A homelabber is strapping 36 Nvidia DGX Sparks together for 4.6TB of local AI memory
The most interesting AI data center on the internet right now sits in somebody's server room. A redditor known as Kurcide — who built what he believes was the first 16-node DGX Spark cluster earlier this year — has posted an upgrade: 20 more of Nvidia's palm-sized AI boxes are going into the rack, taking "the All Spark" cluster to 36 machines and roughly 4.6TB of unified memory. That is enough to hold some of the largest open-weight models in existence entirely on one private machine, no cloud account required.
The build itself reads like a small data center shrunk into a home rack: 36 DGX Spark units wired through a 200/400-gigabit switch with direct-attach copper cables, a dedicated 100-amp electrical panel feeding 240V outlets, and an owner who says the whole thing draws somewhere around 15 kilowatts and is pushing the limits of his home's cooling. Commenters did the back-of-envelope math at something north of $150,000, and the owner didn't argue. Asked why not simply buy a rack of Nvidia's enterprise-class B300 systems instead, his answer was less about money than about physics and control: workstations with big GPUs run hot and loud for a homelab, still lack the pooled memory to serve frontier-scale models, and cloud clusters give up the thing this build exists to protect — zero data leaving the premises.
What makes the post more than hardware flexing is how the cluster is actually used. Rather than one giant inference machine, it's carved into "inference modules" — separate pools of nodes that handle reranking, embeddings, image and video generation, and audio processing simultaneously, all rolled up under a single persistent agent running Hermes with a custom memory sidecar. Sixteen nodes are earmarked for state-of-the-art models like Kimi K3; the rest keep the agent's senses running in parallel. It's a working preview of where local AI is heading: not one model on one box, but a whole agentic stack owned outright. The economics still only make sense for a wealthy enthusiast — but every capability on display here trickles down fast.
What to watch: whether Kurcide publishes real throughput numbers once tensor-parallel runs across 16 nodes are tested — that would settle the loudest debate in the thread about whether clusters of small boxes can ever match one big accelerator.
Would you trade a cloud subscription for a server rack in your home if the price were right? Tell us in the comments.
