Summary
Ahmad Osman of Osmantic delivers a data-heavy conference presentation tracking the rapid compression of AI capability into consumer and prosumer hardware. Drawing on firsthand experience running models since Llama 2 — which required eight RTX 3090s and delivered performance that a 27 billion parameter Qwen 3.5 model now eclipses on a fraction of the hardware — Osman builds a systematic case that “capability density,” not raw parameter scale, is the defining metric of the current AI era.
The talk introduces the concept of a “densing law” (referenced from Nature Machine Intelligence), estimating roughly 50% parameter reduction every 3.5 months for equivalent capability. Osman’s headline prediction: GPT-4o-class intelligence running on a single RTX 5090 with 32 GB VRAM by late 2027 — a figure he describes as conservative. He walks through the current state of the art, including GLM 5.2 (744B total parameters, 40B activated) running on a GX workstation under a desk, Neutron 3 Ultra demonstrating efficient NVFP4 training, and the milestone of running Claude Code-compatible models on a single RTX 1590-class GPU, a threshold that required four RTX 3090s just a year prior.
Beyond technical benchmarks, Osman makes a pointed enterprise argument: as GPT-4o-quality inference becomes feasible on iPhones and workstations, the case for sovereign, on-device AI strengthens dramatically — for data privacy, cost optimization, and long-term resilience. He frames the open-source ecosystem’s commercial viability as dependent on enterprises migrating away from cloud providers and toward owning their full AI stack.
📺 Source: AI Engineer · Published July 21, 2026
🏷️ Format: Deep Dive
