Upscale AI on October 8 launched Token Fabric, a networking platform that links AI accelerators from different suppliers inside the same data centre, with the first shipments due in Q4 2026 and full general availability extending through 2027.

The Santa Clara company pairs its own SkyFabriX scale-up switch with scale-out switches built on Nvidia's Spectrum-X Ethernet chips. The product arrives as AI clusters expand and bandwidth between accelerators becomes as decisive as raw FLOPS.

What does Token Fabric actually do?

Token Fabric connects GPUs, custom AI silicon, and network interface cards from any vendor. It supports open standards including UALink and Ethernet for Scale-Up Networking (ESUN), sidestepping Nvidia's proprietary NVLink as the default scale-up glue.

The hardware is only half the story. Two software layers — SkyOS and SkyCMD — manage the network, spot congestion, and flag hardware due for replacement. The platform separates scale-up links (inside one rack) from scale-out links (across a building) but routes them through a common control plane.

"The network is the backbone of the AI factory, keeping accelerators connected and productive as systems scale," Gilad Shainer, Nvidia's senior vice president of networking, said in the company's announcement.

Who is behind Upscale, and what is the money?

Co-founders Barun Kar (CEO) and Rajiv Khemani (executive chairman) started Upscale in 2025. Funding has stacked up quickly:

• Seed round: more than $100m in September 2025
• Series A: $200m in January 2026
• Series A extension: $190m in June 2026, led by Premji Invest at a $2bn valuation, with Nvidia, Salesforce Ventures, and Temasek joining
• Total raised: roughly $500m

Kar told Reuters he expects Token Fabric revenue next year in the tens of millions of dollars, possibly reaching the low hundreds of millions. The sales target is large cloud providers and neoclouds — smaller firms that rent AI compute by the hour.

How does this fit the broader market?

Nvidia had built a roughly $99bn equity portfolio across AI startups by late July 2026, and Upscale is one of those bets. The strategic logic is straightforward: an open-standards networking layer keeps Nvidia's Ethernet silicon in the scale-out slot even when customers buy non-Nvidia accelerators for scale-up.

Competitors are circling. In July, chip-interconnect firm Eliyan reached a $1bn valuation on a $145m Series C, going after the same bandwidth bottleneck between dies and packages. The race is now to standardise the wiring behind every AI cluster before one vendor's proprietary fabric locks it in.

What changes for buyers?

Data-centre operators get a multi-vendor option at the rack level without abandoning the Ethernet gear they already operate. For Nvidia, the deal preserves a presence in any Token Fabric deployment even if the AI silicon inside is not its own. For AMD, Intel, and the wave of custom-ASIC builders, Token Fabric offers a path to large clusters without negotiating separate interconnect licences with each system builder.

Upscale's revenue guidance is still a fraction of Nvidia's networking business, and the platform has to clear interoperability testing with non-Nvidia silicon before cloud buyers sign at scale. Whether open-standards scale-up can match NVLink on latency and bandwidth per dollar will become clearer once the first Q4 deployments come under live workload.