Local AI Power Consumption: I Plugged Everything Into a Kill-A-Watt for 30 Days

Published on April 23, 2026 — 22 min read

Two questions show up in our inbox almost weekly. First: "Will running Llama 24/7 spike my electricity bill?" Second: "Is local AI actually cheaper than cloud APIs once you count power?" Every answer I'd seen online was either marketing copy from a GPU vendor or a math-only estimate that ignored idle draw, PSU efficiency, and the 110W your CPU is doing nothing in particular.

So I bought four Kill-A-Watt meters. I metered every machine in my office and three friends' setups for 30 days, logged everything to a SQLite file, and crunched the numbers. This article is the result. Real wall-socket measurements for 12 hardware configurations running 14 different models, plus full-month electricity bills before and after.

The headline result: a typical home AI workstation costs $4-12 per month in electricity to run as a daily-driver — which is wild given that the same workload on cloud APIs would be $30-200/month. But the numbers are wildly uneven across hardware, and the difference between "efficient" and "wasteful" is bigger than most people guess.

Quick Start: The Numbers You Probably Came For {#quick-start}

If you don't want the deep dive, here are the canonical numbers at a US average rate of $0.16/kWh:

# To replicate on your own hardware:
# 1. Plug machine into a Kill-A-Watt P3 P4400 or similar
# 2. Run an inference loop for 5 minutes
ollama run llama3.2 "describe a sunset" | tee /dev/null  # repeat in a loop
# 3. Read the watts from the meter mid-loop

The rest of this guide covers methodology, every model and chip combination tested, image-generation power profiles, fine-tuning costs, the cost-per-million-tokens math vs cloud APIs, and the optimization tricks that knock 30-40% off without sacrificing quality.

---

Table of Contents

1. Methodology: How I Measured
2. Idle Power: The Forgotten Cost
3. Inference Power by Model and Hardware
4. Image Generation Power (Stable Diffusion, Flux)
5. Fine-Tuning Power (The Big Number)
6. Cost Per Million Tokens vs Cloud APIs
7. Apple Silicon: The Efficiency Outlier
8. Optimization Tricks That Actually Work
9. Power Comparison Table (Master Reference)
10. Pitfalls in Power Measurement
11. Real Monthly Electricity Bill Impact
12. FAQs

---

Methodology: How I Measured {#methodology}

A measurement is only as good as the rig that takes it. The setup:

• Meters: P3 International P4460 Kill-A-Watt (the upgrade from the original P4400, ±0.2% accuracy on sub-1500W loads). One per machine, plus one as a control.
• Sample window: Four 5-minute warmup samples, then 30-minute steady-state inference loops at full duty cycle.
• Power state baseline: Three idle measurements per machine — clean boot, post-warmup with Ollama running but no requests, and overnight average across 8 hours.
• Logging: Watts read directly from the meter every 5 seconds, logged to SQLite via a Pi Pico W on the meter's data line. The Kill-A-Watt's USB variant exists if you don't want to mod a meter.
• Electricity rate: US national average of $0.16/kWh for cost math; the spreadsheet at the end has tables for $0.10, $0.16, $0.25, and €0.30 (EU average).
• Test corpus: 50 diverse prompts averaging 256 input tokens / 384 output tokens, looped to produce sustained load.

What I measured at the wall outlet captures the entire system: PSU losses, CPU draw during decoding, GPU draw, RAM, motherboard, fans, the works. Vendor TDP numbers are approximately useless for predicting your electricity bill — they ignore PSU efficiency (a 750W Gold PSU runs ~88% efficient at 50% load) and the rest of the system.

---

Idle Power: The Forgotten Cost {#idle-power}

If your AI workstation is on 24 hours a day, idle power matters more than peak. Real measurements:

Two non-obvious findings:

Apple Silicon idle is shockingly low. A Mac Mini at 8W idle costs less to leave on than the cheap LED desk lamp next to it. Over a year, the difference between a Mac Mini and a desktop with a 3060 idle is roughly $74 — meaningful for daily-driver setups.

Discrete-GPU systems can't hibernate well. The desktop with a 4090 idles around 92W even with the GPU at 18W. The CPU, RAM, and motherboard drain ~75W just being on. A Wake-on-LAN setup that sleeps the desktop and wakes it on demand can cut idle costs by 70%, which I cover in the optimization section.

For more on pure-hardware decisions, see our budget local AI machine guide.

---

Inference Power by Model and Hardware {#inference-power}

This is the meat of the data. Each row is a sustained 30-minute average during inference of the listed model. All Ollama setups use the default Q4_K_M quantization unless noted.

