# Don't Buy a GPU Instead Build This 32GB Free AI Server Instead
Canonical: https://social-archive.org/jack_bauer111/UAZ1kRgMqZ
Original URL: https://www.youtube.com/watch?v=IMVjEsKX5UQ
Author: Kai
Platform: youtube
Share mode: full
## Content
Don't buy a GPU yet. Kaggle gives you a Free GPU AI server with 32GB VRAM (two Tesla T4 cards), and every tutorial calls it the best Free AI setup for Local AI coding. I built the AI Server, ran a 27B Model on it with Ollama, and found out what the thumbnails leave out. In this video we look at the viral free Kaggle GPU "32GB AI server", the Ollama Tutorial most people followed, and why two 16GB Tesla T4 cards run a 27B Model at the speed of one GPU. Then we price the free quota in real tokens, and open the free TPU on the same Kaggle menu that runs the same Qwen model about eight times faster. WHAT YOU'LL LEARN - How the free Kaggle GPU setup works: a Kaggle notebook, Ollama, a Cloudflare tunnel, and a coding agent like OpenCode or Hermes Agent - Why 32GB VRAM split across two Tesla T4 GPUs gives you one card's speed (Ollama layer split, about 16 tokens per second) - Why AI Coding agents stall and time out: prefill, 30,000-token system prompts, and Cloudflare's 100-second limit - The real limits: about 30 GPU hours a week, 12-hour sessions, and a 20-minute idle timer - What a week of free quota is worth next to the hosted price of the same Large Language Model - The free Kaggle TPU v5e-8: the Qwen 27B Model at full 16-bit precision, about 130 tokens per second, with an API key - Kaggle's rules, why this doesn't really run AI locally, and when Self Hosted AI on your own GPU still makes sense CHAPTERS 0:00 A Free GPU vs a $5,000 RTX 5090 1:31 How the free 32GB VRAM Kaggle AI Server works 3:02 Two Tesla T4 GPUs, the speed of one 4:34 Prefill, AI Coding agents and the 100-second timeout 6:34 It's a quota, not a GPU 7:16 What a free week is worth in tokens 8:18 The free Kaggle TPU nobody picked 10:26 The catches: Kaggle rules, Local AI and security 12:17 Who the free AI server is actually for 14:11 Verdict: Free GPU, hosted LLM, or your own GPU KEY NUMBERS (checked October 11–12, 2026) - Model: Qwen3.8 27B, a 27B Model, Apache-2.0 Open Source AI, 262K context - NVIDIA Tesla T4: Turing (2018), 16 GB GDDR6, 320 GB/s, 65 FP16 TFLOPS - Kaggle "GPU T4 x2": two T4s, about 30 GPU hours a week, 12-hour sessions, 20-minute idle timeout - Ollama on two T4s: 16.23 tokens per second measured (DevZoneX) - Kaggle "TPU VM v5e-8": 8 chips x 16 GB HBM, about 130 tok/s decode and 10,300 tok/s prefill (kaggle-tpu-lab) - Hosted Qwen3.8 27B: $2.46 per million output tokens (OpenRouter weighted average, Oct 11; prices move daily) WHO THIS IS FOR If you're an AI Developer learning how a Local LLM server works, the free Kaggle setup is a great AI Tutorial sandbox: install Ollama, split a model across two GPUs, set up a tunnel, break it, fix it. If you code every day with AI Coding Tools, a hosted LLM costs less than your waiting time. And if you already pay for ChatGPT or other AI Tools and want Self Hosted AI because your code can't leave your machine, that is the day you buy the GPU. This is a practical Artificial Intelligence breakdown for anyone who wants to run AI locally without guessing: real numbers on the Kaggle GPU, the Large Language Model, and what "free" costs you in time. SOURCES Kaggle notebook limits: https://www.kaggle.com/docs/notebooks Kaggle Acceptable Use Policy (June 22, 2025): https://www.kaggle.com/aup NVIDIA Tesla T4 specs: https://www.nvidia.com/en-us/data-center/tesla-t4/ Google Cloud TPU v5e specs: https://cloud.google.com/tpu/docs/v5e Ollama FAQ (multi-GPU loading): https://docs.ollama.com/faq Cloudflare Quick Tunnels: https://developers.cloudflare.com/cloudflare-one/networks/connectors/cloudflare-tunnel/do-more-with-tunnels/trycloudflare/ Cloudflare error 524: https://developers.cloudflare.com/support/troubleshooting/http-status-codes/cloudflare-5xx-errors/error-524/ llama.cpp CUDA scoreboard (T4 row): https://github.com/ggml-org/llama.cpp/discussions/15013 kaggle-tpu-lab: https://github.com/ARahim3/kaggle-tpu-lab Qwen3.8-27B model card: https://huggingface.co/Qwen/Qwen3.8-27B OpenRouter Qwen3.8 27B: https://openrouter.ai/qwen/qwen3.8-27b Claude Code token overhead report: https://github.com/anthropics/claude-code/issues/52979 DevZoneX build: https://www.youtube.com/watch?v=Lj4fX4TVKw0 STUD TY build: https://www.youtube.com/watch?v=EkA4pqXgta0 Hyperautomation Labs measurements: https://www.youtube.com/watch?v=tSx26jgNyE0 Reddit "gray area" thread: https://www.reddit.com/r/LocalLLM/comments/1vg1adc/ #LocalAI #Ollama #Kaggle --- ## Transcript [0:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=0s) Okay, so picture this. It's 2:00 a.m. [0:03](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=3s) I'm sitting at my desk with my coffee [0:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=5s) gone cold and my laptop fan sounds like [0:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=7s) it is trying to take off from a runway [0:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=9s) all because I'm trying to run a halfway [0:11](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=11s) decent AI model locally. And it's not [0:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=14s) even working since it is putting out two [0:16](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=16s) tokens per second, which is not even a [0:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=18s) speed so much as a medical condition. [0:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=21s) And then I made the mistake of opening [0:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=22s) Twitter where someone was flexing their [0:24](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=24s) RTX 5090 setup with something like, "Oh, [0:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=27s) yeah, 200 tokens a second, silky smooth, [0:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=31s) totally worth the five grand." [0:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=33s) But five grand for a graphics card is [0:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=35s) real money that I just don't have [0:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=37s) sitting around to make an AI go brrr on [0:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=39s) my laptop. So, I went down the YouTube [0:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=41s) rabbit hole like any reasonable person [0:43](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=43s) does and that is where I found the [0:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=45s) video. You've probably seen it, too, [0:48](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=48s) since it's called build a free 32 GB AI [0:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=51s) server with two Tesla T4s, $0. It has [0:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=55s) 260,000 views and the comments are going [0:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=57s) crazy with lines