Operator total cost of ownership
3-year amortized cost of operating local AI at three realistic scales. Includes hardware, electricity (US average $0.16/kWh unless your region differs significantly), operator hours at $50/h, and downtime productivity loss. Compared honestly to the equivalent cloud spend.
We show ranges, not precise numbers. Hardware prices fluctuate and electricity rates vary 5x across regions. Use this as a framework, not a quote.
Hobby — single 4090, single user
RTX 4090 + Ryzen 9 + 64 GB RAM
- Capex
- $2.5-3.5k (build cost 2026)
- Capex amortized (3 yr)
- $70-100/mo
- Power draw
- 350-450 W under load, 30-50 W idle
- Electricity
- $15-30/mo (US average $0.16/kWh, 4 h/day load)
- Operator time
- ~3 h/mo
- Operator cost
- $150/mo @ $50/h
- Downtime assumption
- 1 weekend/year on a driver issue (≈$0 if hobby)
- Total annual cost
- $2,800-3,400
- Cloud equivalent
- $30-150/mo on OpenRouter at hobby usage
Serious operator — RTX 6000 Ada or 2× 4090
RTX 6000 Ada (48 GB) or dual 4090
- Capex
- $8-12k
- Capex amortized (3 yr)
- $220-330/mo
- Power draw
- 500-700 W under load
- Electricity
- $40-80/mo
- Operator time
- ~6 h/mo
- Operator cost
- $300/mo @ $50/h
- Downtime assumption
- 2-3 weekends/year (≈$1k of operator time/yr)
- Total annual cost
- $7,000-9,500
- Cloud equivalent
- $300-1,500/mo for steady production usage
Small team — 4× 4090 homelab, 5 concurrent users
4× RTX 4090 + Threadripper + UPS + 10G LAN
- Capex
- $15-25k
- Capex amortized (3 yr)
- $420-700/mo
- Power draw
- 1.5-2 kW under load (dedicated 20A circuit)
- Electricity
- $120-200/mo
- Operator time
- ~12 h/mo
- Operator cost
- $600/mo @ $50/h
- Downtime assumption
- 5-8 incidents/year (≈$2-3k of team productivity)
- Total annual cost
- $15,000-22,000
- Cloud equivalent
- $3,000-8,000/mo for team usage on Together/Fireworks
What this calculation elides
Resale value. A 4090 three years old still has a market. Cloud spend is gone the month it's billed.
Time-to-first-result. Cloud APIs work in minutes. Local hardware takes weeks of configuration before the first useful output.
Quality gap on hardest tasks. For frontier-tier reasoning, local + open-source models are still behind. The cloud frontier API delta isn't free, but it's real.
Variance in cloud bills. Bursty workloads spike. We assumed steady-state usage; if yours is bursty, cloud cost is 1.5-3x higher than the steady-state numbers above.