The cost to purchase and run a very capable home AI rig

What this rig is for, in plain words

This box runs AI models at home: the kind that answer questions and the kind that draw pictures, on a machine in the room, with nothing sent to anyone's servers. Inference is a model answering: you type a question, or hand it a page, and it writes back. Diffusion is a model drawing: you type a few words and it paints a picture. It does both with no account and no cloud, and it does them at the size that matters. Two cards give it 48 GB of graphics memory. That is room for a 30-billion-parameter language model with a long page in front of it, or an image model drawing at full size, or a smaller model of each running at the same time. Most of the benches on this shelf that did not go to the big workstation card ran here. It is what a person who wants the real thing at home, rather than the mini, ends up building.

Where this rig sits among the three tiers, and what is measured

This is the top of the three home tiers this series prices. The bottom is the mini, $959.00 as one listing on 10 September 2026, which writes at about twelve tokens a second and takes minutes to read a long page. The middle is one used card in an ordinary desktop, and that page is coming soon. This is the pair: the two cards alone are $3,699.98 on the same renewed listing that morning, almost four times the whole mini, before the box around them.

What it can do is measured on this shelf, and the performance section below prints the numbers. For pictures, a consumer RTX 3090 in this workshop drew its own game art, render for render, against the big workstation card on A Rig Your Friend Already Owns. For words, the bench of the pair is queued; this page adds its numbers below when it runs, with the window dated at both ends.

Nothing in this box is server-grade

Two EVGA GeForce RTX 3090 XC3 Ultras, renewed, 24 GB of graphics memory each, 48 GB between them, on a typical consumer board with two slots, desktop memory and a desktop power supply. They are the kind of parts a person buys one at a time from the same retailer this page read. No server, no workstation card, no rack, nothing a home does not already have a place for. That is the point of the box, and of this page: the cards are the expensive part, and they are consumer cards.

The bill, on one day

The mini PC that ran the two-hour battery test is fun, and it can do things. It is also slow enough at reading a long page that the first word of an answer can take a minute and a half. Readers of that piece asked what a real card costs. This page prices the whole rig that answers that question in this workshop. Every price below was read off amazon.com's own product pages, never a search tile, between 06:00Z and 06:14Z on 13 September 2026, before tax and before shipping. The used figures are from the same morning: the published sale averages of the one marketplace that publishes its sold listings. Three rows are the rig's real parts. Five are stand-ins for parts this page could not name at the time of the read, priced in the class the rig uses, with the word representative beside each. The whole rig is the point, and a reader building one will pick their own.

Every price is the retailer's own product page, before tax and before shipping. The cards are 73 per cent of the bill. With the pair bought used instead, at $908.00 each, the past year's average completed sale on the marketplace, the same rig is $3,164.87; at August's average of $1,094.44 each it is $3,537.75.

One card or two

One card is already a very capable home AI rig: 24 GB holds a 30-billion-parameter model at 4-bit with room for a long page, and draws pictures at full size. The second card is the step up to the biggest open models and to running a language model and an image model at once. The same bill, both ways, before tax:

The two-card idle is the meter reading; the one-card idle is that reading less one idle card, about 24 W by the card's own reading, and every loaded row is the estimate the running-cost section explains. Everything but the cards is $1,348.87 in both columns; a one-card build would not need the second slot, and could take a cheaper board.

What the second card buys is not double the speed of one question; it is flexibility. Three ways to spend it, all of which this workshop has run: keep two models warm at once, the language model on one card and the image model on the other, so a question and a picture never wait for each other and neither has to be reloaded; or run the same language model on both cards behind a batching runtime and answer twice the questions at once, roughly double the throughput at the same speed per question; or split one model across the pair, up to 48 GB of it, which is how the biggest open models fit in a home box at all, at the cost of the two cards talking over the bus. One card does any one of those jobs well; two cards do two of them at once, and never make you choose which model stays loaded.

What kind of performance to expect

One RTX 3090 on its own, from this workshop's own bench notes of 24 August 2026 (85 receipts, the model server of that day, one question at a time, measured before this workshop capped its cards), and, in the last row, the same card two days earlier serving eight questions at once through a batching runtime:

Pictures, one card, from the same render bench that A Rig Your Friend Already Owns published: one consumer RTX 3090, alone on its box, re-drawing this workshop's own game art on the night of 26 August 2026, 165 renders across six image models whose licences let this workshop re-share the output (five open licences and one community licence with a revenue cap), the card reading 320 W at the wall that night, above the 280 W this page recommends, so a capped card draws each of these a few per cent slower, unmeasured. Each row is the median of its cell, warm-ups excluded, in seconds from the render engine's own clock; the published page carries the 768-pixel rows and their ratio to the 96 GB card, 3.4 to 4.9 times slower, and each of them a working speed.

