Skip to content

Next Steps


alpha-0.1.0 — Get the Cluster Running

Hardware Needs

Procurement

Ray handles hardware procurement. Ask Vidya for his recommendations on parts and vendors.

Current state: Only the Omarchy box (DNS server), one control-plane node, and zero worker nodes are running. The bottleneck is power supplies — we don't have enough to power the remaining machines.

Item Qty Why Link
OptiPlex 65W chargers (or USB-C chargers with barrel adapters) 4+ Power the 4 already-flashed nodes
DYMO LetraTag Bluetooth label maker 1 Label each machine with hostname, IP, and specs DYMO LetraTag on Amazon

With 4 chargers we can bring up 1 control-plane + 3 workers, which is enough to run real workloads.

Hardware Inventory

Specs vary across the OptiPlex fleet — CPU model, RAM size, and NVMe capacity are not uniform. Before deploying workloads, inventory every machine:

  • CPUtalosctl read /proc/cpuinfo (i3 vs i5 vs i7, core count)
  • RAMtalosctl read /proc/meminfo (8GB vs 16GB)
  • NVMetalosctl read /sys/block/nvme0n1/size (128GB vs 256GB vs 512GB)

Label each machine with the DYMO labeler (hostname, IP, CPU, RAM) so specs are visible at a glance on the physical hardware.

Software Plan

Once the 4 nodes are online (all 4 machines are already flashed and ready to boot):

  • Deploy multiple NemoClaw instances on the cluster
  • Make them available to select members on our floor for early testing

public-beta-1.0.0 — Open to the AI Floor

Scale the cluster to all 40+ machines and open access to all AI floor citizens.

Network Infrastructure

The Omarchy NUC currently serves as the DHCP server for the cluster. This is a single point of failure and not suitable for 40+ nodes. Replace it with a dedicated network appliance:

  • Deploy a Ubiquiti gateway/router (e.g., UniFi Dream Machine or EdgeRouter) to handle DHCP, DNS, and network management
  • Migrate all DHCP reservations (MAC-to-IP mappings) from the Omarchy box to the new appliance
  • The Omarchy NUC can then be repurposed as a general-purpose launcher box or retired

Access Model

  • Bring your own token — each user provides their own AI provider API key
  • We provide a small pool of courtesy API tokens for users who don't have their own
  • Secrets management: build a lightweight app using SOPS + GPG so users can safely upload their tokens, encrypted at rest and inaccessible to other users

Hardware Needs

Procurement

Ray handles hardware procurement. Ask Vidya for his recommendations on parts and vendors.

Item Qty Why Link
Indoor Cat5e cables (flat, flexible — not outdoor-rated rigid cables) 40+ Replace the current stiff outdoor cables Cable Matters 10-pack on Amazon
Pass-through RJ45 connectors 1+ pack Clean, reliable terminations for custom cable lengths Cable Matters pass-through connectors on Amazon
RJ45 strain reliefs 1+ pack Protect connector joints from cable stress Cable Matters strain reliefs on Amazon
NVMe USB enclosures 3–10 Parallelize flashing across multiple drives at once UGREEN NVMe enclosure on Amazon

With this hardware we can bulk-flash the remaining 40+ machines. Help with crimping the Ethernet cables would be appreciated.


Definition of Done

alpha-0.1.0

  • [ ] 4 nodes powered and online (1 CP + 3 workers)
  • [ ] Hardware inventory complete (CPU, RAM, NVMe for every machine)
  • [ ] All machines physically labeled
  • [ ] NemoClaw deployed and accessible to select floor members

public-beta-1.0.0

  • [ ] 40+ machines running and joined to the cluster
  • [ ] BYOT (bring your own token) access model live
  • [ ] SOPS + GPG secrets management app deployed
  • [ ] All AI floor citizens have access
  • [ ] Indoor cabling and clean network terminations in place
  • [ ] DHCP/DNS migrated from Omarchy NUC to dedicated Ubiquiti gateway