Nota's GPU-optimization tech becomes core infrastructure for two Korean national AI programs.
The AMW Read
Extends a known optimization player's momentum (July ICML placement) into embedded infrastructure roles across two national-scale AI programs, with quantified GPU and memory savings.
Nota's GPU-optimization tech becomes core infrastructure for two Korean national AI programs.
Nota AI, a decade-old Korean AI-optimization company, now provides underlying tech for both of South Korea's flagship national AI programs: the government's independent foundation-model project (Dokpamo), led by Upstage, and the nationwide public-AI service Modu's AI, led by SK Telecom. Applied to Upstage's Solar Open2 model, Nota's quantization and pruning cut required Nvidia H100 GPUs from eight to two, reduced memory use 76.5%, and lifted inference speed 1.49x — savings GPUPoet estimates at roughly 280 million won (~$200K) per server. NetsPresso, Nota's optimization platform, spans Nvidia GPUs, NPUs, CPUs, mobile, and robotics hardware, and Nota also joined AMD's Robotics Partner Network.
As AI development shifts from training frontier models to running them profitably, inference-time efficiency has become its own competitive layer — visible in Nvidia's own acquisitions of optimization startups OmniML, Deci AI, and Run:ai. Two structurally different consortia, one model-focused (Upstage), one distribution-focused (SKT), independently chose Nota, suggesting optimization vendors can sit underneath competing programs if they stay hardware-agnostic. Per the AI Market Watch index (coverage of ~5,000 tracked companies, not a census), Nota logged 158 news items in the past 90 days versus 80 prior, following its July third-place ICML finish on Qwen inference speedup.
For builders, quantization and pruning are becoming a procurement filter for large model consortia constrained by H100 supply and server cost, not a pre-launch nicety. For investors, Nota's spread across government, telecom, chip-partner (AMD, Qualcomm), and enterprise channels — on $42.6M in total funding, per the AI Market Watch index — shows inference-optimization vendors can build multi-customer distribution without owning a foundation model.