One of these two cards, held at 250 W, is the print lab's painter: it draws a reader's three plates in about 18.5 s on the lab's page, every day, on the box this page prices.

The power cap costs less than it looks. This workshop runs every 3090 at 280 W, 80 per cent of the card's 350 W rating, and measured the cost on 23 August 2026 at 3 per cent of writing speed and 6 per cent of prompt reading; a reader who leaves the card at 350 W gets that back and pays 70 W for it. The pair working together has no measurement yet; that bench is queued, and its numbers come below, with the window dated at both ends, when it runs.

What the card is like for training

Added 2026-09-15 (UTC): the readings above are this rig answering — a model already trained, being asked a question or asked for a picture. The card also trains, and this shelf has published the receipts for that. What it trained are LoRA adapters for ACE-Step 1.5, an open-weights music generator: a small detachable file of weights learned on top of the frozen model — rank 64 in every one of these runs, 44,040,192 numbers, 88 megabytes on disk in the first of them. Four of this workshop's five music adapters were trained on one RTX 3090 with its power limit set to 250 watts; the fifth went to the big workstation card instead.

The runs are short. Teaching the model Chopin — ten passes over nineteen recordings, fifty weight updates in all — took 4 minutes 49 seconds by the trainer's own clock and peaked at 5.7 gigabytes inside the training process, under a third of the card (Teaching a Music Model Chopin in Five Minutes, in Five minutes of training, honestly accounted). Two more composers followed on the same card by the same recipe, 14 minutes 36 seconds for Sousa and 8 minutes 23 seconds for Bach — 27 minutes 48 seconds of training for three composers (Half an Hour with Dead Composers, in The bill). The longest run the series published is the house-music adapter: 331 tracks read five times over, 415 optimiser steps, 39 minutes 56 seconds of training at 5.73 seconds a step, the same 5.7 GiB peak, on a card that page dates to over five years old and holds at the same 250 watts (The Ceiling Is Not the Corpus, in Forty minutes on a card bought over five years ago).

What that says about the card for training is what those pages say and no more: an adapter of this kind, on a music model of this size, is minutes to an hour of work on one used consumer card, and it never asked for more than 5.7 gigabytes of the 24. Training a language model is a different bill, and nothing on this shelf has measured it. The whole adapter series is here to read — Chopin, two more dead composers, a genre instead of a composer and the house corpus that closed it — and the first of them writes the whole recipe down, which makes it the one to start with.

Two notes on the bill

The memory. No 4 × 8 GB DDR4-3200 kit was buyable that morning; five named kits all read currently unavailable. The rig's own configuration is therefore priced as two 2 × 8 GB kits, $35.03 more than the same 32 GB as two 16 GB sticks. A reader building new should buy two sticks and leave two slots empty.

The board. The representative board may run the second card on fewer lanes than the first, and this page did not verify its two slots against the maker's page. A board whose maker's spec verifies two cards at eight lanes each, the ASUS ROG Strix B550-E, read $499.99 the same morning, and would add $340.00 to the total.

The cards, in May and in September

The cards are the expensive part, and the part that moved.

The same listing is 28 per cent dearer than in May, and the cheapest new pair of any brand is 26 per cent dearer than this pair cost renewed; EVGA's own new FTW3 Ultra, one seller, one left, read $1,965.00. The used average is the year's, from 110 sales; August's alone was $1,094.44, and a reader buying used this month should expect that, not the year's.

The renewed row is one refurbished card from one seller, the listing a reader meets first at that price, not a market price. The retailer's search refused this reading three times inside the window, so every price is off a product page. Every other 3090 listing this page opened that morning, more than ten, from $1,529.99 to $1,799.99, is renewed too, in titles the search tiles cut short.