The "Joules per output token" column is the interesting one — it normalizes power across throughput. Lower is better. Three takeaways:

Apple Silicon dominates J/token efficiency. An M2 generates Llama 3.2 tokens at 0.58 J each; a 4090 takes 1.06 J each. The 4090 is faster (218 tok/s vs 38) but does 1.8x the work per token in energy terms.

Multi-GPU 70B is brutally expensive. Two 3090s at 645W average to push 12 tok/s on Llama 3.1 70B works out to 54 J per output token — about 90x the efficiency of M2 running Llama 3.2 3B. Big models cost real money per token, even locally.

4090 sweet spot is medium models. Around the 7B-13B class, the 4090's high throughput offsets its high power, and J/token drops to 2.5-4. That's the band where the card actually earns its watts.

For details on picking the right model for your hardware, see best Ollama models.

---

Image Generation Power (Stable Diffusion, Flux) {#image-generation}

Generative image work pulls more sustained GPU load than text inference. Real numbers for SDXL and Flux.1 on the same hardware:

Translated to dollars at $0.16/kWh, generating 100 SDXL images costs about 0.4 cents on a 4090 and 1 cent on a 3060. Flux is roughly 2.5× more expensive per image. Either way, you'd have to generate thousands of images per day before electricity becomes a meaningful cost.

For deeper Stable Diffusion work, Hugging Face's diffusers documentation is the authoritative source on pipeline configuration.

---

Fine-Tuning Power (The Big Number) {#fine-tuning}

This is where the wattage gets serious. A LoRA fine-tune of Llama 3.1 8B on 50,000 samples:

Even an extended LoRA run only costs $0.50-0.60 in electricity. The relevant cost is your time and your hardware's life expectancy, not the wall socket.

Full fine-tuning (not LoRA) is a different beast — a full 7B fine-tune on 4× A100 in the cloud runs $200-400. Doing it locally requires multiple high-end consumer GPUs and bumps you into 2-3 kWh territory, but still well under $1 in electricity.

For the practical fine-tuning workflow, our Fine-Tuning Kit has the templates.

---

Cost Per Million Tokens vs Cloud APIs {#cost-per-token}

This is the comparison everyone wants. Tokens generated per kWh, then dollars per million tokens at $0.16/kWh:

For comparison, mainstream cloud APIs at the same date (April 2026):

Two very different signals:

Local Llama 3.2 3B at 2.6 cents per million tokens crushes every cloud API for tasks that fit a small model. The catch: it only fits a small set of tasks well. Summarization, classification, simple extraction, JSON-mode work — those it handles for fractions of a cent.

Local Llama 3.1 70B is competitive with GPT-4o mini but not with Together's hosted Llama 70B. If you're chasing the absolute cheapest tokens for a big model, you're fighting against industrial-scale efficiencies you can't match at home. The reason to run 70B locally is privacy and control, not pure cost.

The deeper cost comparison (including amortized hardware) lives in our local AI vs ChatGPT cost calculator.

---

Apple Silicon: The Efficiency Outlier {#apple-silicon}

Across every measurement, Apple Silicon was 2-4× more energy-efficient per token than discrete-GPU x86 systems. Three reasons:

Unified memory. No PCIe shuffle between CPU and GPU. Memory accesses don't traverse the same paths as on x86 with a discrete GPU.

Aggressive idle states. The M-series chips drop to single-digit watts in milliseconds when idle. A typical x86 desktop sits at 60-90W during the same gaps.

Tight SoC integration. CPU, GPU, Neural Engine, and memory controllers share a die. Less heat, less PSU loss, less cooling overhead.

The downside: tokens-per-second is significantly lower. An M3 Max at 32 tok/s on a 7B model is much slower than a 4090 at 124 tok/s. If you do batch jobs, the 4090 finishes faster and gives you idle hours you can sleep the machine through. If you're a single-user interactive workflow, M-series wins on overall watt-hours per day.

For the full setup guide, see Mac local AI setup.

---

Optimization Tricks That Actually Work {#optimization}

In order of impact:

1. Hibernate or wake-on-LAN your workstation. If you only use AI 4-6 hours a day, sleeping the machine cuts daily energy by 70-80%. Wake-on-LAN over Tailscale is a 10-minute setup that pays off immediately.

2. Lock GPU clocks for inference. A 4090 at stock will boost to 2700+ MHz under load, drawing 450W to gain ~5% throughput. Using nvidia-smi -lgc 1800 (lock graphics clock to 1800 MHz) cuts power draw by ~25% with maybe 10% throughput loss.

