Splitting the cost of running DeepSeek: the honest numbers
We ran the comparison the way a sceptical reader would run it, including the column where we lose.
This is the calculation people in local-model communities do for themselves, out loud, in public, with sources. So here it is done honestly, including the part where owning loses.
The four ways to get DeepSeek-class output
| Path | Real 2026 cost | Where it wins |
|---|---|---|
| Cheap API (DeepSeek) | $0.14 in / $0.28 out per M tokens | Unbeatable on price — but filtered, metered, and your data leaves you |
| Pricier APIs (Kimi, GLM, 405B-class) | ~$1–3.50 per M output | Owning becomes cost-competitive at multi-billion-token monthly volume |
| Renting GPUs 24/7 (8× H100) | ~$11,700/month | Buying the same used server (~$160–200k) pays for itself in 14–18 months |
| Hetzner flat server, rented | ~$965 / month, forever | Easiest by far — but five years is about $58,000 spent and nothing owned |
| Your own machine, hosted by us in Bali | From $900 / month, all-inclusive | Roughly the same monthly bill — except the machine is yours |
Start with the column where we lose
DeepSeek's own API sells output at $0.28 per million tokens. A group-owned rig, at the utilisation a group of friends actually achieves, produces output at roughly $1–2 per million. That is three to seven times more expensive per token, and no amount of framing changes it.
So if your question is "what is the cheapest way to get DeepSeek tokens", the answer is the API, and you should stop reading. We are not going to argue otherwise — the audience for this runs the numbers themselves, and one viral takedown of a dishonest pitch would be entirely deserved.
The three break-evens that are real
- Owning beats renting above roughly 50% sustained usage. Below that, rented capacity you switch off is cheaper than capacity you own and leave idle.
- Owning beats the expensive APIs above roughly 3–8 billion tokens a month. At $1–3.50 per million, heavy volume gets to owned-hardware economics fast.
- Owning never beats the cheapest APIs. Not at any volume we can find. Plan around it.
What splitting actually changes
Splitting does not change the cost per token. It changes whether the machine is reachable at all. One person facing $40,000–65,000 for a Club-tier node does not buy it. Ten people facing $4,000–6,500 each might. The split is an access mechanism, not a discount.
It also changes the shape of the bill. A metered API bill scales with how much you use it and never ends. A shared machine is one large payment, then a flat monthly for hosting divided by however many of you there are. For people who use models heavily and hate watching a meter, the flat shape is worth real money on its own.
The sharpest alternative is not the API — it is Hetzner. A flat-rate dedicated AI server with a single 96GB GPU costs about $965 a month. Split five ways that is roughly $193 each, with zero hassle and nothing to organise. If that machine runs what you need, it is a genuinely good answer and we will say so.
Here is the comparison we would rather you make. Hosting a machine you own in our Bali facility starts at $900 a month, all-inclusive — electricity, internet, monitoring and control, and service. Roughly the same monthly figure as renting Hetzner's single-GPU box. One of the two ends with you holding hardware; the other ends with five years of receipts.
The things the price comparison cannot hold
Every figure above is a cost per token. The reasons people actually self-host are not costs:
- The data never leaves. 44% of organisations name data privacy as their top barrier to adopting AI. On your own machine the question does not arise.
- Nothing is filtered by someone else. No refusals, no policy changes announced by blog post, no model quietly swapped underneath you.
- No meter. The cost of thinking harder about a problem is zero.
- You own it. Three years of rental leaves receipts. Three years of ownership leaves hardware you can sell.
Do the sum for your own group
The cost calculator takes the model class you want and the number of people you have, and prices it line by line — parts, design, assembly, matching and hosting — down to what each person pays, with the Hetzner rental shown next to it, because that is the comparison an honest person makes.