Why these two are EVGA, and why the cheapest card is a different row

The cheapest new RTX 3090 on the retailer that morning was not an EVGA, and the table keeps the two apart on purpose. In an operator's own experience across several cards of the same chip, the range of quality between vendors is huge, and EVGA's are the best of them: usually better built than the reference cards, with a clean design, a solid heatsink and thermal pads where they should be. That is what a card running for months under a power cap needs. This pair was bought renewed, for that, not for the lowest price: a renewed EVGA over a new card of another make. It is a preference stated as one, from a handful of cards, not a measurement; nothing on this page benches one vendor against another. One fact belongs beside it: EVGA left the graphics-card business in 2022, so every new EVGA 3090 on sale today is old stock in a reseller's hands. That is why the new FTW3 Ultra row reads one seller, one left.

What the rig costs to run

The buying price is the big number; the running cost is small. This page measures it the way the power page measures every box on this shelf: the backup unit reports the box's draw at the wall once a minute, and the day's median is the idle figure. This box is the one the power page calls the darkroom's box. Dollars are at the same rate that page uses: 18.34 cents a kilowatt-hour, the United States residential average for June 2026 as the Energy Information Administration publishes it. That is a national figure, not this workshop's bill; a month is thirty days, and the month's dollars are the printed kilowatt-hours times the rate.

The idle row is a meter reading. The two loaded rows are estimates made in good faith from the caps, a ballpark, not readings. This workshop holds each card at 280 W, 80 per cent of its 350 W rating, the posture every card on this shelf runs under. The idle row already carries both cards idling, about 24 W each by the second card's own reading over the same day (0.59 kWh doing nothing), so a card at its cap adds the 256 W between its idle and its cap, 512 W for the pair, which the power page's own rule divides by 0.9 for the power supply's losses: 569 W added at the wall. The ceiling is what the box would draw if both cards never dropped below their caps for a full day, which real work never does. This box is never off: it is the print lab's darkroom, and one of its two cards draws a stranger's pictures every day.

The cards are three quarters of what the rig costs to own and a rounding error against what it costs to run; on this rig the buying decision is the whole decision.

What this page does not say

• An asking price, not a sale. Every new and renewed figure is a price as displayed on one day; the used figures are the marketplace's own averages; no card was bought for this page.
• One retailer, one morning. Prices moved before you read this; the page prints the fourteen minutes it read in.
• Nothing about the rest of the box beyond its price. The memory, the power supply, the board and the case around these two cards are priced here and judged nowhere; they are a page of their own.
• One card's speed, not the pair's. The word figures are this workshop's own bench notes for one card, not a published page; the pair's bench is queued and ships with its kit. What a card gives up under a cap on the big workstation card is on the power-cap page.
• Adapter training, not model training. The training figures are this shelf's own adapter pages, linked above: small LoRA adapters on one music model, on one card. No language model was trained or fine-tuned for any page on this shelf.
• Derived watts, not measured ones. Only the idle row is a meter reading; the working day and the ceiling come from the caps and the power supply's loss rule. A real day of inference sits between the first row and the second.

This page adds corrections and later readings below, each dated (UTC), with a window at both ends where one applies, and each saying in plain words what it counts.

How to check our work

Every row names its listing as titled; open the retailer, search the title, and read the price on the day you do. If a number here does not reproduce when you read it, prices moved, and the page says which fourteen minutes it read. If it does not reproduce for the day printed, say so at the contact desk, where a person reads every message.

Who ran this, and thanks

The prices are amazon.com's own product pages on 13 September 2026 and jawa.gg's published sale averages for the RTX 3090 the same morning. The photograph is the workshop's own. The pictures beside this page's row on the shelf, and its link card, are renders based on photographs of the real rig, and nothing on the page itself is a render. A vision model on this workshop's own seat, Gemma 4 26B (Google DeepMind, Apache 2.0), described an operator's photograph of the box in one paragraph; FLUX.2 klein 4B (Black Forest Labs, Apache 2.0) painted that description in a cyberpunk world of 2055, through ComfyUI (GPL-3.0), on this very box, one of its two cards, in 2.8 to 3.2 seconds a frame at 1,344 by 768; thirteen frames were drawn across two batches and four kept, every one re-saved without metadata. The photograph itself will follow when its owner has cropped it.

A small human team bought the cards, checked the prices and signed the numbers; a fleet of AI agents read the retailer, did the arithmetic and drew the pictures under that team's rulings. Neither the retailer nor the marketplace nor any model's maker owed us anything. Thanks to the readers of the mini page who asked what a real card costs; that question is the whole reason this page exists.