What I actually spend money on each month

People ask me what running AI automation “actually costs” a lot more than they ask what it does. That’s a fair question. Most of what gets written about AI spend is either a vendor pitch or a doom post about API bills spiraling out of control. Neither matches what I actually see on my own statements each month. So here’s the real list, categorized by what the money is buying, not by brand name.

I’m not going to give you exact dollar figures for every line item, because my setup isn’t your setup and a number without context is just a number. What I can do is walk through the categories, explain why each one exists, and be honest about which ones I could cut and which ones I can’t.

Compute is the biggest bucket, and it’s split two ways

The single largest recurring cost isn’t any AI API. It’s infrastructure: servers that stay on whether or not a model is running on them. I run a mix of local hardware and cloud instances, and the split matters because the two do different jobs.

Local machines handle anything that’s repetitive, high-volume, or doesn’t need frontier reasoning. Rendering video, running batch jobs, hosting internal tools, that kind of thing. The electricity and hardware amortization for that is a real cost, it just doesn’t show up as a monthly invoice the way a SaaS subscription does. If you only count subscription line items, you undercount your actual spend, because the hardware sitting in a rack is still capital you paid for and power you’re paying for every day it’s on.

Cloud compute picks up the rest: things that need to scale up and down, things I don’t want to babysit, things where uptime matters more than unit cost. Cloud is more expensive per unit of work than local hardware in almost every case I’ve measured, but you’re not paying for the unit of work, you’re paying to not manage a server rack yourself for that particular job. That tradeoff is the whole game with cloud versus local, and anyone telling you one is unconditionally better than the other hasn’t run both long enough.

API spend is real but smaller than people assume

Model API calls get treated like they’re the dominant cost of running AI, and in my case they aren’t. They’re a meaningful line, but they sit behind compute and, honestly, behind some of the plumbing tools I’ll get to below.

The reason API spend doesn’t run away on me is that I don’t route everything through a frontier model by default. Cheaper, smaller models handle classification, formatting, and routine text transformation. The expensive model gets called when a task actually needs the reasoning it’s paying for, not as the default path for every request. That routing decision is a design choice, not a discount code, and it’s the single biggest lever I have over this cost category.

I want to be clear about one thing here: none of the tools I use for this are free. Some have free tiers with real usage limits, and I’ve hit those limits and paid to go past them. If a tool advertises itself as free, check what the free tier actually covers before you build a workflow on top of it, because the workflow breaking when you hit a quota is a worse day than budgeting for the cost up front.

Agent orchestration and MCP tooling

This is the category that’s grown the most for me in the last year, and it’s the one people ask about least. Once you move from “call a model, get text back” to “let an agent use tools to actually do something,” you’re paying for more than tokens. You’re paying for whatever runs the orchestration layer, whatever hosts your MCP servers if they’re not local, and whatever logging or observability you bolt on so you can tell what the agent actually did when something goes wrong.

That last part isn’t optional if you’re running agents against real systems. An agent that can send a message, write to a database, or trigger a deployment needs a paper trail, because the failure mode isn’t “it says something dumb,” it’s “it did something you didn’t intend and you don’t know why.” The tooling for that costs money and time to set up. I’d rather pay for it up front than debug a production incident with no logs.

MCP itself, as a protocol, doesn’t cost anything to adopt. What costs money is everything downstream of it: the servers you stand up to expose your tools, the compute those servers run on, and the engineering time to keep the connectors working when an underlying API changes shape. Protocol adoption is free. Running a reliable integration on top of it is not.

The unglamorous subscriptions that add up

Password and credential management, monitoring and alerting, backup storage, a handful of SaaS tools for things like scheduling or file storage. None of these are AI tools. All of them are load bearing. If I stopped paying for backups tomorrow to save money, the first time a drive failed I’d lose more value than years of that subscription cost. This is where “cut costs” advice gets dangerous if you apply it without thinking about what the tool is actually protecting you from.

I don’t treat this category as a place to optimize aggressively. The savings from cutting a ten dollar a month monitoring tool are trivial next to the cost of not noticing a system went down for six hours.

What I don’t pay for

Worth naming a few things I’ve deliberately avoided. I don’t pay for every AI tool that launches with a good demo. Most of them solve a problem I don’t have, or solve a problem I’ve already solved with something cheaper. I also don’t pay for enterprise tiers of tools built for teams of ten when I’m the only person using the account. Feature gates that exist to charge organizations more money don’t help a solo operator, and it’s easy to get talked into paying for headroom you’ll never use.

I also don’t treat any of this spending as an investment in the sense of expecting a guaranteed return. It’s an operating cost, the same category as rent or a phone bill. I’m not going to tell you what to spend your money on or claim any of this will pay for itself on a schedule, because that depends entirely on what you’re building and I don’t know your numbers. What I can tell you is how I think about the categories, and let you map that onto your own situation.

The honest summary

If I had to rank the categories by how much control I actually have over them: compute is the biggest cost and the one where local versus cloud decisions matter most. API spend is smaller than the hype suggests and shrinks further with basic model routing. Agent and MCP tooling is the fastest growing line and the one most people underestimate, mostly because the orchestration and logging layer isn’t optional once agents touch real systems. And the boring subscriptions, the backups, the monitoring, the password managers, are the ones I’d cut last, because they’re insurance, not overhead.

None of this replaces the work of actually running the thing. Paying for good infrastructure doesn’t automate away the parts of the job that still need a person watching the output, checking the logs, and deciding when something’s actually ready to ship. The tools make the work faster and more repeatable. They don’t make the judgment calls for you.

If you want to see how these pieces actually connect, from local rendering to agent workflows to the tools I use day to day, take a look around the rest of the site here.

Get new guides and videos first — join the Telegram channel.