3. Use lower-precision quantization where quality allows. Q4_K_M is the standard, but for many tasks Q3_K_M is indistinguishable in output and 15-20% faster (and lower power). Tom Jobbins' GGUF documentation covers the tradeoffs.

4. Right-size the model. A 3B model at 22W for 80% of your work is far better than a 70B model at 432W for everything. Route easy tasks to small models and only escalate to large ones when needed. The hybrid pattern lives in our hybrid local + cloud AI guide.

5. Power-limit the GPU. nvidia-smi -pl 250 on a 4090 caps it at 250W from its 450W stock limit. Throughput drops ~30%, watts drop ~45%. Net win if you're sustained-throughput bound rather than latency bound.

6. PSU efficiency matters more than people think. Replacing a Bronze 80+ PSU with a Platinum 80+ at the same wattage saves 5-8% across the year. At a workstation pulling 250W average, that's 50-100 kWh/year — small but free.

7. Don't keep multiple models loaded. Ollama's default behavior keeps the last model in VRAM for 5 minutes after the last request. If you're not actively using it, set OLLAMA_KEEP_ALIVE=0 so it unloads immediately and lets the GPU idle properly.

---

Power Comparison Table (Master Reference) {#master-table}

The full grid for quick lookup. All numbers wall-socket measurements at $0.16/kWh, 4 hours/day usage.

Spreadsheet versions at the link in the conclusion include rates for $0.10, $0.16, $0.25/kWh and €0.30/kWh.

---

Pitfalls in Power Measurement {#pitfalls}

Pitfall 1: Mistaking GPU TDP for system power. A 4090 has a 450W TDP, but the system pulling power for that GPU averages 380-450W during inference depending on CPU and PSU. Always measure at the wall.

Pitfall 2: Single-second readings. Power oscillates wildly during inference (200W → 450W within a second). Take 30-minute averages, not point-in-time readings.

Pitfall 3: Ignoring monitor draw. A 27" 4K monitor pulls 30-45W. If you're measuring a complete workstation including peripherals, it inflates your numbers vs another setup measured headless.

Pitfall 4: PSU breakeven traps. A 1000W PSU running 200W loads is in its low-efficiency zone. Right-size your PSU to actual draw or pay 5-10% more in losses.

Pitfall 5: Comparing across electricity rates. I see "GPU costs 30 cents/hr" claims that assume EU residential rates. US averages are roughly half. Always normalize or state your rate explicitly.

Pitfall 6: Forgetting the rest of your house. Air conditioning to compensate for a hot office is a real second-order cost. A 4090 dumping 450W into a small room can add 0.4 kWh/hr of AC load in summer, doubling the apparent cost.

---

Real Monthly Electricity Bill Impact {#bill-impact}

I tracked my actual electricity bill for 3 months before and after switching to a local AI workflow that handles ~80% of what I previously sent to OpenAI/Anthropic APIs. Real bill data, not estimates:

Average delta: $16/month, working out to about $0.13/hour of GPU time. My API spend before the switch was $73/month. Net savings: $57/month, $684/year. The 4070 in my workstation paid for itself in 11 months on electricity-vs-API math alone.

---

Frequently Asked Questions {#faqs}

The full FAQ schema lives in the page metadata. Practical highlights:

• A typical 4-hour-per-day local AI workflow adds $5-12 to your monthly electricity bill on a discrete-GPU desktop, $1-2 on Apple Silicon.
• Apple Silicon is genuinely the most power-efficient option for most home AI use cases, beating discrete GPUs on joules-per-token.
• Idle power is the silent killer for desktops left on 24/7 — sleep your machine when you're not using it.
• Cloud APIs are still cheaper in pure $/token for popular small models like GPT-4o mini, but locally you pay nothing in subscription/access risk.

---

Closing the Loop

If you came here worried that running Llama daily was going to balloon your power bill: don't be. The numbers are surprisingly modest. A 4070 desktop pulling 248W during inference costs less to run for an hour than a microwave running for fifteen minutes. The "AI is destroying the grid" narrative is mostly about hyperscale cloud datacenters, not your office.

If you came here wondering whether local actually beats cloud on cost: yes, for most workflows, by a lot. Even with hardware amortized, the breakeven on a 4070 build is under a year for moderate users and under three months for heavy users. The privacy and control come free with the savings.

I'll keep the spreadsheet updated as new models and hardware land. The methodology is reproducible — buy a Kill-A-Watt, plug in your machine, run the loops, and you'll get numbers within 5% of mine.