Field notes / № 1

The real electricity cost of a 70B coding assistant in 2026

You're pricing a rig that can run a 70B model, and someone in a comment section has warned you it will "melt your power bill." Here is the actual math: measured watts instead of spec sheets, honest idle hours, and the same rig priced in eight states. The bill is real. It is also, almost always, the least important number on the page.

All dollar figures computed by the open engine · rates verified 2026-07-05

TDP is marketing. Your wall outlet doesn't read press releases.

An RTX 3090 has a 350 W TDP. Two of them plus a CPU, fans and a power supply "should" pull 850 W or more, and that's the number people plug into scare math.

It's the wrong number. TDP is a thermal design ceiling, not a duty cycle. Generating tokens from a large language model is memory-bandwidth-bound work — the GPU spends much of its time waiting on VRAM, not burning through its power limit. In practice, sustained inference holds a card at roughly 70–80% of TDP. Measured at the wall, a dual-3090 box generating tokens draws about 750 W whole-system — GPUs, CPU, fans, PSU losses, everything. That's the figure our hardware presets use, and it matches what people report with actual meters (see appliancerunningcost.com's local-LLM measurements).

The gap matters more than it looks. Every scare-math article you've read priced the rig at nameplate watts, 24 hours a day. That inflates the annual bill by hundreds of dollars and — conveniently, for sites selling cloud subscriptions — makes ownership look worse than it is. The inverse crowd assumes the GPUs sip power because "it's just inference." Both are guessing.

If you want the truth for your own build, a $20 wall meter settles it in an afternoon. It is the only instrument in this hobby that never exaggerates.

The idle-hours tax nobody counts

Here's the line item scare math skips and rosy math also skips, for opposite reasons: the machine is on far more hours than it is thinking.

Our dual-3090 preset idles at about 90 W whole-system. Suppose you code with the model three hours a day and, like a normal person, never power the box down. That's 21 idle hours a day:

0.09 kW × 21 h × 30 days = 56.7 kWh a month of doing nothing — $10.68 at the national average rate, $26.43 in Hawaii. Every month. For zero tokens.

Compare a disciplined duty cycle — the one our coding preset assumes: 3 hours of inference, 5 hours powered-but-idle, asleep or off the rest of the day. Same rig, same work, and the monthly bill drops from $23.39 to $15.25 at the national rate. Leaving the box on around the clock adds about $8 a month and pushes break-even from month 18.8 to month 20.2. Not ruinous. Not nothing. Suspend-to-RAM is free.

This is also why "electricity cost per token" figures floating around forums disagree by an order of magnitude: some people divide by tokens generated during active hours, others amortize the idle tax across everything. Neither is wrong. They're just answering different questions, which is why the engine asks for both numbers separately.

The kWh math, worked, no hand-waving

Here is every step for the dual-3090 rig running a 70B model at Q4, on the coding duty cycle, at the 2026 U.S. average residential rate of $0.1883/kWh:

Active: 750 W = 0.75 kW × 3 h/day = 2.25 kWh/day
Idle: 90 W = 0.09 kW × 5 h/day = 0.45 kWh/day
Daily total: 2.7 kWh × 30 days = 81 kWh/month
Monthly cost: 81 × $0.1883 = $15.25/month

Now the same duty cycle on a Mac Studio M4 Max with 64 GB of unified memory — the other way people run 70B at home. It draws about 90 W generating tokens and 10 W idle:

(0.09 × 3) + (0.01 × 5) = 0.32 kWh/day → 9.6 kWh/month → $1.81/month.

The Mac's electricity bill is a rounding error — but it costs about $750 more up front and generates tokens at roughly 60% of the dual-3090's speed. At the national rate, the GPUs win on total monthly cost ($77.75 vs $85.14, amortization included). Only at Hawaii-class rates does the efficiency premium pay for itself: there, the Mac's total drops below the GPU rig's ($87.81 vs $100.26). Geography, not fandom, settles that argument.

Same rig, eight states

Electricity rates in the U.S. span 4× — Utah's $0.115/kWh to Hawaii's $0.4662 (EIA residential averages, cross-checked against electricchoice.com, verified 2026-07-05). Here is the identical dual-3090 rig, identical 15M-token coding month, with Claude Sonnet 4.6 ($3/$15 per million tokens) as the honest cloud baseline:

Dual-3090 preset: $2,250 build, 3-year amortization, 750 W active / 90 W idle, 3 h active + 5 h idle daily, 7.5M input + 7.5M output tokens/month vs Claude Sonnet 4.6. Full formulas on the methodology page.

Read the spread. Crossing the entire country — cheapest grid to most expensive — moves the electricity bill by $28 a month and the break-even point by about five months. It nudges Utah and Washington over the OWN line and drags Hawaii deeper into HYBRID territory, but it never once turns a good idea into a bad one, or the reverse. Keep that five-month swing in mind. Something else on this page swings the answer by forever.

The power-limit trick (do it anyway)

You can cut the active number with one line:

sudo nvidia-smi -pl 260

That caps each 3090 at 260 W — about 75% of its 350 W TDP. Because token generation is bandwidth-bound, the throughput cost is typically a few percent; the card was rarely using its full power budget productively anyway. Whole-system draw drops from roughly 750 W to 600 W under load.

Run through the engine: monthly usage falls from 81 to 67.5 kWh, saving $2.54/month at the national rate, $4.34 in California, $6.29 in Hawaii. Break-even improves from month 18.8 to 18.4.

We'll be honest, because that's the whole point of this site: those numbers will not change your verdict. Do it anyway. It's free, it dumps 150 W less heat into your office, the fans stay quieter, and it's gentler on used cards of unknown provenance. The power-limit trick is a quality-of-life feature that happens to pay a small dividend.

The honest conclusion: electricity doesn't pick the verdict

Now the part the "your power bill will melt" crowd and the "electricity is basically free" crowd both miss. Hold the rig, the state, and the usage constant — U.S. average rate, 15M tokens a month — and change only the cloud model you'd otherwise be paying for:

Against DeepSeek V4 Flash, the cloud bill ($3.15/mo) is less than the rig's electricity alone ($15.25/mo). No state in the union fixes that — not even Utah. There is no break-even, ever; you'd buy the rig for privacy and sovereignty, which are fine reasons that deserve their own ledger, not creative accounting.

So: measure your watts, count your idle hours, cap your power limit — and then spend your real deliberation on the comparison that actually decides things. We've written that one up too: Local vs cloud: when it actually saves money. And if your budget is one used GPU instead of two, see Does a used 3090 still pay for itself?

Or skip the reading and run your own numbers — your rate, your hours, your honest baseline. It takes about a minute and it will happily tell you not to buy anything.

Run your break-even test

Rates are EIA residential averages, verified 2026-07-05; your utility bill wins every argument. Hardware draws are realistic whole-system figures, not TDP. Every calculation above is reproducible in the open engine — no secret math, no thumb on the scale.