The Open Networking Stack for the AI Era
Our investment in Upscale AI.
June 22, 2026
- Founders: Barun Kar, Rajiv Khemani
- Sector: AI Infrastructure
- Location: Santa Clara, CA
The AI revolution has placed unprecedented demand on data centers, and networking is where that demand is most visible. A server rack used to draw 30 kilowatts of electricity. Today, it draws a megawatt. A training run used to fit inside a server and now it spans thousands of accelerators that have to behave like a single machine. The bottleneck for AI acceleration is increasingly becoming the fabric that connects compute. That’s why we’re excited to partner with Upscale AI as they build the open-standard networking infrastructure for the AI era.
We’ve invested at the compute layer before — Together AI in 2023 and Crusoe in 2025. What became clear through those investments is that the rest of the data center stack, networking included, was evolving just as fast.
Our investment in Upscale AI reflects a structural shift in how AI compute gets built. For the last few years, buying AI infrastructure meant buying the entire stack from one vendor — GPU, switch, and software, bundled together. Now every major hyperscaler is designing its own accelerator — Google’s TPU, Amazon’s Trainium, AMD’s MI-series, Meta’s MTIA, Microsoft’s Maia, as well as custom ASICs focused on specific workloads. Each of them needs a high-bandwidth, low-latency fabric to stitch hundreds of chips into a single machine. That fabric is the next great open battleground in AI infrastructure.
AI networking is a large and growing market — on track to potentially exceed $200B by 2030. We believe the deepest, most defensible value sits in the switching ASIC: the costliest, hardest-to-replicate part of the box, and exactly where Upscale chose to start. By owning the silicon and building up through systems and software, Upscale can capture far more of the stack than the box-builders a layer above.
Upscale AI is the first company to build a purpose-built scale-up switching ASIC that natively supports multiple open protocols — the networking layer that connects AI accelerators together inside a single server rack. Their scale-up architecture ensures interoperability across open and diverse environments while maintaining performance. In addition, Upscale AI recently announced a collaboration with NVIDIA to enable heterogenous scale-out AI clusters with NVIDIA Ethernet switch silicon, streamlining AI infrastructure at scale.
Upscale AI was founded by Barun Kar and Rajiv Khemani — among the most accomplished operators in the history of networking silicon. Rajiv was COO of Cavium (helped take the company public before being acquired by Marvell), and co-founded Innovium (acquired by Marvell) and Auradine (now called Velaura AI), where Upscale AI was incubated. Barun was employee #10 at Palo Alto Networks, where he helped deliver three generations of custom network and security ASICs. Before that, he was a key contributor behind Juniper’s MX Series router.
We believe Upscale AI is building one of the most important companies in the AI infrastructure stack at the layer that will increasingly determine how and how fast AI scales. We’re proud to back Barun, Rajiv, and the entire team as they build the open networking standard for the AI era.
