The marketing team wanted AI video. The quote for cloud H100 time was a monthly subscription with a per-second meter attached, and the quote for a 4090 box was “also buy Windows”.
We already had four M4 Macs sitting on desks doing email. Spoiler: that's a render farm.
The whole architecture is a folder
No queue daemon, no Redis, no scheduler. One Mac shares ~/RenderFarm over SMB, every other Mac mounts it, and a job is a 1KB text file. Claiming a job is a single mv:
# priority lane first, then the normal queue
for cand in "$QUEUE"/hi/*.job "$QUEUE"/*.job; do
claimed="$RUNNING/$(basename "$cand").$HOST.$$"
mv "$cand" "$claimed" 2>/dev/null && break # only one Mac wins
donePOSIX rename is atomic. Two Macs racing for the same job means one wins and one moves on — no locks, no leader election, no coordination service to babysit at 3am.
The money number
Four Macs, four different clips at once, ~4× the clips per hour. Near-linear, because nothing crosses the network mid-render: a job file goes in, an MP4 comes out. A $20 gigabit switch is plenty. WiFi would do.
What you cannot do is split one render across Macs — that needs NVLink-class interconnect, and your ethernet is about a thousand times too slow. I wrote up why that fails and what to do instead, because it's the first thing everyone asks.
The part that surprised me
Staging ~87GB of models onto the share took 0.68 seconds and 67MB of disk. On the coordinator the share sits on the same volume as the HuggingFace cache, so publishing the models is a hardlink, not a copy:
# same volume -> hardlink; across volumes -> rsync fallback
FARM_ROOT=~/RenderFarm ./seed_farm_assets.sh # 87GB, 0.68s, 67MB usedHF blobs are content-addressed and never modified in place, so sharing the inodes is safe. Workers reading over SMB can't tell the difference — they just pull 60GB off the switch in minutes instead of days from a throttled HuggingFace connection.
The catch
Apple Silicon is not a 4090. A single 1080×1920 hero clip is tens of minutes, not seconds, and MLX memory is invisible to every free-memory reading macOS gives you — which is how we OOM-killed four Macs simultaneously before adding admission control.
So this isn't a speed play. It's a throughput and cost play: unlimited retries at $0 per render, on hardware you already bought, with the models never leaving the building.
The verdict
If you need one clip fast, rent a GPU. If you need forty clips by Thursday and you own four Macs, you already own the farm — you're just missing the folder and the mv. That's what LTX Mac Farm is: the folder, the claim protocol, and an app so nobody has to learn either.