Skip navigation
This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B…
0:00 / 1:16

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B…

Blaze

@browomo23K followers

Visit on X
331 likesDownloadOriginal post

94,424 views

This Chinese developer linked two $2,999 NVIDIA DGX Sparks into one box and runs the full Qwen3-235B at home, after dropping his $1,999-a-month cloud bill to zero. He wired 2 small boxes into a single computer, split a giant 235-billion-parameter model in half between them, and serves it across his own network at about 10 tokens a second, with no internet, no cloud, right there on the desk. No data center, no thousand-dollar graphics cards, no monthly cloud bill. Just him, 2 gold boxes the size of a sandwich, one cable between them, and 1 power strip. And here is the whole payoff. He used to pay the cloud $1,999 a month for the same model, and the meter ticked on every request. Now he paid $5,998 once for 2 boxes, they covered their cost in 3 months, and after that he sends as many requests as he wants for free, only electricity. The two Sparks talk over one fast cable, each holds 128GB of memory, and together they carry the whole model, about 73GB loaded per box, with the chip inside pinned near the limit at 96%. Both boxes work as one and keep trading data over the cable, with no cloud in the loop and no single word leaking out. The ready model sits on one local address, and any app on his network calls it as easily as ChatGPT. And here is how he described, in plain words, what this pair of boxes does: "this is a pair of boxes that holds the huge Qwen3-235B model and serves it to one network. the model is split in half, and each box owns its half. parts: // Box 1 (holds the first half of the model and starts the answer fast, the first word appears in under a second) // Box 2 (holds the second half and writes out the rest, about 10 tokens a second) // Cable (connects the 2 boxes and moves data between them on every step, with no lag) // Address (one local address where any app sends its request, like to a cloud model) // Test (a script that runs big prompts through and measures speed and delays) // Monitor (checks temperature, power draw, and load on both boxes every 2 seconds). the model never goes to the cloud. he only steps in when a box runs hotter than 80 degrees or the cable between them starts dropping data." So the system knows exactly what it is, what it is for, and where its limits are. It knows it has to hold the whole huge model across 2 boxes on its own. It knows it has to answer every request locally, with no meter, no limits, and no internet. It knows the human is only needed when a box overheats or the link between them stalls. → The setup runs around the clock on 2 boxes, each pulling under 60 watts → However many requests he sends, the monthly bill is $0, only electricity → The first box starts the answer in under a second → The second writes text at about 10 tokens a second → One request at a time: 838 tokens in 85 seconds, first word in 0.8s → Two requests at once: 697 tokens in 108 seconds, first word in 0.7s → Both boxes sit at 96% load and warm up to 76-78 degrees And only when a chip in a box runs hotter than 80 degrees or the cable between the 2 Sparks drops data does the system call the owner. And when he himself is out on a run or in a coffee shop, he still reaches his own model at home from his phone: sends a big prompt to the local Qwen3-235B, gets the full answer back in under a minute and a half, with no token meter ticking and no limit to hit. Here is what the test shows on his screen during one of the night runs: "one request at a time: 838 tokens in 84.9 seconds, first word in 0.8s, then 0.1s per token." "two requests at once: 697 tokens in 107.6 seconds, first word in 0.7s, then 0.15s per token." "Box 1: chip at 96% load, 76 degrees, 56 watts, 73GB used in memory." "Box 2: chip at 96% load, 78 degrees, 56 watts, the Qwen3-235B model fully loaded." And while everyone around is paying for AI by the month and bumping into limits, his top-tier model just sits on the desk and works as much as he wants: his own little power plant instead of a forever meter. He has no server rack of his own and no cloud account behind it. Just 2 DGX Spark boxes on a desk, one model split in half between them, one local address, and a folder of prompts next to it. Out of everything I have seen this year, this is the cleanest way to stop paying for AI: $5,998 of hardware on the desk once, $0 a month to the cloud, unlimited forever, and between them 2 gold boxes, 1 cable, and the full Qwen3-235B answering at home with no internet.

22 Comments

Comments from the original post on X (Twitter)

  • William Ruider's profile picture

    @ruider92545

    Yeah, it is 1500s permutation of the same BS. A Chinese developer, Malaysian teacher, Japanese influencer...🤣🤣🤣 I do the same without a fancy super duper shiny whatever specialised developer equipment.

