You paste a very long prompt into the terminal. You want the model to build a complete front-end site: TypeScript, mock data, dark mode, accessibility, unit tests, and it has to run its own commands to check the result. You press Enter. GPU usage on all four machines jumps to 100 percent and the chassis slowly warms up, yet not a single word appears on your screen.
This is not a malfunction. It is what running a very large model locally really looks like: the machines are busy, and all you can do is wait.
The scene comes from a test video by the YouTuber Alex Ziskind. He gave the same front-end job to a four-Mac-Studio cluster he built himself, running the open model Kimi K3, and to the cloud platform Abacus AI Supercomputer. Both sides finished. What differed was the time, the money, and what you are willing to hand over to someone else.
This Is Round Two of Local vs Cloud
Some background first. The author had run an earlier match in which the local side was a single 512 GB Mac Studio, and commenters challenged him: that is not a real local setup. So this time he pushed the hardware as far as he could.
The model was Kimi K3, described in the video as 2.8 trillion parameters, with a raw file of about 1.56 TB, far too large for a single Mac Studio. The hardware was four Mac Studios with 512 GB of unified memory each, 2 TB in total, connected pairwise with Thunderbolt 5 (six cables), with RDMA and MLX distributed splitting the model across the four machines. What actually ran was a reduced quantized version with weights of about 817 GB, but without pruning, so all 896 experts were active. The machine count is four because tensor parallelism splits the model dimensions evenly, which only works with 1, 2, 4, or 8 machines. Three would not work.
One easily missed detail: 817 GB looks like it would fit on two machines, but the author points out that he fed in a very long context, which needs far more space than the weights alone, so two were not enough. The on-screen readings during the run showed the four machines using roughly 208, nearly 200, 177, and 215 GB.
The Cloud Side: Not a Chat Window, a Computer That Never Shuts Down
The selling point of Abacus Supercomputer is that it is really a persistent cloud virtual machine, not just a chat box. According to the video's description, it bundles scheduled tasks, a database, storage, GitHub, and SSH, and lets you switch among more than 100 models from one entry point, from GPT, Claude, and Grok to DeepSeek, GLM, and Kimi K3, with RouteLLM choosing a model based on the situation.
This side used Opus 5 (high) and GPT-5.6 Sol. Note that the two sides did not use the same model, so this is a contest between a whole local setup and a whole cloud setup, not a pure model-versus-model comparison.
There is also a disclosure to make. The video states that it is sponsored content from Abacus. The author himself says in the video that he cannot get too excited because this is a sponsored video. When you read a comparison like this, it is more accurate to treat it as a case with a stance, but whose demonstration you can inspect.
One Prompt, Two Kinds of Waiting
Both sides received the same prompt: build the front end of a silicon compute exchange, using TypeScript and mock data, with dark mode, micro-interactions, accessibility, and unit tests, and with the model verifying its own work. The author also says this was the first time he had actually run this prompt.
By the author's demonstration, the cloud side moved quickly. It produced HTML fast, built a Next.js project inside the virtual machine, and the result could be launched and viewed from that machine. The author's reaction was that the result looked quite good, with a dashboard, live filtering with animation, inventory browsing, a comparison panel, and a booking feature, all from a single prompt (no backend attached).
The local side had a different rhythm. The numbers the author measured were about 238 tokens per second for prompt processing (prefill) and about 14.7 tokens per second for generation. He explains that the Mac's strength is high memory bandwidth, so generation is smooth once processing is done, but the prefill stage on a long prompt is weaker than on Nvidia GPUs. Taken together, the two numbers mean that the longer the prompt, the more patience you need.
The Turn: It Was Not That It Could Not, You Just Had to Walk Away
At first the author thought the local side was broken, because he waited a long time without seeing any output. He switched to a short "hi" test and the model responded. He then asked for a simple web page with a text box and a button, and it finished in about 3 minutes with 147 lines, which he described as quite good looking.
So he stopped recording, fed the full prompt in again, and went for coffee. About 4 hours later he came back to a complete application: filtering, comparison, charts, scheduling, and light and dark modes were all there.
So the answer to whether it can be done locally is yes. But the author's summary is that the same application took about 15 minutes in the cloud and about 4 hours locally. That gap is the real price tag.
