Nvidia-backed networking startup Upscale AI has launched a new platform designed to connect AI processors from different suppliers inside the same data center, targeting one of the infrastructure problems emerging as cloud operators move beyond single-vendor GPU clusters.
Reuters reported that Upscale introduced Token Fabric on October 8, a hardware-and-software platform that lets customers build data centers using AI chips from multiple suppliers without having to integrate separate networking systems themselves.
The launch gives the $2 billion startup a larger role in the competition over the networking layer connecting increasingly diverse AI accelerators.
Token Fabric connects scale-up and scale-out networking
Modern AI systems can distribute workloads across thousands of processors, making the network connecting those chips increasingly important to overall system performance.
Upscale says Token Fabric combines scale-up, scale-out and AI-native software under one architecture. Scale-up networking connects accelerators within tightly coupled computing systems, while scale-out links larger numbers of systems across a data center.
The company’s SkyFabriX switch silicon is designed to deliver 115.2 terabits per second, with a roadmap extending into multi-petabit-per-second performance. Its scale-out systems use Nvidia Spectrum-X Ethernet, while SkyOS and SkyCMD provide the software and orchestration layers.
Crucially, Upscale says the architecture is built around any GPU, any XPU, any NIC rather than requiring customers to standardize their infrastructure around one accelerator supplier.
Nvidia is both investor and infrastructure partner
The company’s relationship with Nvidia makes that strategy particularly interesting.
Upscale raised a $190 million Series A extension in June, led by Premji Invest, with Nvidia, Salesforce Ventures and Temasek participating. The round brought its total funding to roughly $500 million and valued the company at about $2 billion, according to The Next Web.
CEO Barun Kar told Reuters that he expects Token Fabric to generate revenue in the tens of millions of dollars next year, with the possibility of reaching the low hundreds of millions. The company is targeting hyperscalers and neocloud providers building increasingly heterogeneous infrastructure.
Kar co-founded Upscale with executive chairman Rajiv Khemani in 2025.
AI infrastructure is moving toward mixed-vendor systems
The broader opportunity comes from the growing number of alternatives to Nvidia GPUs.
Cloud operators are increasingly evaluating custom accelerators and chips from competing vendors, but mixing processors can create networking and orchestration complexity. Upscale is betting that customers will want an infrastructure layer capable of hiding some of that fragmentation.
The company says general availability for Token Fabric is planned for early 2027, with early-access and joint-validation programs already underway.
That makes networking an increasingly strategic part of the AI stack.
The first wave of the AI infrastructure race focused heavily on securing GPUs. The next may depend on whether operators can efficiently connect different kinds of processors into one functioning compute system.
For builders and cloud providers, Upscale’s proposition is straightforward: customers should be able to choose the best accelerator for each workload without rebuilding the network every time the chip vendor changes.
