A $1,200 AI Training Run Outperformed Postgres Query Planning by 81 Percent

A researcher trained a 4-billion-parameter model on two second-hand RTX 3090s for just $1,200, producing database query plans 81 percent faster than PostgreSQL on the Join Order Benchmark. The model, a LoRA adapter weighing only 42.5 megabytes, was fine-tuned using supervised learning and reinforcement learning on the open-weights Qwen 3.8 4B architecture. Starting from near-useless baseline performance, reinforcement learning with measured execution times as a reward signal pushed the model to a 1.81x geometric mean speedup and a 44.7 percent reduction in total workload latency. The experiment highlights that domain-specific AI breakthroughs may hinge more on training methodology than on large-scale infrastructure, even as the Philippines pursues a $34.4 billion AI infrastructure roadmap targeting a 30-fold expansion in data center capacity by 2033. The contrast between the modest $1,200 experiment and billion-dollar national AI investments raises questions about where the true bottlenecks in AI development actually lie.