“What does a render farm cost” has three answers because there are three completely different things wearing the name. Here is each, and then the question that decides whether you need one.
1. A commercial farm — per node-hour
You upload a project, it is distributed across their machines, you pay for compute by the hour. Priced per node-hour, often with a priority multiplier, and the bill scales with exactly what you use.
The cost that surprises people is not the rate, it is the iteration. A meter running while you decide whether the last version was right changes how people work: you batch up decisions, you avoid exploratory runs, and the creative cost of that is real but never appears on an invoice.
2. A cloud GPU — per second, and still running
Rent the instance, install what you need, pay by the second. More flexible and usually cheaper per unit of compute, with one specific failure mode: it bills while it is on, not while it is working. An instance left up overnight after the job finished is the single most common way this gets expensive.
3. Machines you already own — $0 per render
Not metaphorically. There is no meter. You already bought the hardware, the electricity is rounding error against a rented GPU, and the footage never leaves your network.
The cost moves to setup and operations instead. Models have to be staged on every node — ours move over a plain switch and land by hardlink, 87GB in 0.68 seconds once the file is local, though the first transfer over the wire is a normal large copy. You need admission control, because a worker that claims a job it cannot afford gets killed by the kernel. And you need a queue that re-issues a killed job to a bigger machine rather than handing it straight back to the one that just died.
That is a weekend of work and then it is infrastructure. It is not free in time; it is free in money, permanently.
The question that actually decides it
Not cost. How many jobs do you need, and by when?
A farm parallelises at the job level. Four machines means four clips at once — a real, near-linear 4× on a batch. It does nothing at all for a single clip, because one render cannot be split across machines; the reason is worth understanding before you buy anything.
So: if you need one clip faster, a farm is the wrong purchase at any price. If you need forty by morning, it is the only thing that works, and the cheapest version is the machines already sitting idle in your office after 5pm.
The verdict
Commercial farms sell convenience. Cloud GPUs sell flexibility. Your own machines sell the absence of a meter, which changes behaviour more than the money does. Setting the third one up starts here.