The largest AI data centers are built around one supplier’s chips and one supplier’s way of wiring them together. A Santa Clara startup carrying Nvidia’s money is betting that arrangement is about to loosen.
Upscale AI introduced Token Fabric on Thursday, a networking platform designed to let AI chips from different vendors share a single data center. The launch, first reported by Reuters, takes aim at the default of modern AI infrastructure: a customer who buys Nvidia’s processors usually buys Nvidia’s networking as well, and the two are engineered to fit only each other.
AI clusters depend on two kinds of networking. Scale-up links accelerators inside a single rack, moving data between chips that must stay tightly synchronized. Scale-out links racks across a building. A single training job spans both, and most operators treat them as separate networks bought from separate vendors. Token Fabric’s pitch is to bring the two under one roof.
The twist is that Nvidia is among Upscale’s backers. Nvidia joined the startup’s June financing round, which added $190 million and valued the company at $2 billion. Salesforce Ventures and Singapore state investor Temasek also took part, and Premji Invest, the investment arm of Indian billionaire Azim Premji, led the round. Total funding now sits near $500 million.
Token Fabric is a stack rather than a single product. It pairs Upscale’s own SkyFabriX switch silicon with scale-out switches built on Nvidia’s Spectrum-X Ethernet technology, all run by two software layers called SkyOS and SkyCMD. The software identifies bottlenecks and flags hardware that may need replacing, so that expensive accelerators spend their time computing instead of waiting for data.
The hardware is substantial. SkyFabriX carries a switch capacity of 115.2 terabits per second, with a roadmap toward multiple petabits. The scale-out side runs on Spectrum-X at speeds from 400 gigabits up to 1.6 terabits. SkyOS is built on SONiC, the open network operating system that grew out of cloud computing, and SkyCMD sits on top as the orchestration and monitoring layer.
The architecture rests on open standards. Upscale has committed to UALink, a protocol that lets accelerators read one another’s memory as if it were local, and to Ethernet for Scale-Up Networking, an AI-tuned variant of the familiar standard. The company also contributes to the Ultra Ethernet Consortium. The bet is that customers, especially neoclouds and large cloud providers assembling multi-vendor data centers, will pay to escape a single supplier’s grip.
The pace of the company’s rise shows the size of that bet. Upscale was founded only in 2025 by chief executive Barun Kar and executive chairman Rajiv Khemani, both veterans of Palo Alto Networks and Juniper Networks, and incubated inside Auradine, the chip firm Khemani led. The pair had earlier worked at Innovium and Cavium, two names in the switch-chip trade. The company has grown to roughly 200 employees and reached its $2 billion valuation in under 18 months.
Gilad Shainer, Nvidia’s senior vice president of networking, lent his name to the launch. “The network is the backbone of the AI factory, keeping accelerators connected and productive as systems scale,” he said in the company’s statement. Nvidia’s presence cuts both ways: it gives the project credibility, and it raises the question of how far a startup can challenge the status quo when the status quo is also its backer.
The timing favors the argument. Google, Amazon and Microsoft have each built their own accelerator silicon, which means the largest buyers now run several kinds of chips at once, and no single vendor’s proprietary link covers all of them. A fabric that works across vendors speaks directly to that mixed reality.
The first piece of Token Fabric, linking groups of AI chips inside a data center, arrives in the fourth quarter, with the full platform rolling out in stages through 2027. General availability is planned for early 2027. That schedule is measured for a company that has raised money three times in about a year, and it reflects how hard the underlying problem is.
The problem is real and easy to state. In a large AI system, work is spread across thousands of chips that must constantly exchange data, and when the network is slow those chips sit idle while their cost still counts. Upscale’s pitch is that a single fabric, instead of separate scale-up and scale-out networks from separate vendors, keeps the most expensive part of the data center busy.
Whether customers will abandon a working single-vendor stack for an open one is the open question. Nvidia’s own NVLink keeps its chips tightly coupled, and much of the industry has accepted that lock-in as the price of performance. Upscale is wagering that once the network is fast enough, performance will stop being the deciding factor and cost will start. The company says it aims to generate tens of millions of dollars in revenue next year. For now, that is an ambition on paper, not a market that exists.
