BLK · TCO CALCULATOR

What does it actually cost?

Upfront hardware + electricity + amortization vs cloud API equivalent. Every assumption visible. Every formula sourced. No hidden multipliers — the breakdown is the source of truth.

INPUTS

Default $0.30/M is the conservative 8B-class blended cost across Together / Groq / Anthropic / OpenAI in May 2026.

§ Companion power (optional — desktop rigs)

Typical reference: CPU 65W (idle/inference) – 150W (Threadripper under load) · 27" 4K monitor ~30W · dual-monitor ~60W · audio interface + fans + lighting ~15-30W. Companion components run continuously during active hours (no utilization multiplier applied).

Reset to defaults
HEADLINE · 3-YEAR TCO
Total ownership
$2,088
$1,899 hw + $189 elec
Cost per million tokens
$1.824
at 121.0 tok/s, 60% util
Cloud equivalent cost
$344
at $0.3/M tokens × 1.1B tokens
PLAIN-ENGLISH VERDICT

At your usage pattern, running NVIDIA GeForce RTX 4090 locally costs $1,745 more than the cloud API equivalent over 3 years. Cloud wins unless privacy / latency / offline matter to you.

Break-even tokens/month: 193.4M — below this monthly volume, cloud is cheaper; above, local pulls ahead.

ComponentAssumptionValue
Upfront hardwarecurrent street price$1,899
Average load450W TDP × 60% utilization270 W
kWh / day270W × 4 h/day1.08 kWh
Total kWh (3 yrs)1.08 × 365.25 × 31183 kWh
Electricity cost1183 kWh × $0.16/kWh$189
Representative tok/sextrapolated from bandwidth (8B Q4 class)121.0 tok/s
Tokens produced121.0 tok/s × 4h × 60% × 3yr1.1B
ASSUMPTIONS · LIMITS
What we model
  • Upfront hardware price (street > MSRP > unknown).
  • Electricity over the full ownership horizon at your local rate.
  • Token throughput from measured or bandwidth-extrapolated tok/s.
  • Cloud equivalent cost at a representative $/M-tokens rate.
What we don't model (yet)
  • Cooling/AC overhead (typically +10–30% on top of TDP).
  • Resale value at end of amortization horizon.
  • Income-tax implications (deductible business expense).
  • Privacy / latency / offline benefits — those are non-monetary but real reasons local can be the right answer even when cloud is cheaper.

Full math: /guides/methodology · Want measured tok/s on this rig? Submit a benchmark.