Nvidia poached a founding member from Upscale AI, the startup developing a powerful line of switches for scale-up AI networking.

Rohit Mittal announced the news in a LinkedIn post. According to his profile, he is now at Nvidia working on compute solutions and AI fabrics, holding the position of “systems architect | technical director.”

“We are at a massive inflection point in the history of compute,” Mittal wrote. “As AI models scale, the rack is now the unit of compute and the fabric has become the critical backbone of the entire system. NVLink and NVLink racks are purpose-built for future AI infrastructure.

“Our mission at NVLink Fusion is to ensure our partners and customers can fully leverage this rack-scale technology. Looking forward to driving this next chapter of innovation.”

Mittal has a storied career in data center infrastructure, having previously served as a systems and silicon architect at Google Cloud, where he helped develop features found in the hyperscaler’s Trillium and Ironwood custom TPUs. He also previously led the founding team of an internal startup at Intel, building silicon photonics for data center networks.

Prior to joining Nvidia, Mittal was at Upscale AI, whose SkyHammer line of switches are designed to break bottlenecks in AI infrastructure by supporting high-speed, low-latency connections between GPUs, accelerators, memory, and storage.

SkyHammer was showcased last October, with the startup claiming it was capable of supporting the world’s largest AI clusters, buoyed by an interoperable design leveraging industry protocols like the UALink, Ultra Ethernet (UEC), and the Microsoft-developed Software for Open Networking in the Cloud (SONiC) open source platform.

Based on his LinkedIn post, Mittal looks set to work on NVLink Fusion, an interconnect technology designed to let partner hardware, including network interface controllers (NICs), data processing units (DPU), or scale-out switches, link with Nvidia hardware in rack-scale solutions.

The likes of MediaTek, Synopsys, Fujitsu, and Arm have already come aboard, with Nvidia breaking from traditions and allowing vetted partners to play with its toys so to speak, while also further denting traditional x86+GPU setups.

Shortly before Mittal’s departure, Upscale raised some $200 million from backers including the venture capital arms of Intel and Qualcomm, as well as Tiger Global, Xora Innovation, and Premji Invest.

The January raise added to a $100 million seed round secured last September. The same day as that round saw Mittal speak onstage at DCD Compute 2025, where he said the global AI market was being driven by tokenomics – where infrastructure operators vie to provide the best possible price for running AI models and services without impacting user experiences.

Mittal said that the U.S. market was more capitalist than others when it comes to token cost-modeling, while the European digital infrastructure market’s key differentiator was sustainability.

“Europe is beholden to whoever’s scaled out before,” Mittal said at the show. “It’s not strong much on data, but it is strong on low power. … Build your workforce for low power, and build the story from there.”