One evening this September, the host of Fireship's The Code Report sat down and did some math on his AI bill.
Cursor: $20 a month. Claude Max: $100. GPT Pro: another $100. Gemini Ultra: another $100. That's $320 already โ before counting an 11 Labs voice subscription, DeepSeek, and a pile of API keys he'd signed up for two years earlier and forgotten to cancel. He joked, half-seriously, that his "AI addiction" had gotten so expensive he'd had to give up alcohol just to afford it, and cracked that U.S. alcohol sales were sitting at historic lows because "Big AI" was to blame. That line is obviously a punchline, not an economic claim โ but the underlying feeling, subscriptions quietly stacking on top of subscriptions until the total stops making sense, is one plenty of developers juggling multiple AI tools will recognize.
What actually stopped him wasn't any single charge. It was seeing all four subscriptions laid out at once โ nobody sets out to spend $320 a month, it just accumulates one sign-up at a time until the day you open the invoice. That moment became the premise of his video: instead of adding a fifth subscription, he moved his entire AI workflow onto his own server and rebuilt it with free, open source tools โ cheaper, in his telling, and more productive besides. Here are the five tools he wired together, and the specific problem each one solves.
Keeping prompts private starts with Ollama
Every AI stack begins with the same question: where does the model actually run? For most people, the default answer is to send proprietary code straight to a trillion-dollar company and hope it's handled carefully. Ollama offers a different path โ run the model on your own machine, full stop.
The host describes it as "basically Docker for large language models": a simple command-line interface and API for downloading and running open-weight models, including the various Chinese models that have shown up recently. Running locally has an obvious upside โ prompts never leave your machine, inference costs nothing, and the model keeps working even if your credit card gets declined.
There's a real limit, though: most people simply don't own hardware capable of running anything close to a frontier-scale model. Small models will run almost anywhere, but getting near state-of-the-art performance requires something closer to a small data center. For a developer who still wants top-tier output, that's the next problem to solve.
When your subscriptions max out, route around it
Since local hardware can't carry the heaviest workloads alone, some traffic still has to go to hosted models. Here the host brings in a self-hosted routing layer he refers to (per the transcript) as "Nine Router" โ a proxy that sits between your AI tools and dozens of model providers, all reachable through a single local endpoint.
The problem it solves is concrete: rather than juggling nine separate API keys, you get one door โ an OpenAI-compatible proxy that every tool can point at localhost. What the host finds most useful is its tiered fallback design. Tier one is whatever subscription you're already paying for, like Claude Max. Tier two is a cheap, pay-as-you-go model held in reserve. Tier three covers every free option available โ Chinese models, cloud trial credits, and so on. Once tier one is maxed out, the router automatically falls back to the next tier with no manual switching required. On top of that, it tracks usage and compresses tool output to cut token consumption โ handling both "don't blow the budget" and "don't run out of quota early" at once.
The real culprit behind the bill: wasted tokens
If token usage still looks alarming even with routing in place, the next thing the host points to is context bloat โ illustrated with a vivid example: you ask an AI agent to center a `div`, and to do that, it reads through 50,000 lines of your `package.json`, burning through the computational equivalent of evaporating an Olympic-size swimming pool, only to conclude it needs to install Tailwind CSS.
Headroom is built to sit between your application and the model provider, compressing tool output, log files, and other unnecessary content before it ever becomes a billable input token. The clever part is that the compression is reversible โ compressed content is cached locally, so if the model ever actually needs the original data, it can still retrieve it. Nothing is permanently thrown away.
Wiring the tools together still leaves the question of where they run
At this point the host hits a practical wall: once all of these tools are connected, where do they actually get deployed? In the video, he mentions Hostinger โ the video's sponsor โ and its virtual private server with a built-in Docker app catalog that can install the open source projects mentioned above with a single click, all running together on the same machine. This section is a sponsored segment; readers should evaluate pricing and alternatives independently, and this article doesn't repeat the discount code or referral link.
Infrastructure is only half the story โ Dify is where it becomes an app
Everything up to this point is infrastructure. To build something people actually use, the host turns to Dify โ a visual workflow builder that lets you drag and connect nodes instead of engineering every piece of logic through raw prompts. His example is deliberately tongue-in-cheek: an app called "Horse Tinder," which sends each horse's profile to Dify, retrieves compatible matches from a database through a visual workflow, and uses a large language model to generate a match explanation โ before exposing the entire workflow as an API that the front end calls whenever a user swipes. The joke isn't really about horses; it's about the fact that a workflow like this can become a callable API without writing a traditional backend.
