Local hardware delivers privacy and stability, per speaker

The speaker argues buying AI-capable local hardware today can be the best deal for users who need privacy, stability, or a learning/testing lab. He notes he "paid approximately $3,500 six months ago" and that hardware is now "$1,000 more expensive," but its value increased because it can now run much stronger models — specifically citing "deep-seek flash version 4" and the ability on two machines to handle a "one million contest window." That enables 24/7, private, unlimited token generation without cloud subscriptions.

Three explicit reasons to run AI locally are listed: privacy, stability, and learning/experimenting. Privacy: cloud AI shares user and client data and can be used to train provider models. Stability: cloud models change frequently; local setups remain constant until you change them. Learning/experimenting: local hardware lets you optimize and run custom setups.

On pricing and supply-side constraints, the speaker highlights RAM volatility: "over 300% increase in price," US data-center growth hitting a power bottleneck with a "seven years queue" for electricity, and geopolitical build-outs (Europe, France invested "830 million"). He contrasts US frontier models (claiming "opening eye" spends "$1.6 for every dollar") with China’s push for optimized, open-weight models that are cheaper and 'good enough' for 90% of workloads.

Regarding future hardware, he warns delays and uncertainty: M5 Ultra and "RTX Spark" timing and pricing unknown, "RTX 5090 Super" likely canceled due to RAM costs, and the next-gen "RTX 6090" might arrive 2027 or 2028. Cost examples given: "Two of those machines are around $10,000. $15,000." He reiterates: "You shouldn't buy any hardware to solve a problem that you don't have." Buy local only if you need privacy, stability, or an experimental AI lab; otherwise hybrid/cloud may suffice today, but he expects a drift toward 100% local over time.