Home server and local AI, with the conditions attached.
Guides and calculators for NAS storage, power, noise and running models on hardware you own. Every figure says where it came from, and nothing is called measured unless it came off the bench in my lab.
Start here
Everything →You have never run a model on your own machine.
Run Your First Local LLM with Ollama, Step by StepTokens per second simulatorWhat does 8 tokens per second actually feel like?You need to know whether a model fits your card.
How Much VRAM You Need for Every Model SizeModel fit calculatorWill this model run on my GPU, and at what context length?You are picking drives and a layout for a NAS.
NAS RAID Levels Explained for NewcomersNAS capacity and powerHow much space do I actually get, and what does it cost to run?You want to know what leaving it switched on costs.
What a Homelab Really Costs in ElectricityPower cost calculatorWhat does this machine cost me a year?You are deciding whether to buy hardware or keep renting it.
Cloud GPU vs. Owning Hardware: The Real Cost MathLocal vs cloud break-evenIs running AI locally actually cheaper than the API?The machine has to live in the room where you sleep.
How Loud Is Two Machines in One Room?Noise combinerTwo machines at 38 and 41 dB(A). How loud is that together?You want the lab to survive a power cut.
What Size UPS Does a Home Lab Need?UPS sizing calculatorHow many VA do I need, and how long will it actually last?
Latest Articles
View all →How I Benchmark a Local LLM
The protocol, the commands, and the first QuietWatts row: Qwen3.8 27B on a DGX Spark, measured by a harness rather than a stopwatch on a chat window.
Can You Run AI on a NAS?
Usually yes, and it will be slower than you want. But the electricity is already paid for, which changes the argument more than the speed does.
Proxmox vs. TrueNAS vs. Unraid
Three systems built around three different assumptions about what a home server is for. Picking well means knowing which assumption matches yours, not which is best.
NAS RAID Levels Explained for Newcomers
What each RAID level costs you in capacity, what it actually protects against, and why the rebuild after a failure is the part worth planning for.
What a 32k Context Window Costs You in Memory
Context is not free and it is not fixed. The cache grows with every token you exchange, which is why a model that loaded fine runs out of memory an hour into a conversation.
How Loud Is Two Machines in One Room?
Adding a second machine adds three decibels, not double the noise, and that one fact changes where the money should go when a lab gets too loud.
Tools
All tools →- Model fit calculator
Will this model run on my GPU, and at what context length?
- Power cost calculator
What does this machine cost me a year?
- Local vs cloud break-even
Is running AI locally actually cheaper than the API?
- Tokens per second simulator
What does 8 tokens per second actually feel like?
- NAS capacity and power
How much space do I actually get, and what does it cost to run?
- Noise combiner
Two machines at 38 and 41 dB(A). How loud is that together?
- UPS sizing calculator
How many VA do I need, and how long will it actually last?
Browse by Topic
All topics →Why this exists
Almost nothing written about running models at home is measured. Vendor pages quote peaks under conditions nobody discloses, and the lists that rank for these searches have never touched the hardware.
Method first, verdicts second. How I test, and what has been measured so far.