Run the Numbers Before You Buy a Home AI Box

The box in your cart starts from a familiar fork. Someone on Reddit asked those who spent over $10,000 on a home setup whether they regretted it, gathering 338 upvotes and 347 comments. The trigger was how cheap Chinese plans have become, like Z.AI with GLM-4.6. People weigh whether a home rig justifies its money and time next to a subscription they already pay, and the hesitation lives right in that spread.

Why the used price refuses to fall

The most hunted card is the five-year-old RTX 3090, still the cheapest dollar-per-gigabyte at 24 GB. A tracker of completed eBay sales logged about $639 for January 2026 and roughly $946 by June — a 47.9 percent climb in half a year. By September 16, two trackers reported around $1,400 and $1,415, while owners hunting a second rig cite $1,300 to $1,500 on the open market.

The August claim that every card tripled does not survive the tracker. Of 16 cards followed, only four rose more than 5% in six months while six fell more than 5%. Calling the DGX Spark daily-melting electronic waste is a similar take, not a measurement. The used market is hot for one card, not for all of them.

The math next to your plan

Since resale is a guess, the only hard number next to a box is the plan you already pay. A $4,699 DGX Spark equals about 23 months of a $200 monthly plan, ignoring power, resale and any future hike. A telling break was Claude Code rate-limiting: someone on the $200 Max plan hit the cap so hard they bought a Spark the next day for about $5,000, betting on infrastructure no one can throttle.

Unified memory is what makes these boxes tempting. Spark offers 128 GB at $4,699 in September, up from $4,000. The Framework Desktop with Ryzen AI Max+ 395 and the same 128 GB is $3,449, up from $2,459 in January and $1,999 at launch. A $3,000 no-name mini PC gets a flat rejection: no BIOS updates, no driver updates, no warranty, no thanks. The new Mac Studio with M5 Ultra starts at $5,499 with 96 GB, so the biggest models now fit, at least on paper.

Fitting is not the same as answering quickly. Because each new word forces the model to be read again, speed is a bandwidth job, not a capacity job. Spark delivers 273 GB/s, a Mac Studio comparison is near 800 GB/s, a used 3090 pushes 936 GB/s. At current prices the 3090 gives about 11.5 times more bandwidth per dollar than the Spark; at $946 in June the gap was about 17 times, so the gap narrowed because the card got expensive, not because the box got faster.

Why a coding agent stalls where chat flies

A coding agent is a different animal. It reads your codebase and edits it, but before your prompt it preloads its own instructions: about 33,000 pieces for Claude Code and about 7,000 for OpenCode, varying with the tool. A measured run with a 120-billion-parameter model on Spark reads a prompt at 1,676 pieces per second with 20,000 in context and writes at 36.7; at 48,000, read 1,225 and write 30.4. Those 33,000 starter pieces therefore stall 20 to 27 seconds before the first word. An owner of both Spark and M4 Max Studio says Spark still wins total time because it reads faster, even with more bandwidth on the Mac. Context caching reduces rereading but does not change which box is faster on which model. On a Qwen3 27B mid-size target, Spark and the AMD Max+ 395 box are the worst pair, and a heavily squeezed 1-bit DeepSeek V4 Flash on the same AMD chip in a Strix Halo hobbles at 9 to 13 per second with a long first-word wait, called unrealistic for agent work by its otherwise happy owner. A $48,000 home server owner's ChatGPT versus local anecdote of a few seconds against more than 10 minutes, 10 to 100 times faster, is self-described as a feeling, not a benchmark.

The 3090 fixes speed per dollar but reveals a size ceiling that fuels the next craving. One card holds 24 GB, an RTX 4090 also 24 GB and barely holds a 30-billion model even squeezed, so buyers talk about needing two, then four. The models people admire are larger: the big open GLM family after 4.6 wants over 1 TB. Today a 128 GB box tops out with lower-end Qwen, Gemma and Llama variants and the top answer to whether a 128 GB MacBook Pro will ever match frontier coding is no, not even close. Still, the only way to grow is stacking cards, and that thirst never ends: one owner's regret is buying one instead of two, another with two craves four, and the most discussed climb from a MacBook to an RTX 5090 gaming PC to a 96 GB Mac Studio M3 Ultra and finally to four RTX PRO 6000s with 384 GB in a basement rack fixes one limit and adds something to try, the builder deliberately not computing the break-even date.

What it costs and what comes back

At the top shelf the price walk is steep. RTX PRO 6000 Blackwell 96 GB was listed at $8,565 at launch in March 2025, $13,250 in June 2026 and $16,000 on the vendor marketplace in August, 87% above launch. The RTX 5090 from $1,999 to $6,899, owners of two say worth more than double what they paid, and a $1,700 launch 4090 now feels essentially free. Resale stories depend on when you bought: cheap buyers can sell at a small profit, but you would buy at today's top. Apple trade-in looked low while the same Mac on the open market held near what people paid, a 512 GB M3 Ultra in the UK seen at 21,000 pounds. Power hides in plain sight: a 5090 at 475 to 500 W scorched a connector, a 3090 box idles at 85 to 100 W with two monitors while others see 45 to 50. Leave 100 W idle all year and you burn 876 kWh, about $105 at $0.12 per kWh before a token is written. One buyer stops scaling because the next step needs an electrician and a soaring bill, the winter joke being heat the house with cards, better than gas.

Put together, the fair compare is still the plan. A self-described local believer calculated $10,000 buys about 6.2 billion pieces on a certain pro plan and advises unless your data must stay home to wait for small models to improve and meanwhile try a sensible 24 to 32 GB card for Qwen3 27B. The scale argument sits beside it: a data center busy 90% of the time will almost always run a model cheaper than a home box idle 90% of the time. Waiting has another reason: contract DRAM up 171.8% year over year in November 2025, still climbing, so advice around the new Mac Studio is to keep paying cloud until the memory squeeze passes because 96 GB will be thin in two to three years. A running joke says do not buy yet, AI is about to have its cheap moment, while the opposite doubt says frontier models often get quietly dulled and prices are subsidized and will jump when the subsidy lifts. Both are takes, not measurements.

On money alone the box loses; it wins only on something else: ownership and privacy with a modest model, plus pipelines that loop endlessly. The top-voted reply frames home AI as not a cost saver, arguing it only makes sense for those who value control and privacy — and even then only with a compact model such as Qwen3 27B. Some call their data priceless, some say their Macs are not stealing ideas, some argue absolute privacy rules out any commercial provider, one runs locally so no provider trains on their lifework. Pushback calls privacy a stated reason that often justifies a hobby, while someone with real privacy needs assumes no one drops $10,000 just to save money. The other reason is unlimited looping, automated pipelines that would be pricey in the cloud. Owners do not hide the hobby: one started asking an AI for exam flashcards and ended picking up second-hand MI50 boards from China and fine-tuning models all day while the exam slipped away. Laughs, best addiction, knowledge is never bad. The practical path is modest: rent that PRO 6000 by the hour on RunPod before buying, learn from a water-cooled Threadripper with 768 GB and four 3090s that mostly sits apart due to a bad used board while a plain old PC with one 3090 keeps its owner content, start with a 16 GB AMD card that was almost enough and move to a 32 GB card that covers everything, or grab a used $1,300 PC with a 3090 that runs Qwen3 27B with large context and already makes you crave a second. So the cart question changes shape: not whether local matches cloud, but whether local is good enough for the jobs you actually need — and even then, start small.