⚡ your own tiny ai datacenter

Run AI in your own walls.
Know exactly what it takes.

Tell us your machine and how many people will use it — we'll tell you which open-source models you can run, how fast they'll feel, what to buy, what it costs to keep running, and how to set it all up. No jargon required.

111 open models41 hardware optionscatalogs updated 2026-06-09estimates, not quotes — see assumptions
01 · the calculator

What can you run — or what do you need?

Three simple doors into the same engine. Every answer updates live as you move the sliders.

Your setup

just mefamily / teamsmall company

family, team or company members who will chat with your AI

Machine summary

Usable VRAM
8.00 GB
Bandwidth
256 GB/s
GPUs pooled
1
Active at once
≈2
30% of 5 users

Speed and fit are estimates from published hardware specs — real numbers vary with software versions. Tune the details in Advanced settings.

Best pick for you

GLM Zhipu GLM-4-9B-Chat

9.4B parameters · Q4_K_M

Capable — everyday assistant quality

Feels
instant
Instant — faster than anyone reads
Speed / user
41 tok/s
VRAM used
6.86 GB of 8.00 GB
Max comfy users
16
at reading speed
6.86 GB / 8.00 GB VRAM
Yi 01.AI Yi-9B
Q4_K_M
capableinstant · 43 tok/s6 users
7.46 GB / 8.00 GB VRAM
Yi 01.AI Yi-1.5-9B
Q4_K_M
capableinstant · 43 tok/s6 users
7.46 GB / 8.00 GB VRAM
Qwen2.5-VL-7B-Instruct
Q4_K_M
capableinstant · 46 tok/s16 users
6.41 GB / 8.00 GB VRAM
IBM Granite 3.3 8B Instruct
Q3_K_M
capableinstant · 59 tok/s6 users
7.02 GB / 8.00 GB VRAM
Qwen3-8B
Q4_K_M
capableinstant · 47 tok/s6 users
7.88 GB / 8.00 GB VRAM
Llama 3.1 8B Instruct
Q4_K_M
capableinstant · 48 tok/s6 users
7.49 GB / 8.00 GB VRAM
DeepSeek-R1-Distill-Llama-8B
Q4_K_M
capableinstant · 48 tok/s6 users
7.47 GB / 8.00 GB VRAM
Cohere Command-R7B (8B)
Q4_K_M
capableinstant · 48 tok/s6 users
7.47 GB / 8.00 GB VRAM
Qwen2.5-7B-Instruct
Q4_K_M
capableinstant · 50 tok/s20 users
5.97 GB / 8.00 GB VRAM
02 · money

What it costs to buy and to run

CapEx is what you pay once; OpEx is what you pay every month. Home mode keeps it minimal; enterprise mode adds the team and upkeep costs that on-prem deployments really carry.

Costing: Example — Llama 3.1 8B on a used RTX 3090 (24 GB) (pick something in the calculator above to replace this example)

Buy (CapEx)
$1,040
one-time · SAR 3,900
Run (OpEx)
$86/mo
$1,033/yr · SAR 323
3-year total
$4,140
CapEx + 3 yrs OpEx
Per user / month
$22
5 users, incl. hardware depreciation

One-time costs (CapEx)

Compute hardware (GPUs / machine)
$800
UPS battery backup
rides through power blips so the box never hard-crashes
$150
Networking (router / switch / cabling)
a decent router or a $30 switch usually suffices
$90
Total to buy$1,040

Monthly costs (OpEx)

70 kWh/mo in Saudi Arabia
Electricity (incl. cooling overhead)
70 kWh/mo at $0.05/kWh × 1.35 cooling factor
$4/mo
Internet uplink
$80/mo
Maintenance & repairs reserve
3% of CapEx per year set aside for fans, drives, thermal paste
$3/mo
Total per month$86/mo

Usage pattern

how hard the box works
03 · hands-on

Set it up — from one laptop to a mini-cluster

Pick your level of ambition. Level 1 genuinely takes about fifteen minutes.

Level 1 — One machine, 15 minutes

for 1–5 users

The easiest path. Ollama is a free app that downloads and runs models for you — no configuration files, no code.

  1. 1Download Ollama from ollama.com and install it like any app (Windows, Mac or Linux).
  2. 2Open a terminal and pull a model that fits your machine (the calculator above told you which):
    ollama run llama3.1:8b
  3. 3That's it — you're chatting. For a nicer look, install a web chat like Open WebUI so the family can use it from their browsers.
  4. 4To let others on your Wi-Fi use it, start Ollama so it listens to the network:
    OLLAMA_HOST=0.0.0.0 ollama serve

Level 2 — Add more GPUs

for bigger models or 10–50 users

Two or more graphics cards can pool their memory to hold a model neither could hold alone. The software slices the model across cards — you don't manage that yourself.

