Ahmad Osman says GLM 5.2-class intelligence could run on a single RTX 5090 in 18 months
AI Engineer · youtube · 2026-07-21
In a talk on the “Desktop Frontier,” Ahmad Osman argues that frontier-class intelligence is moving onto personal hardware and that efficiency gains matter more than brute-force model size.
- He predicts GLM 5.2–class intelligence could run on a single RTX 5090 within roughly 18 months.
- The talk emphasizes “impact per parameter” and the shift from server-grade to consumer-grade hardware footprints.
- Osman frames local/open models as a sovereignty issue: users should own the compute stack that runs their models.
- He also walks through the recent evolution from Mistral and Qwen to DeepSeek R1, and why agentic performance and tool calling are becoming central.
- The closing question: if models keep getting smaller and more efficient, does the GPU you buy today become more valuable over time?
More from Infra
- SF Compute founder: buying compute is 'an absolutely awful experience' right now — IgorCarron · 2026-09-11
- SmolVM open-sources persistent computer infrastructure for agents that outlive chat sessions — aniketmaurya · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Local LLM server dilemma: 4x CMP-170HX (price up 53% in 20 days) vs Mac Studio M5 Ultra — rumboll · 2026-09-11
- llama.cpp lands Flash Attention tuning for RDNA4, big prefill gains on AMD — pmttyji · 2026-09-11
- Your p99 latency benchmark may be lying: a deep dive into coordinated omission — Franc0Fernand0 · 2026-09-11