Summary
A post on X by @0xSero, responding to Sam Altman's comments that local inference wastes memory, says DeepSeek V4.1 Flash runs unquantized on four power-capped H200 GPUs. The post claims an enterprise with about $120,000 could serve 16 concurrent workers at over 150 tokens per second each. The snippet also describes a post by Patrick Moorhead showing his son's apartment running four DGX Sparks with DeepSeek V4.1 Flash, plus an HP RTX Spark running Qwen 3.8. These figures are the posters' own claims as shared on X.
AI-assisted summary based on the listed source.
What happened
Following Sam Altman's comments about local inference wasting memory, @0xSero says DeepSeek V4.1 Flash runs unquantized on four power-capped H200s, so an enterprise with about $120,000 can serve 16 concurrent workers at over 150 tokens a second each. Patrick Moorhead posted his son's apartment running four DGX...
Why it matters
Builders and businesses weighing local against hosted AI inference may treat these posts as examples of what on-premise hardware can handle, though the numbers are self-reported. Watch for other operators publishing comparable throughput and cost figures for similar setups. The posts add to the ongoing debate over whether local AI is practical at home and in the enterprise.
Signal Intelligence
Signal Strength 83%
Technical label WATCH
Public Interest 38
Category AI AT WORK
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 60
Practical Impact Score 8
Novelty Interest Score 70
Consequence Score 10
Curiosity Score 0
Shareability Score 55
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
