Field notes / № 2

Cost analysis · Prices verified 2026-07-05

Local vs cloud: when it actually saves money (and when the internet lies to you)

At least eight live calculators promise to tell you whether a local AI rig beats the cloud. Nearly all of them get paid only when you buy the GPU. Here is what the math says when the answer is allowed to be no — and why one input, the cloud model you would actually replace, matters more than everything else on the page combined.

Every calculator has a thumb on the scale

Search "local LLM break-even calculator" and you will find a small industry. SitePoint runs a 36-month TCO calculator inside a blog post. CraftRigs published "the breakeven calculator nobody's built (until now)". There are at least six more. They are not all wrong. They are all leaning.

The lean is structural. A hardware affiliate commission on a $1,500 GPU is real money. A verdict of "keep your API subscription" pays the site nothing. So the defaults drift, quietly, in one direction: the cloud comparison uses the most expensive frontier model, the throughput numbers assume the GPU never idles, and the amortization window stretches until the answer comes out right. No single number is a lie. The sum of the defaults is.

Full disclosure, since we are pointing fingers: this site carries affiliate links too — on both verdicts. If the math says buy, we link hardware. If it says rent, we link API and GPU-rental referrals. The verdict is computed before any link renders, and the entire model is public. The "no" pays us the same as the "yes," which is the only arrangement under which you should trust either.

2026 broke the old answer

Two years ago the story was simple: frontier APIs were expensive, so a home GPU paid for itself quickly if you used it at all seriously. Then open-weight API prices collapsed. Here is the menu as of July 5, 2026 (pricing survey, DeepSeek pricing — always check the provider's live page):

That last row is the one the buy-a-GPU calculators skip. Fourteen cents per million input tokens is not a typo. It is roughly thirty-five times cheaper than the premium tier — and it is an open-weight model, which means it is more or less the same class of thing you would be running on your own GPU anyway. Someone else's datacenter is now serving your home-lab model for less than your wall socket charges to run it. We will show that literally in a moment.

One rig. Three honest answers.

Fix everything except the cloud baseline. The rig: a used RTX 3090 build, about $1,200 all-in at mid-2026 street prices (used-3090 market, GPU pricing). The usage: a daily coding assistant pushing 15M tokens a month — 7.5M in, 7.5M out — with 3 hours of inference and 5 hours of idle per day. Electricity at the EIA national residential average of $0.1883/kWh.

The local side of the ledger never changes: $33.33/mo of hardware amortized over three years, plus 48 kWh of electricity at $9.04/mo, for a flat local cost of $42.37/mo — whatever model you run on it. The only question is what that $42.37 replaces. Every number below comes from the same open engine that powers the calculator:

Used-3090 rig, coding-assistant usage (15M tokens/mo), national average electricity. Computed with the site engine; formulas and assumptions on the methodology page. Prices verified 2026-07-05.

Same rig. Same electricity. Same 15 million tokens. Three different answers, and none of them is a rounding quirk.

Against Sonnet 4.6, the cloud would bill $135 a month. The rig erases that for $42.37 and pays for itself in nine and a half months — call it ten — then banks roughly $93 a month. Clear win. Against Haiku 4.5, the cloud bill drops to $45 and break-even stretches to month 33 of a 36-month amortization. The rig technically pays off just before it is due for replacement, which is less an investment than a photo finish. Against DeepSeek V4 Flash, the cloud bill is $3.15 a month, and break-even does not stretch — it disappears. More on that below.

And if your baseline is the premium tier, the case gets embarrassing in the other direction: swap in Claude Opus 4.8 and the cloud bill is $225/mo with break-even at month 6; GPT-5.5 bills $262.50/mo and the rig pays off in under five months. Which brings us to the whole point:

The cloud model you compare against matters more than any hardware choice you make.

Upgrading from a used 3090 to a 5090 moves this break-even by months. Changing the cloud baseline moves it from five months to never. Every calculator that quietly pre-selects a frontier model as your comparison has made the decision for you before you touch a slider. The honest question is not "what does the cloud cost?" — it is "what would you actually pay for, if you didn't buy the box?" If the honest answer is a $3 open-weight API, the math needs to hear it.

"Never" is a real answer

The break-even formula is one line: months = build_cost ÷ (cloud_per_month − electricity_per_month). Ownership pays off by earning back the build cost through monthly savings. But if the cloud bill is already at or below your electricity bill, the savings are zero or negative — the denominator dies, and no number of months fixes it.

That is exactly what happens here. DeepSeek V4 Flash bills $3.15 for the month. The electricity to serve those same tokens locally costs $9.04. Every month of ownership loses $5.89 on power alone, before the $1,200 of hardware even enters the conversation. For the rig to break even, this workload would need electricity under about 6.6¢/kWh. The cheapest state average in the country — Utah, at 11.5¢ — is nearly double that. In Hawaii, the same rig's power bill is $22.38 against the same $3.15 cloud bill. There is no U.S. state where this pencils out.

This is the case the affiliate calculators structurally cannot show you, because "never buy" earns them nothing. We show it first, because it is true, and because you can check it yourself against Texas, California, or whichever grid you live on.

What buying actually buys you

None of this means local is a mistake. It means local is not always a savings. There are three things a GPU under your desk delivers that no API price cut can touch:

Privacy. Prompts that never leave the machine. For client code, medical notes, legal drafts, or anything under NDA, "cheap" may not be on the menu at all — and once the cheap APIs are off the table, your honest baseline snaps back to the frontier tier, where the break-even math genuinely favors owning.

Offline. It works on a plane, in an outage, and after any provider decides to deprecate, reprice, or "sunset" the model you built your workflow on.

No rate limits. No 429s, no usage tiers, no queue. The marginal token is free once the power is paid, which changes how you use the thing.

These are real, and for many people decisive. But they are value, not dollars, and an honest tool refuses to launder one into the other. That is why the calculator scores sovereignty on its own axis, next to the money and never mixed into it. "Buy it for sovereignty, not savings" is a perfectly good verdict. It is just a different verdict than "this saves you money," and you deserve to know which one you are getting.

The dollars case for owning is real too — it just lives where the incumbents' defaults happen to point: your true baseline is a frontier model, your volume is heavy, and the machine earns its keep daily. A coding assistant replacing Sonnet-class usage clears break-even inside a year on a used 3090, and the electricity is smaller than people fear. If that is your situation, buy the rig with our blessing. The point is to find out which situation is yours before the money leaves.

Run your own numbers

Your electricity rate, your token volume, and — above all — your honest cloud baseline. The engine takes about thirty seconds and is happy to tell you not to buy anything. That is rather the point.

Run the break-even test

Estimates, not quotes. All prices verified 2026-07-05 and drifting since; the methodology page lists every formula, assumption, and source so you can audit the math or run it by hand.