An Nvidia-funded startup has just unveiled a network designed to let chips from rival manufacturers live in the same cluster — exactly the kind of mix that NVLink, Nvidia's proprietary interconnect, makes difficult.
An AI cluster has two networks. The scale-up network connects the GPUs inside a rack so they act as one giant accelerator; at Nvidia, that's NVLink, a proprietary link. The scale-out network connects racks to each other, usually over Ethernet (at Nvidia, Spectrum-X). Mixing accelerators from different manufacturers on the same network is hard because each vendor brings its own interconnect fabric.
Upscale AI, an Nvidia-backed startup, introduced Token Fabric, a networking platform that aims to tie together chips from rival vendors in the same data center. According to The Next Web and Network World, it combines:
The idea is to extend existing protocols rather than replace them: scale-up traffic uses ESUN and scale-out uses standard Ethernet with RoCE, with a single control plane across both layers. Its CEO, Barun Kar, proposes measuring the network with AI metrics — time to first token, tokens per second, tokens per dollar and tokens per watt — instead of packet metrics.
Upscale was founded in 2025 by Barun Kar and Rajiv Khemani, has raised close to US$500 million across three rounds and was valued at US$2 billion in June 2026, with investors including Nvidia, Salesforce Ventures and Temasek. The first part of Token Fabric arrives in the fourth quarter of 2026, with the full platform rolling out in stages through early 2027. Kar expects revenue in the tens of millions of dollars in 2027, possibly up to the low hundreds of millions.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel