The shift from 'General' to 'Specific' compute
While headlines focus on $100B training clusters, a quiet rotation is happening in venture funding the deployment layer of AI. Two major rounds total nearly $270M, betting the future isn't just bigger models but running them cheaper and locating them precisely in the real world.
Taalas: Hardwiring the model into the chip
Toronto-based Taalas secured $169M led by Quiet Capital to bake model weights directly into silicon. By hardwiring parameters (e.g., Meta's Llama 3), Taalas claims 1000x better energy efficiency than a GPU for inference. The risk: if the model architecture changes, the chip is obsolete.
ZaiNar: The 'GPS for Indoors' finally arrives?
After nine years in stealth, ZaiNar announced a $100M round valued at $1B. The company uses existing 5G and Wi-Fi signals to track devices with sub-meter accuracy in 3D, solving the indoor positioning problem for Physical AI agents (robots, drones, autonomous logistics).
