Daniel Svonava, co-founder of Superlinked, argues that production AI is shifting from single large models to fleets of task-specialized small open source models, and that self-hosting them on a handful of GPUs with pull-based gateways, centralized queueing, model packing, and lazy loading delivers far lower cost and latency than managed embedding APIs. He says small models are no longer dumb but require slicing workloads per task, running evals, and adapting with LoRAs, which breaks the standard top-down inference router. Superlinked's Apache 2.0 answer is a gateway-plus-shared-queue topology where workers pull work and form their own batches; centralizing the queue roughly doubled cluster throughput versus the vLLM/SGLang router approach, which struggles to push GPU utilization past 20–30% under constant load with many small requests. The talk includes performance figures on NVIDIA L4, A100, H100, and RTX Pro 6000, a three-runtime strategy (PyTorch, a Rust/Candle path, and SGLang as baseline), and an auto-research loop shipping pre-tuned configs. It closes with an 80-cent LoRA that improved retrieval quality on German legal text by 18%. Svonava defines "small" as models fitting on a single GPU two to three generations old, citing Qwen 3 27B around GPT-5.1 performance.