How to Do the Math: One-Time Purchase vs Monthly Fee
The cost comparison in the video is worth taking apart. On hardware, each Mac Studio is about 16,000 US dollars, and the author guesses it is no longer sold (his words were "I think"), so prices would only go up. Once you buy, it is yours to use any way you like. On the cloud side, Abacus starts at 10 US dollars a month, 7 dollars for the first month. The author's point is that on monthly fees alone, it would take a very long time to catch up with the hardware cost. On routing, the cloud plan lets you use a custom router that sends easy work to cheaper models and saves the expensive models for hard tasks.
These figures are all spoken in the video, and we have not independently verified them, so check official announcements for actual prices. More important is what the comparison leaves out: the cost of waiting 4 hours, electricity, and maintenance time. What you save is the monthly subscription, and what you pay for it is time and management burden. There is no standard answer to that trade.
For teams that use APIs often, this is the same kind of choice you face when reading your usual Token bill: expensive models handle the hard problems, and cheap models handle the odd jobs.
When Should You Stay Local?
The author's own conclusion is fairly restrained, and he even says it is not what the sponsor wants to hear. If you want high quality, fast results, and fully managed one-step delivery, the cloud plan wins. But if your data cannot leave the office, or you simply like running things yourself, a cluster still has a reason to exist. He believes both markets exist and it is not an either-or choice.
His own way of working is to use agents to build all kinds of one-off small tools, which need to be fast and runnable right away, and for that the cloud suits him. It is also a reminder for readers: first work out what your task actually looks like, then decide where to put your resources.
FAQ
Q1: Did the video compare local AI and cloud AI using the same model?
No. The local side ran Kimi K3, while the cloud side used Opus 5 (high) and GPT-5.6 Sol. What is compared is the completion speed and output of whole setups, not a ranking of single-model capability.
Q2: Is running a large model locally always slower?
It depends on the hardware and the task. The conditions in this video were a very large model and a very long prompt, with a measured prefill speed of about 238 tokens per second, which is why it was especially slow. The author also explains that on the Mac, the generation stage after the prompt is processed is relatively smooth. With a smaller model or a shorter prompt, the conclusion could be completely different.
Q3: Can 4 hours vs 15 minutes be treated as a general speed gap?
We do not recommend it. This was a single test, the author himself says it was the first run of this prompt, and 4 hours is the rough time he saw when he came back after leaving his seat. Treat it as an observation for this scenario, not a benchmark.
Q4: Why four Mac Studios and not three?
The author explains that tensor parallelism spreads the model dimensions across nodes, and the dimensions must divide evenly, so the machine count can only be 1, 2, 4, or 8.
Source Statement
This article is a rewrite and summary, not a sentence-by-sentence translation. Its content is reorganized from the subtitles of the YouTube video below. All hardware specifications, speeds, durations, and prices are the video author's spoken figures, which we have not independently retested. The subtitles are auto-generated, so some model names may be misspelled. The video states that it is sponsored content from Abacus AI, and the descriptions of the cloud product's capabilities are self-reported by the sponsor and the author.
Video title: I Gave Local AI and the Cloud the Exact Same Job. Channel: Alex Ziskind. Link: https://www.youtube.com/watch?v=ujs0_cpAnaw. Publication date: 2026-08-31.
Further Reading
If this topic interests you, we have published related articles on this site: a hands-on comparison of Opus 4.8 and Fable 5 building the same app with the same instructions, a comparison of GPT-6 Astra and Fable 5.1 across 15 real work tasks, why AI Tokens are so expensive when you break down the autoregressive mechanism behind coding agents, and five tested tips for reducing the Token cost of AI coding, using Claude Fable 5 as the example.
Before You Press Enter Next Time
Back to the opening scene. Now you know that four machines at full load with no output does not mean they are broken, and it does not mean the model is slow-witted. It only means you chose to trade time for control.
The questions worth asking before you press Enter are these. Does this job need to be delivered in minutes, or can it run for an afternoon? Can the data leave your premises? Are the hard steps and the easy steps worth handling with the same model?
If you are still undecided whether this job belongs locally or in the cloud, pick one real task of your own, run it once through the same key on AI Token King, note the Tokens actually used, and then come back to do the math. Try AI Token King free: https://aitokenking.com.tw/home
