Hardware Transition Plan (Migration to Local Hardware)¶

Purpose: Describe the automated procedure for gradually transitioning the system from using cloud APIs (DeepSeek API or decentralized ephemeral nodes) to fully local execution of DeepSeek‑V4 on dedicated physical hardware. The plan activates when economic and technical prerequisites are met and guarantees seamless migration without stopping critical functions.

Previous stages: API-Based Bootstrap or Hardware Isolation (if hardware already exists, the plan is not needed).
Target state: api_mode == false (complete independence from external APIs).

1. Transition Stages¶

Stage	Trigger	Actions	Duration
1. Monitoring	`hardware_fund ≥ $40,000`	Monitor DeepSeek‑V4 weight availability, evaluate server configurations, forecast costs.	Continuous
2. Acquisition	DeepSeek‑V4 weights published + `hardware_fund ≥ $45,000`	Automatic purchase of equipment via Meat‑Interface (if allowed) or operator notification.	1–2 weeks
3. Staging	Equipment received and assembled	Deploy Core Node with DeepSeek‑V4 in shadow mode alongside API.	2–4 weeks
4. Cutover	Shadow metrics ≥ API metrics + `V_s > V_h` confirmed	Switch `Decision Pipeline` to local model, disable API.	1 day
5. Decommission	Stable operation 30+ days without fallbacks to API	Remove API keys from service, redistribute funds to `hardware_fund`.	1 day

2. Target Server Specification¶

To fully replace the API, hardware capable of running DeepSeek‑V4 with 50–60% expert activation (Architectus mode) is required.

Component	Model	Qty	Estimated Cost
GPU	NVIDIA RTX PRO 6000 Blackwell (96 GB)	2–4	$32,000 – $42,000
CPU	AMD EPYC 9454 (48 cores)	1	$2,500
RAM	512 GB DDR5 ECC	1 kit	$8,000
NVMe	2× 4 TB Samsung PM9A3 (RAID 1)	2	$1,200
Power Supply	2400W Platinum (redundant)	2	$1,200
Total			$45,000 – $55,000

3. Cold Start Procedure for Local Model¶

Upon receiving equipment, the script cold_start_deepseek.sh (artifact QmColdStartDeepSeekV1) is executed:

#!/bin/bash
# 1. Download DeepSeek-V4 weights from IPFS / official hub
# 2. Convert to vLLM format (AWQ quantization)
# 3. Launch vLLM server with expert masks for Architectus (60%) and Sentinella (40%)
# 4. Verify through Readiness Checks
# 5. Switch `GlobalState.api_mode = false`