Building and running AI on hardware you own — GPUs, clusters, interconnects, power, and the measured numbers behind them.https://www.privateai.blog/enhttps://www.privateai.blog/en/posts/teaching-my-logo-to-wavehttps://www.privateai.blog/en/posts/teaching-my-logo-to-waveA fully self-hosted pipeline: script to voiceover to render, no cloud in the path. Two mascots, worst-to-best, the 3090/4090/DGX numbers, and why the right model depends on your character.Sun, 30 Aug 2026 00:00:00 GMTon-prem-aivideo-generationwancomfyuirtx-4090rtx-3090benchmarksopen-sourcemascotcampaign:launchhttps://www.privateai.blog/en/posts/who-made-the-weights-you-runhttps://www.privateai.blog/en/posts/who-made-the-weights-you-runNvidia is reported to be buying Hugging Face. A good week to learn what is really on the shelf: master weights, vendor builds, and the community quants most of us run.Sun, 30 Aug 2026 00:00:00 GMTon-prem-aihugging-facequantizationnvfp4fp8bf16glmvllmopen-weightsmodel-archivesupply-chainhttps://www.privateai.blog/en/posts/gpu-passthrough-cost-me-nothinghttps://www.privateai.blog/en/posts/gpu-passthrough-cost-me-nothingA year after virtualizing my edge AI supercomputer, I finally measured what the hypervisor costs: single-GPU work is free — the VM even won — and four-way tensor-parallel serving pays a third.Sun, 23 Aug 2026 00:00:00 GMTon-prem-aivfiogpu-passthroughproxmoxvllmnccltensor-parallelsdxlrtx-3090benchmarksvirtualizationcampaign:launch