The AI scale-up switch space is inundated with big-name players, from Broadcom to Arista, Cisco to Marvell, and the unrelenting juggernaut that is Nvidia.

Upscale AI, though, is betting there’s still room for an AI‑only, open‑standards networking player with its SkyHammer switch line.

Having raised $100 million last September and a further $200 million in January, the startup is gearing up for a product launch later this year. Speaking to SDxCentral on the sidelines of GTC, Deepti Chandra, Upscale’s VP for product and marketing, said the startup was looking at some time “close to [the] second half” of the year.

The Intel- and Qualcomm-backed firm didn’t appear out of nowhere. Founders Rajiv Khemani and Barun Kar previously launched Auradine, which sells servers specially designed for bitcoin mining.

Using their expertise in hardware and computing, they came together to tackle one of the biggest bottlenecks in AI infrastructure: ensuring high-speed, low-latency connections across GPUs, accelerators, memory, and storage.

Since emerging from stealth, Chandra told SDxCentral the firm has now grown to around 200 staff as it starts to “build to scale internally.”

That team’s initial efforts focused on AI for scale-up, or interconnectivity within the rack. However, just days before GTC, Upscale revealed a collaboration with one of the market leaders it aims to challenge: Nvidia.

Having joined Nvidia’s Partner Network, Upscale announced plans to offer scale-out (cabinet to cabinet) Ethernet systems built on the tech giant’s Spectrum-X switch silicon, paired with the startup’s AI-optimized software based on the software for open networking in the cloud (SONiC) standard.

Chandra said the shift to scale-out was about “diversifying” the startup’s portfolio, adding: “We are executing strongly on scale-up. What we are doing is enhancing the strength of the ecosystem.

“Nvidia is a great partner because of everything that they bring to the ecosystem, from a compute standpoint, from all the different solutions that they have built with storage, etc. It just made sense to unify those synergies and diversify the scale-out portfolio as well," Chandra said.

Just a month before the partnership was struck, one of Upscale’s senior product leaders departed for Nvidia. Rohit Mittal joined the tech giant to work on compute solutions and AI fabrics.

Despite his exit, though, Chandra emphasized that the architecture team that was “focusing and executing” remains unchanged, adding that the AI networking world is small enough that they’ll likely work together again.

It’s now full speed ahead for the startup, with its Intel- and Qualcomm-backed efforts framed by the VP as part of a broader push toward heterogeneous XPU environments, stressing that Upscale’s switches are built on open standards to support multiple accelerator vendors rather than locking into a single ecosystem.

And since its hardware is built to comply with emerging standards around AI fabrics, including the Ethernet for Scale-Up Networking (ESUN) consortium and the SONiC work group for scale‑up networking, she argued that, unlike traditional networking vendors, Upscale doesn’t carry “baggage” from service provider or campus markets and can design its silicon and software purely around AI workloads.

“We are liberated and we have this advantage of pure play where we can build the architecture from the ground up so it's not taking silicon, it's not taking software, it's not taking an architecture that was built for something else in mind and then retrofitted to the scale-up problem,” Chandra said.

“The scale-up in itself is a very nuanced problem with memory semantics and the type synchronization that you need between GPUs," Chandra added. "There is a lot that goes on under the hood to make it work in a very synchronous manner. That is a prong that we are thinking end-to-end, and it's going on strong.”