Skip to content
Hack Your WorldSoftware · Infrastructure · Home automation

Upgrade decisions / buy the right comparison

Should you replace an old PC to save electricity?

Only if the replacement does the same job and the measured electricity saving repays the complete purchase cost soon enough. A more efficient computer can still be a bad purchase.

I run two kinds of computer workload that make a bad comparison look tempting. Home Assistant wants a small, dependable host that remains available all day. Local Ollama inference on an RTX 5070 Ti wants a capable GPU when I ask it to work. A tiny replacement may beat the GPU machine at idle power while failing the only job that made the GPU necessary.

So I start with the work, not the age of the case. If the new machine cannot provide the same storage, acceleration, ports, reliability, or software support, the electricity calculation is comparing different products.

Compare your measured wattage and purchase cost

The arithmetic is the easy part

For two machines doing the same work, the annual electricity saving is:

(old watts − new watts) × hours per day ÷ 1,000 × price per kWh × 365

Divide the complete net purchase cost by that annual saving to get simple payback. This does not include financing, future electricity prices, or a discount rate. It is a blunt test designed to stop a purchase from being called “free” because the new box uses fewer watts.

Example: 80 W replaced by 15 W at $0.20/kWh
Daily runtime Annual saving $250 payback Five-year net saving
24 hours $113.88 2.20 years $319.40
16 hours $75.92 3.29 years $129.60
8 hours $37.96 6.59 years −$60.20
4 hours $18.98 13.17 years −$155.10

Those are labeled example inputs, not claims about a particular old desktop or mini PC. The point is how violently runtime changes the decision. An always-on service creates far more opportunity for savings than a computer used for a few evening hours.

First, prove both machines can do the same job

I write down the workload before looking at replacement hardware:

For Home Assistant, reliability and access to the required radios or USB devices may matter more than peak performance. For the local Ollama assistant, GPU memory and measured token speed are part of the job. A low-power machine that moves inference back to a paid cloud service did not necessarily save money; it moved the bill.

Measure the whole system at the wall

I do not compare a processor’s TDP with a mini PC product page. Neither number is the electricity bill. I want average wall power under comparable work, including the power bricks and external devices needed for the job.

If the old machine uses internal storage while the replacement needs a separate enclosure, the enclosure belongs in both the wattage and purchase totals. If a scheduled GPU job runs for an hour, a week-long measurement captures that better than a two-minute idle reading.

The always-on cost guide shows how I convert a meter’s kWh reading back into average watts and why an apparently idle service may still need to remain available.

The purchase price is not the sticker price

My net purchase cost includes what I actually need to pay to make the replacement usable:

I keep setup time out of the first calculation because it is hard to price honestly, but I do not pretend it is zero. A two-year electrical payback can still be unattractive if the move creates a fragile system or consumes several weekends.

Try the no-purchase options first

Before replacing a working computer, I check whether the waste can be removed without buying another one:

  1. Let it sleep. A machine that is needed eight hours a day should not be modeled as a 24-hour load unless it truly stays awake.
  2. Move or remove the forgotten job. Old services, duplicate backups, and abandoned containers can keep hardware running for no current reason.
  3. Consolidate carefully. Combining two lightly loaded hosts can remove an entire base load, but it also combines their failure mode.
  4. Limit the expensive workload. A GPU workstation may not need to remain fully awake merely because local inference is useful occasionally.

These changes can produce savings with no payback period because there is no new purchase to recover. They can also expose the actual requirement for the replacement if one is still needed.

When I would replace it anyway

Electricity is not the only legitimate reason to retire a computer. Unsupported software, unreliable storage, excessive noise, missing performance, or unavailable replacement parts can force the decision. In that case, I use the energy result to compare viable replacements rather than asking electricity savings to justify the entire purchase.

That distinction matters. “I need a supported machine and this one also saves $80 a year” is an honest decision. “This $600 computer pays for itself” is not, unless the measured saving actually reaches $600 within the period being claimed.

My stop/go check before buying

  1. I measured both systems at the same boundary or have clearly labeled estimates.
  2. The replacement can perform the same workload.
  3. The purchase total includes every required part.
  4. The runtime reflects how the machine is actually used.
  5. The payback finishes before I expect another replacement for unrelated reasons.
  6. The result still works when I make the assumptions slightly less favorable.

If the decision collapses when electricity moves by two cents or the replacement draws five more watts than hoped, I do not have a sound reason to buy yet.

How I calculated the break-even point

The worked comparison uses the same arithmetic as the payback calculator. It is a transparent example, not a test of a named mini PC. My local-AI link documents a real machine and measured inference performance, but it does not supply an energy claim for this page.

Disclosure: This page contains no affiliate links. If I later link to a replacement computer or measurement device that pays a commission, I will label that relationship next to the link.