The Entire Operation Costs ~$10 a Month
Alflora Research · 01 Aug 2026
Tl;dr
- The whole platform — two servers, an Arc A770, git, CI, MinIO, status page, and an AI chat — costs ~1000₽ (~$10)/month in electricity.
- Hardware is owned outright, so the electricity bill is the entire operating cost.
- Server 1: Xeon E5-2686 v4, 64GB DDR4, Arc A770 16GB. Server 2: dual Xeons, 128GB RAM, bulk storage.
- Last decade's enterprise hardware is this decade's homelab bargain — the cloud rents you less for more.
The inventory
Server 1 is the daily driver: an Intel Xeon E5-2686 v4 — eighteen cores of 2016 enterprise silicon — with 64GB of DDR4 and an Intel Arc A770 16GB, on a Chinese X99 board that cost less than a decent dinner. It runs the web stack, the databases, the AI chat, the CI jobs, and this website, and it has opinions about none of it. Server 2 is the heavy lifter: dual Xeons, 128GB of RAM, and enough storage to make MinIO feel welcome.
None of this hardware was new when it arrived. All of it was enterprise gear doing a second tour of duty, bought for a fraction of what a single year of equivalent cloud instances would cost.
The bill, line by line
At full power, 24/7, both machines together pull enough electricity to cost about 1000 rubles a month — roughly ten dollars. That is the entire recurring cost of the platform: the web hosting, the git server, the CI, the object storage, the status page, the AI chat's body (its brain is rented separately, per request). No instance hours, no egress fees, no per-seat pricing, no surprise invoice because something autoscaled.
For comparison: a single mid-range cloud VPS with 64GB of RAM costs several times that per month, and it comes with a fair-use CPU policy and a stranger's hands on the hypervisor. Here, the same money buys the electricity for two full machines and leaves the hardware itself already paid for.
Why owned beats rented
The cloud's offer is convenience: someone else racks it, cools it, replaces the dead fan at 3am. The homelab counter-offer is arithmetic and control. The arithmetic you just read. The control is that the hardware profile is exactly what the workload needs — a 16GB GPU for inference experiments, ECC-adjacent memory headroom for databases, real storage for object storage — instead of whatever shape the instance catalog happens to sell.
And when something dies, the failure is physical, diagnosable, and fixable with a screwdriver, not a support ticket.
The catch, stated honestly
This math only works because the labor is free — the operator is also the beneficiary, and the 3am rule applies. Power goes out, disks will eventually die, and there is no SLA except personal stubbornness. But that's the trade at the heart of this whole project: rented power is convenient and fragile in someone else's hands; owned power is cheap and fragile in yours. For ~$10 a month, we'll take ours.