When Adebayo Alonge demonstrated his Rx Scanner spectrometer in Cape Town in 2019, the cloud-based AI failed — latency from a U.S. data center over limited bandwidth stretched a single pill scan to five minutes. His engineers compressed the model to run on an Android phone in two hours, creating an offline version that authenticates medication without broadband, computers, or reliable electricity. This pivot mirrors a broader shift: according to a World Bank report, only 0.7 percent of internet users in the poorest countries have used ChatGPT, versus 25 percent in developed nations.

Small AI — models with a few billion parameters versus trillion-parameter LLMs — now powers drone-based cashew disease detection in India, ant infestation monitoring in Uruguay, malaria mosquito identification across multiple nations, and Arduino-based electrocardiograms in rural Brazil. Marcelo Rovai of the Federal University of Itajubá calls this "the most important area in AI nowadays." These systems rely on pruning, distillation, and quantization (32-bit to 8-bit) to shrink models for specific tasks.

Hardware advances accelerate deployment. Counterpoint projects 45 percent of 2025 smartphones will run generative AI, exceeding 50 percent in 2026. Open-weight models like Google DeepMind's Gemma 4 and Alibaba's Qwen 3.5 enable domain retraining — Rovai demonstrated a $50 Arduino UNO Q with a Qualcomm chipset running a language model that analyzes sensor data to detect mosquito breeding pools at three watts.