← niklaslenz.com NIKLAS LENZ Follow the build ↗

Free PDF · The Jarvis Setup Pack

What a private AI setup actually costs.

You saw the reel. This is the honest, number-by-number answer — five budgets, from €0 to the full two-machine setup, with parts lists at every tier, benchmarks I measured on my own hardware, and the three things that broke while I built mine. No affiliate links. No hype.

Free. Enter your email and confirm — double opt-in. Unsubscribe anytime. See the Privacy Policy. No spam, no resold list.

What’s inside

Six things you’ll know by the last page.

A map, not a sales pitch — so you can decide what to buy for your budget, even if that budget is zero.

  1. 01

    The honest receipt

    The “$10,000 setup” is really two machines, verified at €6,982 (System A) and €2,627 (System B) — ≈ €9,608 together. Not €15k of monitors and furniture; only the two boxes that actually run models.

  2. 02

    Five budgets, €0 → ~€10k

    From the laptop you already own, through a first GPU, an always-on daily driver, a single heavy machine, up to the full two-box system — with concrete parts lists and prices at every tier.

  3. 03

    Benchmarks I measured myself

    Qwen3.6-27B at ~70 tokens/second on the RTX 5090; FLUX.1-dev images in ~12 seconds; HunyuanVideo 1.5 5-second clips in ~2 minutes. The only speed numbers in the pack that are mine — everything else is clearly marked community data.

  4. 04

    The 2026 memory reality

    DDR5 RAM at roughly 3–4× its 2024 price, NVMe SSDs at 2–3×, GPUs well above list. Why this is a bad year to overbuy — and how to build around it (including a used-platform route that sidesteps the worst of it).

  5. 05

    What local hardware is actually for

    Not “replacing ChatGPT.” Frontier models still do the hardest thinking. Local wins a specific set of jobs — high-volume agent tasks, privacy, and image/video generation — and memory beats model size: a smaller model plus your own indexed data out-answers a giant model working blind.

  6. 06

    The three things that broke

    The newest CUDA was the broken one on the RTX 5090; a Linux service died every time I logged out; a mainboard failure took System A down for five days — and the whole time, the 24/7 work kept running on a separate machine. The lessons, not the drama.

I’m Niklas Lenz — I build local-first AI systems on my own hardware and document every step in public: the wins and the parts that broke. The Setup Pack is the map; the build itself plays out on the homepage.
More about the build →

Technical information, not purchase advice. Prices and specs are as of July 2026 for the German market (EUR) and move fast — especially during the current memory shortage. Verify current prices before buying. System figures reflect my own setup; media-generation times are p50 benchmarks measured on my RTX 5090 (July 2026). Community benchmark numbers for other hardware are marked as such in the pack.