The last step: replacing yourself
By the end of the video, the host says โ half-joking, half-not โ that writing code by hand started feeling outdated after Q2 2026, once OpenHands arrived. It's an open source autonomous coding agent, described as a top-performing tool on SWE-bench Verified, the benchmark that measures whether models can resolve real GitHub issues. Instead of fixing a bug yourself, you open a GitHub issue and let OpenHands handle it.
In practice, it works as a command center: a self-hosted fleet of AI agents running continuously in the background on your own server. Because everything runs on infrastructure you control, you can point it at OpenAI or Anthropic models, or route it through the local models you already set up with Ollama. That's the real turning point of the whole video โ it starts as "I don't want to pay for one more subscription" and ends as "the entire development pipeline is mine to run, and the model behind it can change any time I want."
Subscribe or self-host was never really the question
Looking at all five tools together, the problem the host actually solved wasn't "open source is cheaper" โ that's too thin a takeaway. It was "where, exactly, is my money and my prompt data actually going?" โ the question most people only ask once they've stacked enough subscriptions to notice. Self-hosting isn't free in the way it sounds; it costs time to maintain servers and work around hardware limits. What it buys instead is visibility: every call, every dollar, accounted for.
Whichever path you take, the underlying logic is the same โ you can't make a real choice about your AI spend until you can actually see it laid out. That's part of why interest keeps growing not just in self-hosted, open source stacks, but also in tools like AI Token King that consolidate usage and billing across multiple model providers into one place. The two approaches solve different problems โ one drives down the cost of each call, the other clarifies where the money is going in the first place โ but they both start from the same premise: you need a bill you can actually read before you can decide what to do about it.
FAQ
Q1: Can I still use Ollama without a powerful local GPU? Yes. Small models will run on almost any machine; it's only frontier-scale models that require hardware closer to a small data center. Most people can start with a smaller model and decide later whether hardware investment is worth it.
Q2: Is self-hosting an AI stack actually cheaper than paying for subscriptions? The video reflects the host's own experience of canceling four subscriptions and switching to self-hosting, without providing a direct cost comparison. Whether it pays off depends on your actual usage, your ability to maintain the infrastructure, and the hardware you're willing to invest in โ this is worth evaluating case by case rather than assuming someone else's outcome applies to you.
Q3: What do the "fallback tiers" in Nine Router actually mean? As described in the video, provider access is split into three tiers: tier one is a subscription you already pay for (like Claude Max), tier two is a cheap pay-as-you-go model, and tier three covers free providers. When the current tier is exhausted, the system automatically falls back to the next one without manual intervention.
Q4: How is a visual workflow tool like Dify different from just writing prompts? The difference is in structure and maintainability. Dify lets you assemble steps โ data retrieval, model calls, output formatting โ by connecting nodes visually, and can expose the entire workflow as a callable API, rather than compressing all of that logic into a single prompt. That structure tends to be easier to maintain for anything called repeatedly or exposed as a service.
Q5: Can an autonomous coding agent like OpenHands fully replace engineers? The video describes it as a top-performing autonomous coding agent on SWE-bench Verified, capable of handling GitHub issues and running continuously in the background โ but it doesn't claim to replace human engineering judgment or review. Keeping a human in the loop for oversight is still advisable in practice.
Source
This article is adapted from a video published by the YouTube channel Fireship on September 7, 2026, titled "5 open source tools that replaced my $320/mo AI stack..." (view count verified at the time of writing: 1,247,954). Original video: https://www.youtube.com/watch?v=Y5rSSvXfL4g (https://www.youtube.com/watch?v=Y5rSSvXfL4g). This article is reorganized and rewritten directly from the original English transcript โ it is not a translation of any other article, and it does not reproduce the transcript verbatim. Tool names, cost figures, and scenarios described here are drawn from the transcript as stated, without adding numbers the transcript didn't provide. The name "Nine Router" reflects the automatic speech-recognition transcript as generated; the transcript text itself ("Instead of juggling nine different API keys, Nine Router provides one door") suggests this may be an intentional pun on "nine API keys" rather than a simple mis-transcription, but the official product name and spelling have not been independently confirmed and should be verified before republication. The segment about U.S. alcohol sales and AI spending is the video host's own exaggerated joke, not an economic claim made by this site.
Further Reading
Related coverage on aitoken.com.tw (Chinese-language, not yet available in English) includes pieces on what it actually costs to self-host a frontier-scale AI model, what an AI gateway like OpenRouter does, when prompt caching can make an AI bill go up instead of down, and why a shocking AI invoice is often just a mirror held up to actual usage.
Try It Now
Whether you end up wiring together all five of these tools yourself, or just want to see where your own AI spending is actually going every month, the first step is the same: get a record you can actually read. Try AI Token King for free (https://aitokenking.com.tw/home) and bring the usage and billing scattered across all your different accounts into one place.
