part of getting value from AI infrastructure. Many large systems use a single vendor’s connections, such as Nvidia’s NVLink. Upscale instead supports shared standards including UALink and Ethernet for Scale-Up Networking, or ESUN, to let rival manufacturers’ chips share a network. Nvidia is both an investor in Upscale and a technology supplier for Token Fabric, even as the platform seeks to support processors from competing companies.
Who it affects: Upscale is targeting large cloud providers and neoclouds, smaller companies that rent out AI computing power. Reuters reported that the startup is betting on demand from customers expanding data centers and seeking chips from multiple suppliers. Upscale was valued at $2 billion in June, when a $190 million funding extension brought its total financing to about $500 million. Chief executive Barun Kar told Reuters he expects Token Fabric revenue in 2027 to reach tens of millions of dollars, potentially the low hundreds of millions. Those figures are expectations, not reported sales.
What to watch: Reuters reported that the first component, connecting groups of AI chips across a data center, is planned for the fourth quarter of 2026, with the full platform arriving in stages through 2027. The Next Web reported that Upscale’s announcement puts general availability in early 2027 and says early-access programs are under way. Buyers should distinguish initial availability from delivery of the complete platform and validate interoperability and performance with their intended hardware mix. Pricing and independent performance results were not reported.
Our take
Cross-vendor networking could give infrastructure buyers more flexibility. Teams should validate interoperability and performance with their actual hardware mix before committing.