(The following scenario is an illustrative fictional case. The protagonist, "Wen Zixuan," is a representative figure used for explanation, not an actual customer.)
The Boss Dropped a Line in the Group Chat
"The company next door already bought their own GPUs to run AI. Why are we still paying Claude's monthly bill?"
Wen Zixuan is the technical lead at a startup of a dozen-odd people, and staring at the line his boss had just dropped into the group chat, he couldn't think of a comeback on the spot. Their Claude API bill really had been climbing every month, and lately the internet was full of claims that "a single Mac Studio can outrun Claude." His boss followed up with one more line — "Budget meeting next week. Bring me a comparison: buy our own hardware, or keep paying for the API" — and just like that, the whole question landed on his desk.
First, Find a Stand-In You Can Actually Calculate
The first thing Wen Zixuan hit when he started digging was that nobody actually knows how many GPUs it would take to run Claude Fable 5, because Anthropic has never published its parameter count, architecture, or weights. There was a whole Reddit thread guessing at how many H100s it would take to run something Fable-class — one person guessed four, another guessed eight B200s, someone else just said "you'd need a whole rack." The answers were all over the place, because nobody had actually run the numbers.
He eventually found a video that did the math using Kimi K3 as a stand-in, and that's when it clicked: this was the only way to actually calculate anything. Kimi K3's own materials publish 2.8 trillion parameters and a 1-million-token context window, and its launch materials directly compare it with Fable 5 on coding and agent benchmarks — though Kimi's own blog admits it still trails the strongest closed models overall, and a benchmark score isn't the same as doing the actual work well. The point isn't to claim Kimi K3 can replace Fable 5; it's that Kimi K3 is the only stand-in that's both frontier-scale and has a published, transparent size — which makes it the only thing you can actually do arithmetic with.
Once He Ran the Numbers, the Answer Showed Up Immediately
Kimi K3 stores its weights in a 4-bit format called MXFP4. 2.8 trillion parameters times half a byte gives you about 1.4 terabytes just for the raw weights — and that's before you count the inference software, the conversation context, or any safety margin. It's the floor, not the total.
Wen Zixuan worked through that math against every option his company could actually afford, and hit a wall every time. He first checked the company's top-spec Mac Studio — 512 GB of unified memory, not even a third of the floor value, out immediately. He remembered a coworker saying "four H100s should be plenty" — he ran the numbers: 80 GB each, four of them is 320 GB, even less than the Mac Studio, also not enough. Moving up, a DGX B200 with eight B200 GPUs has 1.44 TB of GPU memory — on paper that finally clears the floor — but after the weights, less than 40 GB is left, and the context window, conversation cache, and inference software all have to squeeze into that 40 GB with no room to breathe. He kept climbing and found the DGX B300: eight cards, 2.3 TB, leaving nearly 900 GB after the weights — theoretically, finally, workable. Except Kimi's own recommendation for "efficient inference" calls for 64 or more accelerators — which works out to eight servers, not one.
Run a single DGX B300 flat out and its published power draw is 14.5 kW; run eight of them and you're pushing close to 116 kW before cooling. He also found someone in the thread who said their company runs several dozen B200s, and powering them on "sounds like sitting behind an Airbus A380." He ran a rough estimate using today's public rental prices — eight nodes like that would run somewhere between $313,000 and $577,000 a month. He knew this was just an order-of-magnitude figure, not a quote, but it was already enough to show that the scale was nothing like what he'd originally imagined.
He Thought He Was Pricing a One-Time Purchase — He Was Actually Pricing a Data Center
Reading this far, Wen Zixuan realized his original assumption had been completely wrong. He'd assumed his boss wanted "buy some GPUs, pay once, never touch the API bill again" — a one-time capital expense. But once the math was laid out, it turned out that a self-hosted setup capable of running a Fable-class model isn't a "buy it once" hardware purchase at all — it's a full data center's worth of ongoing operations: power, cooling, network infrastructure, and someone watching over the machines around the clock. And that monthly cost was an order of magnitude higher than what they were currently paying Claude.
Even more important, the video pointed out that the Reddit thread had gotten its own question wrong from the start: "100 users" is not the same as "100 concurrent requests." If 100 employees each send one request every 10 minutes, and each response ties up the system for 30 seconds, the average number of requests actually running at once works out to around 5 — AI agents can push that number up fast, sure, but when Wen Zixuan went back and checked his own team's actual concurrent usage, it was nowhere near as extreme as he'd imagined.
