π Platform & InfrastructureΒ· 8 min read
Facility Planning
Site selection, power procurement, construction timelines, phased buildout, and sustainability considerations for AI data centers.
facility-planningsite-selectionconstructionpower-procurementsustainability
Building or leasing a facility for AI infrastructure is a multi-year, multi-million-dollar decision. The biggest constraint is usually power availability β not space, not cooling, not networking. Getting site selection wrong means 12-24 months of delay while you wait for utility upgrades.
1. Site Selection Criteria
| Factor | Weight | Key questions |
|---|---|---|
| Power availability | Critical | Can utility deliver MW-scale power? How fast? |
| Power cost | High | 0.03 to $0.15+) |
| Fiber connectivity | High | Multiple carriers, low-latency paths to users |
| Climate | Medium | Free cooling hours, natural disaster risk |
| Water availability | Medium | For cooling towers (if applicable) |
| Land cost & permitting | Medium | Zoning, environmental review timelines |
| Carbon intensity | Growing | Grid mix (hydro/nuclear vs fossil) |
| Talent pool | Medium | Local ops and engineering staff |
| Tax incentives | Medium | State/local data center tax breaks |
Top Regions for AI Data Centers
| Region | Power cost | Carbon | Climate | Notes |
|---|---|---|---|---|
| Northern Virginia | $0.06-0.09 | Medium | Moderate | Largest DC market, constrained power |
| Texas (ERCOT) | $0.04-0.07 | Medium-high | Hot | Cheap power, renewable growth |
| Quebec | $0.03-0.05 | Very low (hydro) | Cold | Excellent for free cooling |
| Iceland | $0.03-0.05 | Very low (geothermal) | Cold | Cheap power, remote location |
| Nordics (Sweden, Finland) | $0.04-0.07 | Low | Cold | Good power, EU data sovereignty |
| Singapore | $0.10-0.15 | High | Hot/humid | APAC hub, moratorium on new DCs |
2. Power Procurement
The single biggest constraint and longest lead time:
Power procurement timeline:
Month 0-3: Site survey, utility discussions
Month 3-6: Power contract negotiation
Month 6-12: Utility infrastructure build (substation, feeders)
Month 12-18: Facility construction (if new build)
Month 18-24: Equipment install, commissioning
New facility from scratch: 18-30 months
Colo expansion (existing utility): 3-6 months
Cloud: minutes (but at 5-10Γ the cost)
Power Contract Types
| Type | Mechanism | Best for |
|---|---|---|
| Utility direct | Contract with local utility | New facility, large scale |
| PPA (Power Purchase Agreement) | Buy from specific generator (solar/wind) | Renewable targets |
| Behind-the-meter | On-site generation (solar, fuel cells) | Supplemental, backup |
| Wholesale market | Buy on spot market | Flexible, risk-tolerant |
3. Phased Buildout
Donβt build for year-5 capacity on day 1. Phase the deployment:
Phase 1 (Month 0-6): 256 GPUs (1 SuperPOD)
Power: 250 kW
Cooling: 2 CDUs
Network: 2 spine switches
Staff: 2 SREs (part-time)
Phase 2 (Month 6-12): 1024 GPUs (4 SuperPODs)
Power: 1 MW
Cooling: 8 CDUs
Network: 8 spine switches
Staff: 3 SREs (full-time)
Phase 3 (Month 12-24): 4096 GPUs
Power: 4 MW
Cooling: dedicated CDU room
Network: multi-tier fat-tree
Staff: 8 SREs + 2 network engineers
Each phase:
βββ Pre-order hardware 4-6 months before needed
βββ Pre-provision power and cooling for NEXT phase
βββ Validate at current scale before expanding
4. Colo vs Build vs Cloud Decision
| Factor | Cloud | Colocation | Own facility |
|---|---|---|---|
| Time to deploy | Minutes | 1-6 months | 12-30 months |
| Upfront cost | $0 | $1-5M | $10-100M+ |
| $/GPU-hour (H100) | $5-12 | $3-5 | $2-4 (amortized) |
| Flexibility to scale | Instant | Months | Months-years |
| Flexibility to shrink | Instant | Contract term | Stuck |
| Control | Low | Medium | Full |
| Min practical scale | 1 GPU | 100 GPUs | 500+ GPUs |
Decision tree:
βββ <100 GPUs β Cloud
βββ 100-500 GPUs β Cloud or Colo
βββ 500-5000 GPUs β Colo (sweet spot)
βββ >5000 GPUs, >70% utilization β Own facility
βββ Uncertain scale β Cloud (pay for flexibility)
5. Construction Checklist
site_selection:
- Power capacity assessment (utility letter of commitment)
- Fiber connectivity audit (at least 2 diverse paths)
- Environmental assessment (flood risk, seismic)
- Zoning verification (data center permitted use)
- Tax incentive evaluation
design:
- Power distribution design (N+1 or 2N UPS, generator sizing)
- Cooling design (DLC from day 1 for AI)
- Structured cabling plan (IB cable trays, fiber routing)
- Physical security (fencing, access control, cameras)
- Fire suppression (clean agent, not water, near servers)
construction:
- General contractor with DC experience
- Electrical contractor (medium voltage certified)
- Mechanical contractor (DLC experienced)
- Network cabling contractor (IB certified)
- Commissioning agent (independent verification)
commissioning:
- Power load testing (full rack power draw)
- Cooling capacity testing (worst-case heat load)
- Network burn-in (IB link stability under load)
- UPS transfer testing (simulate utility failure)
- Generator runtime testing (72-hour continuous)
6. Sustainability Planning
| Strategy | When to implement | Impact |
|---|---|---|
| DLC from day 1 | Design phase | PUE 1.1 vs 1.4 |
| Renewable PPA | Power procurement | 50-100% carbon offset |
| Heat reuse | Design phase (if viable) | District heating, agriculture |
| Waste heat recovery | Retrofit possible | Additional revenue stream |
| Free cooling optimization | Site selection | Reduces mechanical cooling hours |
| Water-free cooling (DLC/immersion) | Design phase | Eliminates water consumption |
Self-Check
- Evaluate a data center site on power, connectivity, climate, and cost
- Create a power procurement timeline from site survey to commissioning
- Design a phased buildout plan from 256 to 4096 GPUs
- Compare cloud vs colo vs own facility at different GPU scales
- Build a construction checklist covering power, cooling, and network
References
- Uptime Institute: Data Center Site Selection
- JLL: Data Center Market Outlook
- For power delivery details, see Power Delivery
- For cooling design, see Cooling Systems
- For network topology, see Physical Topology