README
Project NEXUS
A private, bare-metal AI/HPC cloud built on two compute planes: a small dedicated Core, and several hundred borrowed DIU classroom PCs.
Start here
| Document | What it answers |
|---|---|
ULTIMATE-PLAN.md | What and why. The laws, the physics, the hardware model, the risk register, the acceptance gates, the roadmap. |
CAMPUS-FABRIC.md | Plane B — the harvest fabric. How borrowed lab PCs become efficient, consented, measured capacity. Required reading before any *B phase. |
ARCHITECTURE.md | How the pieces fit. The eleven-layer stack, component catalog, ADRs, and the Plane A/B layer deltas (§0.4). |
phases/README.md | How to build it. The execution protocol, the phase index, and the Beachhead path. |
The two planes
CONTROL PLANE (3 nodes: etcd · Kueue · Argo CD · Availability Oracle)
│
┌──────────────┴───────────────┐
▼ ▼
PLANE A — CORE PLANE B — CAMPUS HARVEST
4–8 dedicated racked nodes 150–400 existing DIU classroom/lab PCs
25–100 GbE RoCEv2, Ceph Campus 1 GbE, borrowed outside class hours
Multi-node training, storage, Sweeps, batch inference, preprocessing, CI,
all state, tightly-coupled single-node training, rendering
Cost: procurement Cost: electricity only ($0 capex)
The thesis: the campus already contains more idle compute than the department could afford to buy. The engineering problem is not acquiring capacity — it is harvesting it without waste and without ever inconveniencing a human.
Where to begin building
The Beachhead path (phases/README.md §8) reaches real users in ~8 weeks with zero hardware purchase:
00 → 01 → 01B → 04B → 03B → 06 → 07 → 08B → 09 → 12 → 13 → 15 → 14 → 14B → 25B → 30 → 31B
🚧 Phase 04B is a hard barrier. No campus machine is woken, netbooted, or enrolled before that lab's participation agreement is signed and recorded. See
phases/README.mdRule C1.
Repository layout
| Path | Contents |
|---|---|
phases/ | Numbered work orders (00–56 Core, *B Campus) |
evidence/ | Per-phase proof: preflight, acceptance, handoff, deviations |
inventory/ | Machine-readable source of truth — nodes, racks, power, network, campus/ |
clusters/nexus-prod/ | GitOps-managed cluster state |
docs/ | Facility, security, and campus/ documentation |
policies/ | Kyverno policies and campus/ consent records |
tools/ | Validation, discovery, and campus/ operational tooling |
benchmarks/ | Suites, runners, baselines, reports |
runbooks/ | Operator procedures, one per alert |