Upscale AI raises $190M to challenge Nvidia's networking lock-in

This funding signals a critical move toward hardware agnosticism in the AI stack. By building a high-speed networking fabric that allows heterogeneous GPUs to communicate at full speed, Upscale AI is targeting the proprietary bottleneck that currently forces hyperscalers into single-vendor ecosystems. For the streaming and cloud industries, success here means lower infrastructure costs and greater flexibility in sourcing compute for massive encoding and inference workloads. As AI infrastructure spending is forecasted to hit $100 billion annually by 2030, the ability to mix and match accelerators from AMD, Intel, and Nvidia without performance penalties will define the next phase of data center efficiency. Watch for the first real-world performance benchmarks of the Skyhammer architecture against Nvidia's Blackwell Ultra platform in late 2026.

The push for open AI networking standards is accelerating as hyperscalers attempt to diversify their hardware dependencies. Per EE Times (June 2026), the UALink Consortium recently ratified the UALink 2.0 Specification, which introduces 'in-network compute' to reduce latency between accelerators. This standard is specifically designed to manage the memory-semantic traffic required for training multi-trillion parameter models, providing a direct alternative to Nvidia's proprietary NVLink. While Nvidia remains a dominant force, the consortium now includes over 90 companies, reflecting an industry-wide effort to standardize the 'scale-up' fabric inside the server rack.

Competition in the pure-play AI networking space is intensifying alongside these standards. Per SiliconAngle (March 2026), direct rival Nexthop AI closed a $500 million Series B round at a $4.2 billion valuation, positioning itself similarly as an open-source switching alternative for neocloud providers. Simultaneously, traditional Ethernet is closing the performance gap with InfiniBand. According to a June 2026 report from Dell'Oro Group, Ethernet switch sales for AI back-end networks more than doubled in Q1 2026, now accounting for two-thirds of all AI-cluster port shipments. This shift is largely driven by AI-optimized platforms like Nvidia's own Spectrum-X, which ports InfiniBand-style features like adaptive routing into standard Ethernet environments.

Nvidia’s participation in the Upscale AI round highlights a dual-track strategy. While maintaining its proprietary Blackwell and NVLink ecosystem, the chipmaker is also investing in the very technologies that facilitate heterogeneous computing. Per Network World (June 2026), Nvidia’s gilad Shainer noted that the 'AI factory' model requires multiple types of specialized networks. By backing Upscale AI, Nvidia secures a stake in the emerging UALink-based market while continuing to push its vertically integrated stack for the highest-performance training tiers.