Why I Bought the M5 Ultra Mac Studio for Local AI Agents

I bought a Mac Studio, and the funny part is I can admit right from the beginning that I absolutely did not need one. My MacBook Pro already handles pretty much everything I normally throw at it. I can write my books, edit audio and video, work on the site, record the podcasts, use all of the AI services I already pay for, and do the normal day-to-day stuff without really pushing it hard enough to justify replacing it. So if I was looking at the Mac Studio as a normal computer upgrade, it would make almost no sense. The reason I bought one has a lot more to do with how my workflow has started changing and how much more I have been using AI agents instead of just using AI as something I open, ask a question, and close again.

I have been experimenting more with agents over the past few weeks, and that has slowly turned into me giving different AI systems actual jobs instead of just using one assistant for everything. I have an agent that helps with editing my books, another that focuses more on publishing, one that can look at marketing, and another that I use when I am working through new business ideas. The point is not that these agents magically run everything for me because they do not. I still have to give them instructions, check what they are doing, and make the final decisions, but they can take a larger task and keep working through it instead of me sitting there manually doing every part myself. Once I started using them that way, I realized pretty quickly that this was probably going to become a much bigger part of how I work.

The thing that really pushed me toward buying hardware for it was actually my editing workflow. I had been using Genspark for one of my editing agents, and the process was working surprisingly well. I would finish a book, run it through ChatGPT first to look for issues and help me create a prompt for the editing pass, then give that prompt and the manuscript to the agent. It would go through the book, make the changes, give me the edited file and a change log, and then I could run the new version through another check afterward. On the $25 plan I was using, one of those full passes took around 3,000 credits, and because the plan gave me around 10,000 credits, that seemed perfectly reasonable. In my head, that meant I could edit around three books a month using the agent without having to jump into a much more expensive plan.

The problem was that I was not actually seeing normal credit usage. When I looked closer, Genspark was running a promotional credit rate where I was only being charged around 10% of what the task would normally use. That means a job that looked like it took around 3,000 credits could actually require closer to 30,000 once that promotion disappeared. At that point, the math changed completely for me because I publish too often for that to be something I can just ignore. A book is not guaranteed to immediately make back whatever I spend on editing it, and if I am putting out multiple books every month, I cannot build a workflow around something where one editing pass could eat through an entire paid plan by itself.

I could obviously solve that by paying for more cloud AI. There are bigger Genspark plans, higher Claude tiers, and plenty of other services that would let me keep everything online. That is actually the easiest option because I would not have to worry about hardware, local models, setup, or maintaining anything myself. But that is also when I started looking at how much all of these subscriptions can add up to if I am using agents for editing, publishing, marketing, analytics, and other parts of what I do. If I am going to keep increasing what I spend every month just to rent access to more compute, eventually I have to ask whether it makes more sense to put some of that money toward hardware I actually own.

That is where the Mac Studio started becoming a real option instead of just something I wanted because I like computers. I have wanted a Mac Studio for a while, so I am not going to pretend there was not already some temptation there, but the local AI angle finally gave me an actual use case for it. Apple also has its newer upgrade program, which made the cost a little easier for me to justify because instead of dropping the entire amount at once, I can spread it out monthly and still pay it off early if I want to. Once I decided I was probably going to do it, the next problem became figuring out which Mac Studio actually made sense.

I originally started looking at the M5 Max version with 128GB of unified memory because memory is one of the first things you look at when you start talking about running larger AI models locally. The problem is that once I upgraded the processor and memory to where I wanted them, I was already getting close to the price of the base M5 Ultra. The Max configuration I was looking at had an 18-core CPU and 40-core GPU with 128GB of unified memory, while spending around $500 more moved me into the Ultra with a 30-core CPU, 64-core GPU, and 96GB of unified memory. That meant giving up around 32GB of memory, which definitely matters for local AI, but I would gain 12 CPU cores, 24 GPU cores, and roughly double the memory bandwidth.

The memory bandwidth was probably the thing that pushed me over the edge because the Max configuration is around 600GB per second while the Ultra gets up around 1.2TB per second. For what I want to do, I decided I would rather have the additional processing power and bandwidth even if it meant stepping down from 128GB to 96GB of unified memory. I am not trying to run the biggest model that exists just so I can say I did it. I want something that can realistically run the kinds of models I need for agents, editing, research, automation, and all the other things I already use AI for without making the entire machine feel like it is sitting there waiting on one task.

