Upscale aims to challenge Nvidia's dominance in AI networking infrastructure with an open solution that builds on SONiC and other standards-based initiatives.
AI workloads have fundamentally changed networking requirements, creating an opening that Upscale AI aims to exploit.
The startup emerged from stealth this week with $100 million in funding to challenge Nvidia’s dominance in AI networking infrastructure. The seed round was co-led by Mayfield and Maverick Silicon, with participation from StepStone Group, Celesta Capital, Xora, Qualcomm Ventures, Cota Capital, MVP Ventures and Stanford University. Upscale AI was incubated by Auradine, a blockchain and AI compute infrastructure company. The company is led by networking veterans from Palo Alto Networks and Innovium.
“You need a network, which not only has very low latency, it has predictable latency, it has high bandwidth needs and is a memory semantic load-store network,” Barun Kar, CEO explained to Network World. “Nvidia has shown us the way with NVLink, but it’s kind of a closed standard, right? So what we are trying to do here is a completely open, standards-based solution in this space.”
A new inflection point in networking
The genesis of Upscale can be traced to a shift in the demands placed on data center networks.
“Networking has gone through several inflection points,” Kar said. “We feel there is an inflection point happening right now as we speak, driven by AI. The networks we need today are no longer just CPU-to-CPU or CPU-to-storage—they’re GPU-to-GPU, with high bandwidth and ultra-low, predictable latency.”
The company was incubated within Kar’s previous company, Auradine, which is building out AI and blockchain technologies. As his team spoke with customers, they identified a clear need for a new, ground-up solution to better tackle the challenges that AI brings to networking. As customer interest and engineering work progressed, the effort also attracted investor interest, and the decision was made to spin the effort out as Upscale.
“What we are trying to do here is a completely open, standard-based solution, so that any compute vendor, any accelerator, or even hyperscalers, can bring their own compute and seamlessly plug into the rack,” Kar said.
How Upscale is building on an open standards foundation
Upscale AI’s technical approach leverages multiple open standards initiatives rather than developing proprietary protocols.
The company’s stack builds on the SONiC (Software for Open Networking in the Cloud) network operating system, Ultra Ethernet Consortium (UEC) specifications, Ultra Accelerator Link (UALink) standards, and the Switch Abstraction Interface (SAI).
UEC specifications specifically target AI networking challenges by adding congestion management, advanced telemetry and predictable latency features that traditional Ethernet lacks. UALink provides standardized interfaces for accelerator interconnects, breaking dependence on proprietary solutions like Nvidia’s NVLink, while SAI provides hardware abstraction.
Kar explained that those standards will provide the foundation for what Upscale is building, but he’s not stopping there. “We are upgrading the stack for both SAI and SONiC for scale up,” Kar noted.
Full-stack integration strategy
While some networking vendors focus on a specific part of the stack, the goal for Upscale AI is to cover everything. “We are fully vertically integrated,” Kar said. “We do silicon, system, software, everything.”
The company is developing custom ASICs for scale-up applications while partnering with existing manufacturers for scale-out and management functions. This approach includes thermal management expertise from cryptocurrency mining experience.
Building all that technology will take time and money, which is what the $100 million will be used to support. The company plans to first develop scale-out products using merchant silicon and out-of-band management, followed by scale-up solutions with custom ASICs and then complete racks.
“What we are trying to do is build an open-standard networking technology to basically democratize,” Kar said. “Democratization means faster innovation with more choices essentially and flexibility.”
Upscale AI at a glance:
- Founded: 2024 (incubated by Auradine)
- Funding: $100 million
- Investors: Mayfield, Maverick Silicon, StepStone Group, Celesta Capital, Xora, Qualcomm Ventures, Cota Capital, MVP Ventures and Stanford University
- Headquarters: Palo Alto, Calif.
- CEO: Barun Kar
- What they do: Develop high-performance AI networking infrastructure using open standards.
