Local AI cost calculations: power, usage and payback
A workstation saves money when it replaces work that would otherwise incur a bill. Its ability to generate tokens around the clock is not enough. Those tokens need to be usable, and the alternative cost needs to be one the owner would actually pay.
Research checked 31 October 2026. Compiled from published tests. No hardware benchmarks were run for this article; measurements belong to the credited testers.
Eugene Petrenko’s February 9, 2026 analysis makes utilization central to the calculation. His model combines published benchmarks with assumptions about cloud equivalents and hardware prices. It explicitly excludes maintenance and depreciation, and its main calculation excludes electricity. That makes it a scenario model, not a record of savings achieved by a typical owner. Its Q1 prices should not be used as October quotations. Petrenko’s analysis and calculator.
Power measurements need a boundary
Puget Systems measured 480W across two R9700 GPUs while serving Qwen3.6-27B in stock vLLM FP8 at eight concurrent users, with aggregate throughput of 156.21 tokens per second. The article adds an estimated 300W of system overhead for its cost model. That overhead is not a wall-meter measurement of the whole machine. Dustin Moore, Puget Systems.
Using those inputs gives about 1.39kWh per million output tokens: 0.78kW multiplied by the time required to generate one million tokens at 156.21 per second. At an illustrative tariff of $0.15/kWh, that is about $0.21 in electricity. This derived figure excludes purchase cost, idle periods, cooling outside the workstation, maintenance and unsuccessful work. It also assumes the measured throughput is sustained.
For an always-on machine, idle consumption belongs in the budget. Jeff Geerling measured roughly 30W idle on the Dell GB10 system he reviewed in December 2025. If sustained all year, that alone would be approximately 263kWh. That calculation is an extrapolation from his reading, not a measured annual bill. Geerling’s Dell review.
Use avoided spending, then subtract ongoing costs
Consider a hypothetical $5,000 purchase:
| Monthly spending actually avoided | Monthly local running costs assumed | Simple payback |
|---|---|---|
| Avoids $100/month | $25 | About 67 months |
| Avoids $300/month | $25 | About 18 months |
| Avoids $500/month | $25 | About 11 months |
These are arithmetic examples, not observed owner results. Simple payback divides purchase cost by net monthly savings and ignores financing, resale and changes in prices or workload. If the model cannot complete the work acceptably, the supposed avoided spending may never materialize.
Keeping an existing subscription while buying a local workstation may add capacity, but its savings must come from other bills actually reduced. An existing suitable GPU also has different incremental economics from a new complete workstation.
Privacy, offline access and control over model versions can justify owning hardware independently of payback. The financial case is strongest when the owner can identify recurring work the local system handles successfully and a corresponding expense that disappears.
Read the series
- Mac Studio and RTX cards: reading published LLM tests
- DGX Spark 64GB: what published tests can tell us
- AMD local LLM tests: the configuration behind the score
- Why local LLM benchmark charts disagree
- Local AI cost calculations: power, usage and payback (this article)
For memory planning, use the model-fit guide. Browse all Local AI notes.