like "This changed my [1:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=60s) life. I can't believe this is free and [1:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=62s) my RTX 3090 is crying." And honestly, I [1:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=65s) believed that enough to spend an entire [1:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=67s) Sunday setting it up, getting it [1:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=69s) running, and feeling like a genius right [1:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=72s) before reality hit like a cold shower. [1:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=74s) Now, that is exactly what we are going [1:16](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=76s) to talk about today. There is a lot more [1:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=79s) to the story than the thumbnail tells us [1:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=81s) including something on that exact same [1:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=83s) free menu that literally nobody picked [1:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=86s) and it absolutely destroys the T4 setup. [1:29](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=89s) So, let's get into it. So, first, let's [1:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=92s) walk through what this whole free 32 gig [1:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=95s) AI server thing actually is in case you [1:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=97s) haven't seen those videos. There is a [1:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=99s) platform called Kaggle, which is [1:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=101s) Google's data science playground and [1:43](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=103s) basically where data nerds go to train [1:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=105s) models and compete in ML competitions. [1:48](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=108s) And Kaggle has this absolutely wild free [1:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=111s) tier where you verify a phone number, [1:53](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=113s) just a phone number, and get about 30 [1:56](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=116s) GPU hours per week for free with no [1:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=119s) credit card, no trial, nothing. And one [2:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=122s) of the options on that menu is GPU T4 [2:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=125s) X2, which is two Nvidia Tesla T4 cards [2:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=129s) at 16 gigs each. So, 32 gigs of VRAM [2:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=132s) total. Now, on that hardware, people [2:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=134s) have been running Qwen 3, specifically [2:16](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=136s) the 27 billion parameter version. And [2:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=138s) that is a serious model, Apache [2:20](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=140s) licensed, out since August, and [2:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=142s) downloaded 6.8 million times on Hugging [2:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=146s) Face in a single month. So, it can code, [2:29](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=149s) it can reason, and it is the real deal. [2:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=151s) The workflow this video show is to spin [2:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=153s) up the Kaggle notebook, download the [2:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=155s) model through Ollama, punch a hole [2:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=157s) through Cloudflare to make it public, [2:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=159s) and paste that URL into a coding agent [2:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=162s) like Cloud Code or Hermes. At that [2:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=164s) point, we have a free 32-gig AI server [2:46](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=166s) that we can point a whole workflow at. [2:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=169s) And the comparison they always make is [2:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=171s) free versus a $1,300 graphics card. So, [2:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=175s) free wins before the video even starts, [2:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=177s) right? [2:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=178s) That is also exactly where the story [3:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=180s) starts to get interesting. I remember [3:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=182s) watching the benchmark and Dev Zone X [3:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=184s) video, which showed 16 tokens per [3:06](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=186s) second, and thinking it wasn't blazing [3:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=188s) fast, but it was workable enough to ship [3:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=190s) code with. But once I actually started [3:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=192s) using it as a daily coding server, [3:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=195s) something started to feel off. [3:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=197s) Two GPUs sounds like two GPUs. So, [3:20](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=200s) wouldn't you expect double the power or [3:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=202s) at least twice the speed of one card? [3:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=205s) But that is not what actually happens [3:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=206s) here. The model file for Qwen 3 27B at [3:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=210s) 4-bit quantization is about 17 GB, while [3:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=213s) each T4 on Kaggle gives us 15.3 GB of [3:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=217s) usable VRAM. So, the file doesn't fit on [3:40](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=220s) one card, and what Ollama does instead [3:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=222s) is split the model across both cards by [3:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=225s) layers, where half the model lives on [3:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=227s) card one and half on card two. Now, to [3:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=229s) generate one token, card one reads its [3:52](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=232s) half