Upscale AI has launched a networking platform called Token Fabric that aims to let customers build data centres using artificial-intelligence chips from different suppliers without stitching together separate networking systems for each one.
The startup, valued at about 2 billion dollars in June, combines its own networking chips, equipment using Nvidia’s Spectrum-X technology and software that identifies bottlenecks and flags equipment that may need replacing. The aim is simple to state and hard to achieve: keep expensive AI processors working instead of waiting for data to move between them.
Modern AI training is as much a data-movement problem as a computing problem. Thousands of processors must exchange results constantly. If the network slows, costly chips sit idle while still consuming power, cooling and capital. Upscale says operators should judge a network by useful output, such as tokens produced per dollar and per watt, rather than by headline port speed alone.
The company’s bet is that customers want to avoid dependence on a single chip supplier. Large cloud providers, newer specialist clouds known as neoclouds, accelerator makers and big enterprises are all building AI capacity. Some may want a complete system; others may want only silicon, switch equipment or software. Upscale says it can sell the full stack or individual pieces.
Chief Executive Barun Kar told reporters he expects Token Fabric revenue next year in the tens of millions of dollars, potentially reaching the low hundreds of millions. Those are company expectations, not guaranteed results. They do, however, show how quickly the firm believes demand for multi-vendor networking will grow.
Upscale has substantial backing. In June it raised a 190 million dollar extension round, bringing total funding to about 500 million dollars. Premji Invest led the round, with participation from Nvidia, Salesforce Ventures and Singapore’s Temasek, among others. Nvidia’s involvement is notable: a company famous for its own tightly integrated AI systems is also an investor in a firm selling interoperability.
The first part of Token Fabric, which connects groups of AI chips across a data centre, is planned for release in the fourth quarter of 2026. The full platform is expected to roll out in stages through 2027. That timetable leaves time for competitors, including established networking vendors and the chip companies themselves, to respond.
The launch captures a wider shift in the AI industry. The first race was to obtain enough processors. The next race is to connect them efficiently enough that the investment pays. If Upscale can make rival chips work as one system, it could claim a valuable position between the companies that design accelerators and the customers who operate them.
Upscale’s challenge will be execution across many partners. A network that promises to connect rival chips must prove itself in real data centres, where a failure can idle millions of dollars of equipment. Customers will test whether Token Fabric keeps processors busy under full training loads, not in demonstrations. Larger networking companies are unlikely to concede the role of neutral connector, and chip makers themselves sell competing systems. Still, the direction of travel is clear: as AI clusters grow, whoever controls the connections influences the whole market. Upscale is asking to be that Switzerland of AI hardware, armed and well funded.
Related reading: SpaceX Spectrum Deal Aims to Turn Starlink Into a Full Mobile Carrier · Apple Reportedly Plans Touchscreen OLED MacBook Pro Launch on October 27 · Microsoft Unveils New Surface Laptop, Windows Updates at San Francisco Showcase