What He Brought Into the Meeting Wasn't a Quote — It Was a Question
At the following week's budget meeting, Wen Zixuan didn't walk in with a purchase order for eight DGX B300s — that would have been jumping the gun. Someone in the video who had once run an LLM service for thousands of users said a line he'd written down: "A model that's good enough and that you can actually scale beats a super-model that can only serve a couple of people." It matched another pattern he'd found — teams that have actually run local models on their own are almost never "fully self-hosted": private or repetitive work goes to a small local model, and the hardest tasks still go to Claude or GPT.
What Wen Zixuan brought into the room instead was a question: "Do we actually know how many tokens, and how many concurrent requests, we really use every day?" His suggestion was to hold off on deciding whether to self-host at all, and first lay out the team's real usage — which tasks are high-volume but simple, which are low-volume but heavy — before deciding whether trading the API bill for the operational responsibility of an entire data center was actually worth it.
Frequently Asked Questions (FAQ)
Q1: What is "Kimi K3" mentioned in this article? What does it have to do with Claude Fable 5?
Kimi K3 is a large language model with a publicly disclosed parameter count (2.8 trillion parameters, mixture-of-experts architecture, 1-million-token context window), and its launch materials directly benchmarked it against Claude Fable 5. Because Anthropic has never published Fable 5's parameter count or architecture, this article (and the video it's adapted from) uses Kimi K3 as a stand-in that's "similar in scale but transparent in its published specs" to estimate the hardware needed for local deployment — it is not a claim that the two models perform equally.
Q2: Does an ordinary company actually need to buy its own GPUs to run a top-tier AI model locally?
In most cases, no. The source video's conclusion is that teams who have actually run local models are almost never "fully self-hosted" — small local models are well suited to private or repetitive work (like document search, classification, or routine coding), while the hardest tasks still go to cloud services like Claude or GPT. Whether self-hosting is worth it depends on your actual concurrent usage and workload pattern, not a knee-jerk instinct to cut costs.
Q3: Are the purchase and electricity figures in this article (like $313,000–$577,000 a month) precise quotes?
No. These figures are rough order-of-magnitude estimates the source video's creator worked out from publicly available hardware rental prices and specs — they are not a formal quote, nor an official price set by Anthropic, Nvidia, or any cloud provider. Actual costs vary significantly by region, vendor, and contract terms; readers assessing their own situation should get an actual quote.
Q4: What's the difference between "100 users" and "100 concurrent requests," and why does it matter so much?
"Number of users" is how many people could potentially use the system; "concurrent requests" is how many requests the system actually has to process at any given moment — and the two can differ enormously. The article's example: if 100 employees send requests at an even, spread-out pace, the number running at the same instant might be as low as around 5, far below the intuitive number "100." This distinction directly determines how much hardware you actually need to buy, and it's the point in this article most likely to be misread — and most likely to lead to over-buying.
Q5: Instead of guessing at hardware specs, how should a company actually decide whether to move AI work to local infrastructure?
The source video's suggestion is: don't rush to place a purchase order — rent hardware for a couple of weeks and see how real users actually respond, which tells you more than any spec sheet or benchmark chart. And before even renting to test, the more fundamental first step is to get a clear picture of your team's actual token usage and concurrent request patterns — you need to know how much you're really using, and where, before you're in a position to decide whether trading that cost for the operational burden of an entire data center is worth it.
Source Note: This article is adapted from a video published by the YouTube channel Kai, titled "What Would It Cost to Run Claude Fable 5 Locally?", rewritten in narrative form and not a word-for-word translation. The protagonist in this piece, "Wen Zixuan," is a representative figure used for explanation, not a real customer case; his company's background and situation are an illustrative reconstruction. All specific figures in the article (memory conversions, GPU specs, power draw, rental cost estimates) are order-of-magnitude estimates the source video's creator worked out from publicly available specs, not data independently verified or audited by this site or any third party, nor official pricing. Readers should rely on actual hardware specs and vendor quotes when assessing their own situation.
Start Now
Whether you keep using the API or start considering self-hosting, the thing you actually need to figure out first isn't "how many GPUs to buy" — it's "how much are we actually using right now." Instead of guessing at fixes for a headache you haven't diagnosed, lay out the real token usage for every model and every task first. Try AI Token King free today, and let us help your team lay out its real usage, so you can decide with data whether to trade it all for an entire data center — not with a line like "the company next door is doing it."
