Upscale AI raises $190M to build open-standard AI networking fabric

Upscale AI is attacking the least glamorous but increasingly decisive part of the AGI stack: the network between accelerators. As context windows stretch and mixture‑of‑experts models explode in parameter count, the constraint is no longer just FLOPs or HBM, but how fast and predictably data can move across tens of thousands of chips. Upscale’s bet is that an open, memory‑semantic “scale‑up” fabric—compatible with standards like UALink and ESUN—can provide NVLink‑class behavior without being tied to a single GPU vendor.

The investor roster reads like a who’s who of global capital and AI infrastructure: Premji Invest and Temasek from Asia, NVIDIA and Salesforce from the US, Prosperity7 from the Gulf, and a bench of top-tier venture firms. That breadth underscores how central networking has become to the AI profit stack. If Upscale can deliver on its promise of zero‑packet‑loss, open-standard fabrics, it will both pressure incumbent Ethernet vendors and give hyperscalers a neutral alternative to proprietary interconnects.

For the race to AGI, better networking is pure accelerant. It lets labs stitch together larger effective “super‑GPUs” out of commodity hardware and reduces the tax paid in idle cycles while models wait on gradients. That doesn’t solve alignment or safety, but it does make it cheaper and faster to train whatever architectures researchers can imagine.