The main software I want to experiment with on it is Hermes. What caught my attention with Hermes is not just that it can run local models, but the whole idea of giving an agent memory, allowing it to build skills, and eventually letting it create more specialized agents for different jobs. That is much closer to what I actually want than simply having another chatbot running on my computer. If I am going to dedicate a machine this powerful to AI, I want it doing things continuously and becoming part of how I already work instead of just being a local replacement for opening ChatGPT in a browser.

A good example is everything that happens after I finish recording one of these podcasts. Right now I record the episode, get the transcript, create the subtitle, write the Apple Podcasts description, create the show notes, turn the episode into a full site post, and sometimes make another version that works better on Medium. I use AI throughout that process already, but I am still the one moving everything from one step to the next. In theory, I could eventually have an agent that understands exactly how Tek With Josh show notes are supposed to look, another that knows how I write the site posts, and another that understands how I want something rewritten for Medium. Once the transcript exists, those agents could start working through those jobs instead of me rebuilding the same workflow every single time.

The same thing applies to the publishing side. I already use agents to go over numbers with me every Monday for the books, podcasts, and site. They can look at rankings, pull together new reviews, keep spreadsheets updated, compare performance from one week to another, and help me notice when something is starting to move up or down. None of that is something I could not do myself. I can go to Amazon and check a ranking. I can open analytics and look at site traffic. I can update a spreadsheet. The value is that I do not necessarily need to be the person spending time on every one of those little repetitive jobs when I would rather be writing another book, recording something, working on another project, or just doing something that actually requires me.

That is probably the biggest reason the Mac Studio makes sense to me because I am not buying it for one single AI task. If all I wanted to do was run one local chatbot, spending this kind of money would be ridiculous. I already have the publishing side, the podcasts, the website, different business ideas, analytics, social media, and all of these little things happening at once. I also already have a Mac mini running as a server, and when the Studio gets here, I plan on moving a lot of that over as well. That means the Studio can handle the AI agents while also taking over server duties instead of me keeping another Mac around doing only that.

I also want to get back into making more videos. I already have the Books By Josh YouTube channel where I can talk more about writing and publishing, I bought a green screen, and I have been slowly getting everything ready to start doing more with it again. Tek With Josh technically has a YouTube channel too, but right now it is mostly another place where the podcast lives. Having something this powerful sitting there makes me want to experiment more with actual video for both channels because I can edit everything on the same machine that is already handling the AI side of the workflow. At that point, the Studio becomes less of an expensive computer I bought just to run one model and more of the central machine behind a lot of what I do.

I am also completely aware that when this thing gets here, I am probably going to have no idea what I am doing for a while. I need to move my server setup over from the Mac mini, get Hermes installed, figure out which models actually make sense on 96GB of unified memory, decide what I want running all the time, and work out how much access I am comfortable giving different agents. There are going to be things I break, things that do not work the way I expect, and probably moments where I wonder why I did not just pay somebody else for cloud compute and avoid the headache. That is part of the reason I want to do it, though, because I like figuring this stuff out and seeing how far I can push something once I have it.

I also do not think everyone interested in AI should go out and buy a Mac Studio. If you use ChatGPT a few times a day, pay for one subscription, and everything you need already works in the cloud, there is almost no reason to spend thousands of dollars trying to replace that. Even for me, this is not automatically cheaper just because I own the hardware. I still paid for the machine, I still have electricity, and I am probably going to continue paying for some cloud AI anyway because there are models and services I like using. The difference is that I am reaching the point where enough of my normal workflow can potentially live locally that owning the hardware starts to make more sense than it would for someone who just wants to ask an AI a few questions.

That is really why I bought the Mac Studio. My MacBook Pro was not too slow, nothing broke, and I was not sitting around waiting for Final Cut to export something. The thing that changed was how I use AI and how much more I want these systems to do without me manually babysitting every step. I went from using AI mostly as something that answers me to experimenting with agents that can actually take responsibility for a specific part of my workflow, and once that started happening, I began looking at computing power differently too.

Maybe I will get the Studio, spend a few months with it, and decide that I completely overdid it. I would not even be shocked if that happens. But I think there is also a real chance that this ends up being one of those purchases where the interesting part is not how much faster it is than my MacBook Pro, but how much it changes what I actually expect a computer to do for me. Instead of just being the machine I sit in front of to get work done, I want to see what happens when it becomes the machine doing some of that work alongside me.

Thanks,

Josh