PRIVATE AI
COST CALCULATOR.
What it costs to run your own AI model on hardware you own — and whether it pays back against the seats and API bills you have now. Runs in your browser, no signup.
Oct 2026 price. Get a live quote.
What you pay, not list.
Needs about 47 GB (39 GB of weights plus working room) on a box with 96 GB. Speed ceiling: about 46 tokens/sec per user — memory bandwidth divided by the weights read per token. Expect roughly half to two-thirds of that in practice.
Over 3 years: own hardware $51,970 vs current AI $45,000. Owning works out to $1,444 a month with the purchase spread over its life.
Show as table
| End of | Own hardware | Current AI |
|---|---|---|
| Year 1 | $38,657 | $15,000 |
| Year 2 | $45,313 | $30,000 |
| Year 3 | $51,970 | $45,000 |
A model's weights must sit in GPU or unified memory, plus room for the conversation. If it does not fit, nothing else on this page matters. Lower precision (4-bit) roughly quarters the memory a 16-bit model needs, at a small quality cost you should test on your own work.
Every word the model writes reads its active weights from memory once. So a box with lots of memory but slow memory, like the DGX Spark at 273 GB/s, can load a 70B model and still answer at single-digit tokens per second. Mixture-of-experts models read only a fraction of their weights and stay fast.
Hosted AI is cheap per word because providers share each GPU across thousands of customers. Owning wins when your data must stay inside your walls — and pays back only when it replaces real, recurring spend. If the payback lands after the hardware's life, buy it for control or not at all.
On the numbers: hardware prices are October 2026 list and street prices. A memory shortage pushed them up sharply this year — Nvidia’s RTX PRO 6000 went from about $8,565 at launch to $16,000, and the DGX Spark from $3,999 to $6,950 — so treat every default as a starting point and replace it with a quote. Power assumes full draw during busy hours and ignores idle draw. Read how we size these deployments →
FAQ.
Is running our own AI model cheaper than ChatGPT or an API?
Per word, usually not. Hosted providers share each GPU across thousands of customers, so open models sell for cents per million tokens. Owning pays back when it replaces meaningful recurring spend — many seats, or a large API bill — and the hardware lasts long enough to recover its price. The stronger reason to own is control: the data never leaves your building and nobody changes the model under you.
What hardware do we need to run a 70B model?
At 4-bit precision a 70B model's weights take about 39 GB, plus working room, so a single 96 GB card such as the RTX PRO 6000 Blackwell holds it comfortably. A 32 GB card does not. A 128 GB box like the DGX Spark also holds it, but its memory bandwidth limits it to a few tokens per second per user.
Why is a box with lots of memory still slow on big models?
Writing each token means reading the model's active weights from memory once, so speed is capped at memory bandwidth divided by the size of those weights. The DGX Spark has 128 GB but 273 GB/s of bandwidth; a 96 GB RTX PRO 6000 has about 1,792 GB/s. Same model, roughly six times the ceiling. Mixture-of-experts models like gpt-oss-120b read only a small part of their weights per token, so they stay fast on big, slower memory.
Why have AI hardware prices gone up in 2026?
A shortage of memory chips. Nvidia's RTX PRO 6000 Blackwell launched at about $8,565 and is listed at $16,000 on Nvidia's store in autumn 2026; the 128 GB DGX Spark went from $3,999 at launch to $6,950. Consumer RTX 5090 cards sell well above their $1,999 list price. Quotes go stale in weeks, so the calculator lets you replace every price.
Can we run DeepSeek on our own server?
The full DeepSeek V3 or R1 model has 671 billion parameters and needs hundreds of gigabytes of GPU memory — a data-center server with eight H200-class GPUs. The 'DeepSeek' models people run on a single workstation are distilled versions: smaller Qwen or Llama models trained on DeepSeek's outputs. They are useful, but they are not the same model.
What does 'setup and training' cover?
Choosing the model for your tasks, installing a production inference server, connecting it to your documents (retrieval) and your apps, and, where it pays off, fine-tuning it on examples of your own work. The step that decides success is a test set of your real tasks with correct answers, so every choice is measured rather than guessed.
Is this calculator free, and is anything stored?
Free, no signup, and nothing is stored or sent anywhere. Every number is calculated in your browser.