Architecture & Financial Trade-Off: Hetzner 2-Node GPU Cloud vs. Local $1,500 SFF CapEx

Date: August 19, 2026

Document ID: SEN-4-SUPPLEMENT

Topic: 2-Node Dedicated Infrastructure (Web/Sales Edge + AI/Compute GPU Node) vs. Local Hardware CapEx

1. Financial Reality: CapEx vs. OpEx Analysis

Spending $1,500 upfront on physical mini PC hardware has hidden costs:

• Hardware depreciation (3-year lifecycle).
• Residential power & heat (24/7 continuous draw at California tier pricing ~$0.35–$0.45/kWh ≈ $15–$25/mo).
• Residential ISP upload bottleneck (affects remote webhook latency and large artifact sync).
• Single point of physical failure in lab.

Hetzner Equivalent Runway ($1,500 Capital Budget):

• Hetzner Cloud VPS (Web/Sales Edge - e.g. CPX31 / CAX21): ~$7.00 – $15.00/mo.
• Hetzner GPU Server / Dedicated GPU Node: ~$50.00 – $110.00/mo (e.g. RTX 4000 Ada / RTX A4000 / Server Auction GPU).
• Runway: $1,500 capital provides 14 to 24 months of enterprise datacenter compute with 1Gbps/10Gbps unmetered bandwidth, 99.9% uptime, automated snapshots, and zero physical maintenance.

2. Proposed 2-Node Sovereign Architecture

                                  [ Internet / Clients ]
                                             │
                       ┌─────────────────────┴─────────────────────┐
                       │                                           │
                       ▼                                           ▼
          [ Polar.sh / Webhooks ]                     [ Public Research Portal ]
                       │                                           │
                       └─────────────────────┬─────────────────────┘
                                             │
                                             ▼
                     ┌───────────────────────────────────────────────┐
                     │ NODE 1: Web & Sales Edge (Hetzner VPS)        │
                     │ - Caddy Reverse Proxy & SSL                   │
                     │ - Public Briefings & Research Hub             │
                     │ - Inbound Webhook Receiver (Polar / Stripe)   │
                     │ - n8n Workflow Automation Engine              │
                     │ - Public DMZ Security Barrier                 │
                     └───────────────────────┬───────────────────────┘
                                             │ (Private Tailscale / WireGuard)
                                             │ (Encrypted Sovereign Mesh)
                                             ▼
                     ┌───────────────────────────────────────────────┐
                     │ NODE 2: AI & Compute Worker (Hetzner GPU Node)│
                     │ - Dedicated NVIDIA GPU (16GB–24GB VRAM)       │
                     │ - Ollama / vLLM High-Throughput Server        │
                     │ - ComfyUI Serverless Rendering Engine         │
                     │ - Heavy Batch Agentic Trajectories            │
                     │ - Isolated from Public Internet               │
                     └───────────────────────────────────────────────┘

3. Key Advantages of the 2-Node Hetzner Strategy

1. Clean Separation of Concerns (DMZ vs. Compute):

- Node 1 (Web & Sales): Hardened public-facing server. If compromised, it contains no private LLM weights or internal sovereign databases.

- Node 2 (Compute/AI): Completely private backend reachable only via Tailscale mesh. Handles raw token generation, ComfyUI rendering, and background agents.

2. Elastic Scaling & Capital Preservation:

- Preserves $1,500 cash reserves for business formation and marketing ramp.

- Allows upgrading/downgrading GPU tiers as model sizes evolve without getting stuck with depreciated local hardware.

3. Residential Lab Offloading:

- Orchestrator Node and Local Inference Node in the lab stay lightweight, cool, and dedicated to interactive development and local testing.