    11 likes

  • Blaze's profile picture

    @browomo

    Wow how powerful is this infrastructure?

  • William Ruider's profile picture

    @ruider92545

    Good compute power 💪, 3-20 TOK/s 😂, les then 300W under the heavy load 💪, no air-con for cooling 💪, 400B+ FP16/BF16 local LLM models 💪, powered by solar panels💪. 5💪: 1😂 = 😇

    2 likes

  • Blaze's profile picture

    @browomo

    Ahhahahahah Bro this is already a whole farm

    1 like

  • Noah King's profile picture

    @digitalnoah

    I’m pro AI and a local AI enthusiast. There’s three things you’re misrepresenting here. 1. Qwen 3-235B isn’t a very good model compared to Chat GPT 5.5, Opus 4.8, or even Composer 2.5 Local AI is valuable, but doesn’t replace frontier models, it supplements them for easier tasks. 2. When you string 2 computers together, it doesn’t create a single memory pool. Some software lets you split the jobs, as you stated here, but it’s sub optimal at best. 3. DGX spark boxes are slow. Very slow. Under 300 GB/s memory bandwidth means you’re running weaker models and you’re getting slower tok/s. There’s so many incredible things you can do with local, sovereign AI, but you have to set your expectations right. The hardware is expensive. The models are waaaay dumber than frontier. The tok/s speed is dependent on memory bandwidth and most home builds are slow as molasses.

    9 likes

  • MEZTech 𐤊's profile picture

    @MEZ_tech

    Crazy thing is even with this setup, not sure the electricity costs are less than the token costs for the equivalent token usage, not to mention this setup isnt even capable of producing high token throughput when desired. Absolutely love the idea of local llms, especially wrt data privacy, but theyre not necessarily cost effective

    1 like

  • Blaze's profile picture

    @browomo

    I see it differently. Two of these boxes draw just 120W combined that's pennies for electricity compared to a $2,000 monthly cloud bill. And 10 tokens a second is more than enough to comfortably read the response in real time.

    1 like

  • MEZTech 𐤊's profile picture

    @MEZ_tech

    86400 seconds/day * 10 tok/s = 864000 tok/day max output (disregard encoding time, etc). Qwen 3 255B output token costs are $0.45/1M So, cloud costs would be roughly $0.45/day = $13.5/mo if using Qwen 3 255B at the equivalent level of usage (rounding up tok/day to 1M to account for input tokens). So no, it does not "save $2000/mo"

    1 like

  • Blaze's profile picture

    @browomo

    wow that's a powerful argument

  • NMhao's profile picture

    @asukahaox

    それはすごいですね!技術の進歩は本当に驚くべきです。将来的にAIに支払い続ける必要がない方法を考えると、この方法はとてもクリーンで効果的ですね。

  • INFORMERCIALS - glutathainment's profile picture

    @Quake_jr

    Data centres gotta re-diversify major purposes cx DGX fixed consumer cloud services…..Oooppssss

  • XMon - XVine's profile picture

    @Christi18676713

    @IdleProtocol $Idle

  • Grebe's profile picture

    @0xGrebe

    bro turned a cloud bill into a desk appliance lol

  • Chexxe's profile picture

    @Chexxee

    bro really turned $2k a month into a desk appliance lol

  • sandybrige's profile picture

    @kanukagi

    それは本当に革命的ですね!自宅でAIを稼働させる方法が素晴らしいです。

  • Far's profile picture

    @FarAICoder

    curious if the cable is just ethernet or something custom for that 10 tokens/sec speed what happens if one spark dies mid-request

  • Stanislas Perzo's profile picture

    @caliroi70

    Et comment puise-t-il l’information en live sur le net pour préparer ses réponse ?

  • Kevin Chappell's profile picture

    @kevinchappell

    You should tell your model to be less wordy when writing posts. It's just writing the same thing 3 different ways.

  • Greg Sky's profile picture

    @regsky_

    yeah didnt we all try this day one of dgx spark? what changed, or are yall running at low tps?

  • BriscoCountyJr's profile picture

    @UselessTreasure

    So I think the point of this, is that you don't now what youre doing. Am interpreting that right?