AI hardware ROI for a personal AI workstation: comparing the complete purchase price

Add the PSU you swapped because the card pulls more power than the old one could push. Add the fan you replaced for the coil whine. Add the second NVMe once quantized weights filled the boot drive. None of that was in the number you quoted.

A personal AI workstation is one desktop: a capable GPU, RAM enough to hold context, steady cooling and power, built for interactive coding, private document work, experiments, and the odd overnight batch. The alternative is API calls blended with whatever subscription you're already carrying and the machine already on the desk. Price it per month of verified developer work supported.

What actually goes on the bill

Capital cost is every component, tax line, shipping, spare part, and whatever hardware the new box displaces:

installed_capex = compute + host + memory + storage + power + cooling + network + tax + setup

Build a bill of materials against that formula, then subtract anything you'd have bought anyway: a laptop for ordinary work stays off the AI ledger, a memory upgrade bought solely for a model does not. Split fixed from variable too, hardware and installation sit still, electricity and tokens move with usage, and human review swings either way.

The measurements that decide it

Run the same workload through local, rented, and API routes in local-first cascade order, and log accepted tasks against failures, the input/cached/reasoning/output split, and hours active versus idle. Price the finished result: a cheap model needing two tries and five minutes of correction is not cheap, and a local model that nails one narrow job doesn't belong up against a frontier API at full reasoning effort.

The failure mode here is low utilization and hobby time booked as business savings. Model low, base, and high demand, and include the daily peak: a box at full tilt has no room for an interactive request. The comparison people botch: a bare used GPU against a finished, cooled, powered computer, or a cloud endpoint bundling someone else's ops labor.

Where the payback line sits

Payback month is when cumulative discounted benefit clears cumulative cost; ROI is that benefit over discounted cost across the horizon, needing a conservative useful life since a card can outlive its usefulness once the model lineup moves on. Productivity savings need a discount of their own: minutes saved only count as revenue if they avoid a hire, cut outsourced spend, or clear a binding queue. Run it once with that term at zero: a purchase that only clears the bar there is a warning, not a green light.

The line you commit to

State the decision as a boundary, not a recommendation: accepted tasks per month, productive GPU hours, a ceiling on API cost per task, or a minimum useful life to clear. Boundaries get revisited; narratives get defended. Put a retest date on the calendar: rerun when volume shifts, the quality gate changes, API prices move, or a part ages out of support.

Price the machine that gets plugged in and turned on. Not the card that shipped in the box.