# Modular Data Centers MOC ## Cooling Economics - [[DX vs Chilled Water Crossover]] - [[Condenser Cable Economics]] - [[PUE at Altitude — ASHRAE Derate]] ## High-Density Compute - [[Density Tiering Principle]] - [[Delta AI CDC Container]] - [[Rear Door Heat Exchangers]] ## GPU Vendor Dynamics - [[Nvidia Gatekeeper Model]] - [[Tenstorrent — Alternative Silicon Path]] ## Power Architecture - [[48V DC and OCP Bus Bar]] - [[PV as IT Load Headroom]] - [[UPS Autonomy Philosophy — Generator Bridge]] ## Factory & Deployment - [[Factory Pre-Testing and CFD Validation]] - [[Transport Constraints as Design Inputs]] - [[FK-5-1-12 Fire Suppression]] ## Monitoring - [[Delta DCIM — U-Level Granularity]] --- # Directional Arrows of Progress **Cooling Architecture Selection** `Air → DX (dedicated condensers) → Chilled Water → Hybrid` DX wins below ~1MW on cost. The crossover isn't about performance — it's about economics. Long condenser cable runs ($2K each × 24 cables) still cost far less than the €200-300K chilled water upgrade. **Rack Density Separation** `One-size-fits-all → Tiered envelopes (colo / inference / training)` Don't try to cool 160kW racks in the same thermal envelope as 20kW colo racks. Separate them physically — different containers, different cooling, different operational models. The Delta approach: standard MDC for ≤20kW, AI CDC bolt-on for 30-40kW, and a separate conversation entirely for GB300 at 160kW. **GPU Access Model** `Open hardware → Certified installers → Gatekeeper lock-in` Nvidia's B200/B300/GB300 requires authorized installation partners (Asus, Super Micro, Dell), 6-month training, blanket refusal otherwise. This is hardening, not softening. Alternative silicon (Tenstorrent at ~12.5kW/chip, no gatekeeper) becomes strategically important even at lower performance. **Power Distribution** `AC PSU per server → Rack-level 48V DC → OCP bus bar (eliminate CRPs)` The OCP bus bar architecture converts 48V→12V at the rack, removing individual server power supplies. Fewer components, fewer failure points, better efficiency. **Monitoring Resolution** `Rack-level → Per-server IPMI → U-level DCIM` Delta's DCIM platform goes to individual U position with IPMI via ASRock AST2600. Replaces Zabbix-style polling with purpose-built DC infrastructure management. --- # Key Principles 1. **Cooling is the cost lever below 1MW** — not compute, not power. The DX vs chilled water decision cascades into cable routing, condenser placement, redundancy topology, and total cost. 2. **Transport constraints are design inputs, not afterthoughts** — 3.5m width, 20m length max without escort, ~1 month sea freight. These shape the physical architecture from day one. 3. **Factory beats field** — pre-fabrication with full thermal simulation (CFD) and on-site condition replication before shipping. 6 months build + 1 month ship vs 12-18 months traditional. 4. **Solar is engineering headroom, not a sustainability checkbox** — PV offsets 2kW/rack, creating 22kW usable IT load from a 20kW grid allocation. It's a power budget input. 5. **Minimize stored energy, maximize mechanical generation** — 5-minute UPS autonomy bridges to generators. Lead-acid, 10-year cycle, 2N redundant. Extended battery autonomy is expensive and unnecessary when generator start is reliable. 6. **Vendor lock-in is structural** — Nvidia's gatekeeper model means your choice of GPU dictates your choice of installer, timeline, and cost. Plan around it or pick different silicon. ### Related - [[Data Center MoC]] - [[Navon MoC]] - [[Modular Data Center Design Principles]] - [[Open Source Hyperscaler MoC]]