of the weights while card two sits [3:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=234s) there doing nothing. Then it hands off [3:56](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=236s) the result and card two reads its half [3:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=239s) while card one sits there doing nothing [4:01](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=241s) and only then do we get one token. It's [4:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=244s) a relay race, not a team lift. Engineers [4:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=247s) call this a layer split, which means two [4:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=249s) T4s give us the capacity of two cards [4:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=252s) and the speed of exactly one. The T4 [4:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=254s) itself launched in 2018 and it moves 320 [4:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=258s) GB per second through its memory bus, so [4:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=261s) that is really our ceiling here and [4:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=263s) DevZoneX measured 16.23 tokens per [4:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=266s) second with multi-token prediction on [4:29](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=269s) and one commenter said they could not [4:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=270s) get past 18, which is not a coincidence, [4:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=273s) it is physics. Now, 16 to 18 tokens per [4:36](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=276s) second for writing code is survivable [4:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=279s) and I've used worse, but the part that [4:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=281s) absolutely destroyed me when I figured [4:43](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=283s) it out is the thing nobody in those [4:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=285s) tutorial videos mentions. And every time [4:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=287s) I'd fire up my coding agent and give it [4:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=289s) a task, [4:50](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=290s) there'd be this pause. At first I blamed [4:53](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=293s) the tunnel or maybe my internet being [4:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=295s) weird, so I refreshed and waited and [4:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=298s) nothing happened until the words [4:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=299s) suddenly showed up. What's happening [5:01](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=301s) during that silence is called prefill. [5:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=304s) Before a model writes a single word [5:06](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=306s) back, it has to read everything we sent [5:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=308s) it, which includes the message, the [5:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=310s) attached files, the previous [5:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=312s) conversation and the agent's system [5:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=315s) prompt. And the system prompts in tools [5:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=317s) like Claude Code are reportedly around [5:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=319s) 30,000 tokens before anyone has even [5:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=321s) typed a single word. Reading is a [5:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=323s) different kind of work than writing, [5:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=325s) since it's math-heavy rather than [5:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=327s) memory-bound, and the T4 has 65 [5:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=330s) teraflops of compute for this. So when [5:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=332s) we scale the numbers for a 27B model, we [5:34](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=334s) are looking at roughly 325 tokens per [5:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=337s) second during prefill. That puts a [5:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=339s) 30,000 token agent prompt at about 92 [5:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=342s) seconds of complete silence before the [5:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=344s) first word comes out. And the beautiful [5:46](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=346s) part is that Cloudflare's quick tunnel, [5:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=349s) the one every single one of these [5:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=351s) tutorials uses, automatically kills any [5:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=354s) request that hasn't responded in 100 [5:56](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=356s) seconds. That is the timeout limit on [5:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=358s) their free tier, and it can't be changed [6:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=360s) below their enterprise plan. So, we've [6:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=362s) got 92 seconds of prefill and 100 [6:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=365s) seconds before Cloudflare gives up and [6:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=367s) throws a 524 error, which leaves us 8 [6:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=370s) seconds of margin. One of the creators, [6:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=372s) A Stud TY, literally says in their own [6:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=375s) video that the first message might time [6:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=377s) out and to just send it again. And if we [6:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=379s) scroll through the comments on these [6:20](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=380s) videos, there is a greatest hits [6:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=383s) collection of 502 errors, 504 errors, [6:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=387s) and provider unavailable. So, it is not [6:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=390s) actually broken. It's just math on a [6:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=392s) whiteboard that nobody drew for us. [6:34](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=394s) And this is where I had my own no [6:36](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=396s) moment, since I was actually planning to [6:38](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=398s) use this for my day-to-day work. I've [6:40](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=400s) been thinking about it as a GPU, when [6:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=402s) really it is a quota. [6:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=404s) We get about 30 GPU hours a week, and [6:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=407s) Kaggle caps a single session at 12 [6:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=409s) hours. And if we walk away from the [6:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=411s) laptop and the session goes idle, which [6:53](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=413s) happens in as little as 20 minutes, the [6:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=415s) whole thing dies, including the model [6:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=417s) download, Ollama, the Cloudflare tunnel, [7:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=420s) and the URL we pasted into the agent. [7:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=422s) And the next time we start, we get a [7:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=424s) completely new URL, so the config needs [7:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=427s) updating again. [7:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=428s) Some people in the comments say each [7:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=430s) cold start costs them 15 minutes, and [7:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=432s) some say 30, and those minutes come out [7:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=434s) of a 30-hour weekly quota, too. So, I [7:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=437s) did the math the thumbnails don't want [7:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=438s) us to do. Now, 30 hours a week is [7:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=441s) 108,000 [7:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=443s) seconds. At 16.23 tokens per second, if [7:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=447s) the server was generating tokens [7:28](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=448s) literally every single second with no [7:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=451s) setup time, no idle time, and no reading [7:34](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=454s) time, it would produce about 1.75 [7:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=457s) million tokens in a week. Now, that same [7:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=461s) model, Qwen 327B, is sold on OpenRouter, [7:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=465s) where about 20 different companies host [7:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=467s) it. And the average price customers [7:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=469s) actually paid is $2.46 per million [7:52](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=472s) output tokens, and 1.75 million tokens * [7:56](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=476s) $2.46 [7:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=477s) is $4.31. [7:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=479s) So, our free server, at the absolute [8:01](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=481s) best case with every second perfectly [8:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=484s) optimized, is worth $4.31 a week. And [8:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=487s) the $4.31 version doesn't time out, [8:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=490s) doesn't need a phone number, and doesn't [8:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=492s) die when we close a tab. That is the [8:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=494s) comparison I really needed to see from [8:16](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=496s) the beginning. Now, this next part [8:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=499s) genuinely annoyed me when I discovered [8:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=501s) it, because it's sitting right there on [8:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=502s) the same page, and somehow nobody in the [8:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=505s) whole wave of 32 gig server videos [8:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=507s) picked it. If we scroll down on the [8:29](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=509s) Kaggle accelerator menu, below the T4 [8:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=512s) pair, there is another free option that [8:34](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=514s) says TPU VM. That is eight of Google's [8:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=517s) own TPU chips with 16 gigs of HBM each, [8:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=521s) so 128 GB total. And each chip moves [8:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=524s) about 800 GB per second, 2 and 1/2 times [8:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=527s) the T4's memory bandwidth, with 197 [8:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=531s) teraflops of 16-bit math per chip. [8:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=534s) Until August, there was actually a good [8:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=535s) reason nobody was using it for this, [8:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=538s) because Qwen-3 uses a type of attention [9:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=540s) that TPU software could not handle. But [9:03](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=543s) in August, vLLM's TPU backend added [9:06](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=546s) support for it in version 0.28.0. [9:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=550s) And shortly after, a project called [9:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=552s) Kaggle TPU Lab put everything together, [9:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=555s) and it is MIT licensed and sitting on [9:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=558s) GitHub with about 579 stars. What it [9:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=561s) runs is Qwen-327B [9:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=563s) at full 16-bit precision, not the 4-bit [9:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=566s) quantized version that fits on the T4s, [9:28](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=568s) but the full 55.6 GB model with the full [9:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=572s) 262,000 token context window. [9:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=575s) And the numbers from the project's own [9:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=577s) benchmarks are kind of embarrassing for [9:39](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=579s) the T4 setup. Writing speed is around [9:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=582s) 130 tokens per second against the T4 [9:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=585s) pair 16, so that is eight times faster. [9:48](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=588s) Prefill, the reading speed, comes in at [9:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=591s) about 10,300 tokens per second against [9:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=594s) roughly 325 on the T4 pair, which makes [9:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=598s) it 30 times faster. [10:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=600s) So that 30,000 token agent prompt that [10:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=602s) takes 92 seconds of silence on the T4s [10:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=605s) takes about 3 seconds on the TPU. It [10:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=608s) also hands out an API key, so the [10:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=609s) endpoint is actually secured, and the [10:11](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=611s) whole thing is live about 22 minutes [10:13](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=613s) after pressing run. And it's the same [10:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=615s) free quota page, the same phone [10:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=618s) verification, and the same notebook [10:20](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=620s) format. Just a different drop-down menu [10:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=622s) option that literally nobody in 260,000 [10:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=625s) views clicked. Now, to be fair, this is [10:28](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=628s) not about swapping one set of hype for [10:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=630s) another because both of these setups, [10:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=633s) the T4s and the TPU, come with real [10:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=635s) catches that nobody reads out in any of [10:38](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=638s) these tutorials. The first catch is [10:40](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=640s) Kaggle's acceptable use policy. The [10:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=642s) version in force since June 2025 [10:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=645s) explicitly lists server farming and [10:48](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=648s) activity unrelated to ML data science as [10:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=651s) abuse of their resources. And when it [10:53](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=653s) comes to wiring a Cloudflare tunnel to a [10:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=655s) notebook and running a keep-alive script [10:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=657s) to dodge the idle timer, one Reddit [11:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=660s) builder who published exactly this setup [11:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=662s) called it a gray area. We could not find [11:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=665s) a documented ban, but it is still your [11:07](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=667s) account and it's still their rules. [11:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=669s) The second catch is that it is not [11:11](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=671s) local, even though every thumbnail says [11:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=674s) free local AI. The moment the code goes [11:16](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=676s) through a Cloudflare tunnel to Google [11:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=678s) servers, it stopped being local. And [11:20](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=680s) Cloudflare themselves say quick tunnels [11:22](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=682s) are for testing and development with no [11:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=685s) uptime guarantee. That matters because [11:28](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=688s) the one real reason to run AI locally is [11:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=691s) keeping code private and off someone [11:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=692s) else's machine. And that is exactly what [11:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=695s) this setup can't give us. [11:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=697s) One commenter summed it up perfectly [11:38](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=698s) with free local AI look inside VPS, [11:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=702s) which is really just a VPS with extra [11:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=704s) steps. [11:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=705s) The third catch is that the defaults on [11:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=707s) the T4 tutorials are genuinely kind of [11:50](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=710s) chaotic. Ollama has no login out of the [11:52](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=712s) box, and most tutorials put it behind a [11:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=715s) public link with no API key, so the URL [11:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=718s) is essentially the password. And the [12:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=720s) model most of them pull is an uncensored [12:02](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=722s) reupload from a single hugging face [12:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=724s) account rather than Qwen's official [12:06](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=726s) file. And the TPU project also wants a [12:09](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=729s) Kaggle API token to launch from a [12:11](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=731s) terminal, but it's open source, so we [12:13](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=733s) can read every line before handing that [12:15](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=735s) over, and we should. [12:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=737s) Now, I've been sitting on this for a [12:18](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=738s) while, and I think the honest answer is [12:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=741s) more nuanced than the thumbnail suggest. [12:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=743s) The free Kaggle setup, T4s or TPU, is [12:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=747s) genuinely perfect for learning. If the [12:29](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=749s) goal is understanding how an inference [12:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=751s) server works, how to split a model [12:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=753s) across cards, how to configure Ollama, [12:36](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=756s) and how to set up a tunnel, this is an [12:38](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=758s) amazing sandbox. We install the engine, [12:41](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=761s) watch things break, and fix them. And [12:43](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=763s) nothing we own gets hot, which is [12:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=765s) genuinely valuable. It is also what [12:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=767s) Kaggle was built for, which is training [12:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=769s) and evaluating models. Batch jobs, [12:52](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=772s) fine-tuning, running 500 prompts through [12:55](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=775s) a model for an eval, and saving the [12:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=777s) output aren't a gray area at all. [12:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=779s) Because that is the exact thing the [13:01](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=781s) platform is designed around. And for a [13:04](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=784s) student with no budget and no credit [13:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=785s) card, the TPU route gives something [13:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=788s) real. A strong coding model running at [13:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=790s) 130 tokens a second, API key and all, 20 [13:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=794s) hours a week, completely free. For a lot [13:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=797s) of people in a lot of places, that is [13:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=799s) not a rounding error, but a meaningful [13:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=801s) tool. But for a daily coding workflow [13:24](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=804s) that just works, where we open the [13:26](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=806s) laptop at 9:00 a.m. and the agent [13:27](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=807s) responds immediately, a full week of [13:30](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=810s) this free server is worth less than [13:32](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=812s) lunch. The hosted version of the same [13:34](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=814s) model, on Cerebras, for example, runs at [13:37](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=817s) 618 tokens a second for $1.49 [13:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=822s) per million tokens. At that price, the [13:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=824s) 30-hour weekly Kaggle quota, converted [13:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=827s) to tokens, costs less than five bucks. [13:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=829s) So, buy $5 of credit and move on. [13:52](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=832s) And if the real reason for wanting local [13:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=834s) AI is that the code literally cannot [13:57](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=837s) leave the machine because of security, [13:59](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=839s) compliance, an NDA, whatever, then every [14:03](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=843s) option in this video is out, including [14:05](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=845s) the free ones. That is the actual reason [14:08](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=848s) to own hardware, and it's a completely [14:10](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=850s) legitimate reason. So, after all of [14:12](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=852s) this, where I actually landed is pretty [14:14](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=854s) simple. If you want free, skip the 2T4s, [14:17](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=857s) run the TPU recipe, and treat it like a [14:19](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=859s) lab you open for an afternoon rather [14:21](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=861s) than a server you leave running. If you [14:23](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=863s) code every day, point your harness at [14:25](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=865s) the hosted model because a full week of [14:28](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=868s) the free server is worth $4.31. [14:31](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=871s) That is two cups of coffee, and one of [14:33](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=873s) them will go cold while the agent prompt [14:35](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=875s) pre-fills. And buy the GPU the day your [14:38](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=878s) code is not allowed to leave your desk [14:40](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=880s) because until that day, free is just [14:42](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=882s) someone else's computer with a timer on [14:44](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=884s) it. If this broke the way you were [14:45](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=885s) thinking about this whole thing, then [14:47](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=887s) good. That was the point. Drop a comment [14:49](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=889s) if you've actually tried either setup [14:51](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=891s) because I want to know what your real [14:53](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=893s) numbers looked like. And if you haven't [14:54](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=894s) subscribed yet, that is fine, but you [14:56](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=896s) probably should since next time we are [14:58](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=898s) going to look at what actually makes [15:00](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=900s) sense for a coding workflow when you're [15:01](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=901s) not made of money. So, I'll see you [15:03](https://www.youtube.com/watch?v=IMVjEsKX5UQ&t=903s) then.