  1. 1Physical: your motherboard needs a free full-length PCIe slot per card, and your power supply needs headroom (~350 W per big card — the calculator's PSU estimate includes this). Leave a finger of air-gap between cards for cooling.
  2. 2Same brand, ideally same model of GPU keeps things simple (e.g. 2× RTX 3090). Mixed cards work but run at the slower card's pace.
  3. 3Switch from Ollama to vLLM — the serving software built for many users — and tell it how many cards to slice across:
    vllm serve meta-llama/Llama-3.3-70B-Instruct --tensor-parallel-size 2 --quantization awq
  4. 4The flag --tensor-parallel-size 2 is the whole trick: it means "split every layer across 2 GPUs." Use 4 for four cards. Everything else is automatic.

Level 3 — A few machines

department / small enterprise

When one box isn't enough, you run several and need a coordinator. That's where Kubernetes comes in (explained below) — for nano scale, use its lightweight cousin k3s.

  1. 1Install k3s on your main machine (one command — it becomes the "manager"):
    curl -sfL https://get.k3s.io | sh -
  2. 2Join each extra machine as a "worker" with the join command k3s prints. Now all your machines act as one pool.
  3. 3Install the NVIDIA GPU Operator so the pool knows which machines have GPUs, then deploy vLLM as a service. Ready-made recipes exist (search "vLLM production stack helm").
  4. 4What you gain: if one machine dies, work moves to another; updates roll out without downtime; one web address serves everyone.
Do I actually need this?

Honest answer: probably not until you have two+ machines AND uptime requirements. A single strong box with vLLM serves 50+ light users. Kubernetes earns its complexity when downtime costs money or you juggle several models across several machines.

04 · plain language

Every concept, explained like you're smart but busy

VRAM, quantization, Kubernetes — the nine ideas behind this whole page, one card each.

AI model & parameters

The "brain" you download. Bigger = smarter but hungrier.

Tell me more

An open-source model is a single big file of numbers ("parameters"). An 8B model has 8 billion of them. More parameters generally means better answers — and more memory needed. The open-source community (Meta, Mistral, Qwen, DeepSeek…) releases these free for anyone to run.

VRAM

Your graphics card's own memory — the #1 thing that decides what you can run.

Tell me more

The whole model must sit inside the GPU's memory (VRAM) to run fast. A card "with 24 GB" can hold roughly a 30B-parameter model at 4-bit compression. Regular RAM can substitute, but it is ~10× slower — that's why VRAM is the first number this calculator asks about.

Quantization

Smart compression: the same brain in half the memory, barely dumber.

Tell me more

Models are trained with very precise numbers (16-bit). Quantization rounds them to 4–8 bits, shrinking memory 2–4× with only a small quality loss. Q4 (4-bit) is the community sweet spot — it's why a 70B model fits on two consumer GPUs at all.

Tokens & speed

AI writes in tokens (~¾ of a word). 12 tokens/sec feels like reading speed.

Tell me more

Every response streams out token by token. Below ~5 tok/s users get impatient; 10–15 tok/s matches reading speed; 30+ feels instant. Generation speed is mostly limited by how fast the GPU can re-read its memory — which is why memory bandwidth, not raw compute, rules this calculator.

Inference engine (Ollama vs vLLM)

The software that serves the model. Ollama = easy, vLLM = many users.

Tell me more

Ollama (built on llama.cpp) is a one-click app, perfect for one to a handful of users. vLLM is server software that interleaves many conversations at once ("continuous batching"), serving 10–100× more users from the same GPU — at the cost of a slightly more technical setup.

Multi-GPU (tensor parallelism)

Two cards pool their memory; the software slices the model across them.

Tell me more

With "tensor parallelism" every layer of the model is split across your GPUs, so 2× 24 GB cards behave like one 48 GB card (with a small efficiency tax). You enable it with a single flag — the setup guide shows exactly how. Separate computers can NOT pool memory this way; that requires the cluster tools below.

Kubernetes

An "operations manager" for a fleet of machines. Overkill for one box.

Tell me more

Kubernetes (K8s) is software that treats several computers as one pool: you declare "I want the AI service running with 2 copies", and it decides which machine runs what, restarts crashes, rolls out updates without downtime, and gives users one stable address. In a nanodatacenter its role begins when you have 2+ machines and uptime matters — and the lightweight k3s edition is the right size. One home PC running Ollama does not need Kubernetes, the same way one food truck does not need an airport control tower.

Context window (KV-cache)

How much conversation the AI remembers — it costs memory per user.

Tell me more

The context window is the AI's short-term memory (2K tokens ≈ 3 pages, 128K ≈ a novel). While answering, the GPU keeps a "KV-cache" for every active conversation — so memory needs grow with both context length AND simultaneous users. That's why the calculator asks how many people use the system at once.

PSU, UPS & cooling

Feed it clean power, and remember: every watt becomes heat.

