Upscale AI raises $190M extension at $2B valuation; NVIDIA backs AI-native networking switches

Upscale AI, a Santa Clara-based networking infrastructure startup, raised $190 million in Series A-1 extension funding at a $2 billion post-money valuation. The round was led by Premji Invest (India-based VC), with new investors NVIDIA, Salesforce Ventures, Seligman Ventures, and Temasek joining existing backers Maverick Silicon, Mayfield, Prosperity7 Ventures, StepStone Group, and Tiger Global. Total funding now stands at $500 million. Upscale develops silicon, systems, and software for AI data center networking, building AI-optimized scale-up and scale-out switches designed around NVIDIA's Spectrum-X silicon.

AI cluster workloads expose a structural bottleneck: the network connecting thousands of GPUs must coordinate model training and inference across synchronized, high-throughput pipelines. A single latency spike cascades, delaying all downstream computations. Upscale is betting that AI-native networking—built for transformer compute patterns rather than generic data center traffic—can unlock 10-50% efficiency gains. The company uses open standards (ESUN, SONiC) to avoid vendor lock-in, positioning against NVIDIA's proprietary switches, Broadcom, Arista, and Cisco's enterprise networking stack.

For infrastructure architects, this round validates networking as its own category in the AI compute stack. Global AI data center switch spending is projected to exceed $100 billion annually by 2030. Upscale's $2B valuation and NVIDIA's participation signal confidence that the networking layer, long a forgotten abstraction, is now a first-class optimization surface. Teams shipping large-scale training/inference should evaluate whether commodity data center switches are losing headroom to AI-optimized fabrics.