Tell me more

A big GPU draws 300–450 W under load; the power supply (PSU) should have ~1.7× headroom. A small UPS (battery) rides through flickers so the machine never hard-crashes. All consumed power turns into heat in the room — in hot climates the air conditioner then spends extra energy removing it (the "cooling factor" in Advanced settings).

05 · perspective

Nanodatacenter vs full-scale datacenter

Hyperscale planners calculate megawatts, hectares and water rights. At nano scale the same physics collapses into different, friendlier questions — here's the side-by-side.

Power
Full-scaleMegawatts. Grid substations, medium-voltage feeds, diesel generators, N+1 redundant UPS halls. Power is THE design constraint.
NanoHundreds of watts to a few kW. The real question: can one wall circuit (typically 1.8–3.5 kW) feed the box, and is a small UPS attached?
Area & space
Full-scaleThousands of m² of "white space", raised floors, hot/cold aisles, loading docks. Land price and floor loading matter.
NanoA desk corner, a closet, or one 12U rack (~0.35 m²). The question is noise and airflow, not square meters.
Cooling
Full-scaleIndustrial chillers, cooling towers, liquid loops; PUE (total power ÷ IT power) engineered from 1.1–1.5. Water use is a siting issue.
NanoThe room's AC or a good airflow case. We model it as a simple "cooling factor" on electricity — 1.1 in a temperate room, up to 1.4 in Gulf summers.
Redundancy
Full-scaleTier III/IV: everything duplicated (power paths, cooling, network). Downtime is measured in minutes per YEAR.
NanoAccept that a reboot happens. A UPS plus good backups covers a home; an enterprise nano-rack adds a spare PSU and maybe a second box with k3s.
Network
Full-scaleMultiple fiber carriers, BGP peering, DDoS scrubbing, private links to clouds.
NanoYour existing broadband. For remote access, a simple VPN (Tailscale) beats opening ports.
People
Full-scale24/7 NOC shifts, security guards, facilities engineers, compliance teams — dozens of salaries in the OpEx model.
NanoHome: nobody, you reboot it on Saturday. Enterprise: a slice of one engineer (the % FTE slider in the cost panel).
Cost basis
Full-scale$/MW built, $/kWh at industrial tariffs, land, taxes; ROI over 15–20 years.
Nano$/GPU bought, home or SME electricity tariff, hardware depreciated over ~4 years. CapEx dominates; talent only appears at enterprise scale.
Siting & geography
Full-scaleChosen for cheap power, cool climate, water rights, fiber routes, political stability.
NanoYou host where you already are — geography only shows up in your power price and how hard your AC works (both tunable in Advanced settings).

The takeaway: hyperscale calculators optimize megawatts, land and water; a nanodatacenter calculator optimizes VRAM fit, felt speed per user, one power circuit, and a realistic monthly bill. Same physics, very different knobs.

06 · advanced settings

Assumptions & geography — tune the model to your reality

Everything above is an estimate built on the assumptions below. If you know your real electricity tariff, cooling situation or workload shape, set it here and every number on the page updates.

🌍 Geography

drives electricity, cooling & currency

your actual tariff beats our preset

1.1 temperate room · 1.4 hot climate with AC

Hot climates (Gulf summers) make the air-conditioning work harder for every watt the GPUs draw — that's the cooling factor. Cheap electricity (Saudi Arabia ≈ $0.05/kWh) can make self-hosting dramatically cheaper to run than in Germany (≈ $0.40/kWh).

⚙️ Model & workload

how the sizing math is tuned

of total users, typing at the same moment

~12 tok/s matches human reading speed

What the calculator assumes (full list)
  • Memory math: model weights = parameters × bytes-per-parameter for the chosen compression; conversation memory (KV-cache) scales with context length × concurrent users; +15% framework overhead on top.
  • Speed math: generation is limited by memory bandwidth — tokens/second ≈ bandwidth ÷ active model size. Multi-GPU adds ~85% of each extra card's bandwidth. Real speeds are usually 60–90% of this ceiling.
  • Concurrency: not everyone types at once — default assumes 30% of users are active at any moment. vLLM batches many users efficiently; Ollama is best for a handful.
  • Quality bands: parameter count is used as a rough quality proxy (starter <4B, capable 4–15B, strong 15–80B, frontier 80B+). Benchmarks vary per task — always test your use case.
  • Power & cooling: wall power = IT power × average load × cooling factor. PSU sizing uses GPU TDP × 1.7 + 150 W for the rest of the box.
  • Prices: hardware prices are street estimates (June 2026 refresh) and drift constantly; electricity presets are national blended averages. Enterprise staffing figures are illustrative defaults, not quotes.
  • Geography: only electricity tariff, cooling burden, internet cost and currency are modeled. Import duties, VAT, and local hardware availability are NOT included — add them to CapEx mentally or via a custom price.