Mac Studio vs €200/Month for AI: What's Actually Worth It?

Every few weeks someone asks me the same question in DMs, and it’s always some version of: should I buy the machine or keep paying the subscription? It’s a fair question. It’s also the wrong shape, and I keep answering it badly because I answer it from my situation instead of theirs.

So let me do the boring, useful thing and put my actual numbers on the table first. Not as a flex. The opposite. I have an R&D budget, I’m lucky enough to be allowed to spend it, and a decent chunk of this spend is me finding out what doesn’t work so I can tell you before you pay for the same lesson. Most of the people reading this don’t have that budget, and the honest version of this essay has to say that out loud.

Sorted Pixels is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

What I actually pay

A few lines, and they behave completely differently.

Claude Max, €200 a month. That’s the product. Chat, Claude Code, the whole experience layer. It’s flat, it’s predictable, I don’t think about it. If it disappeared tomorrow I’d feel it in my hands within an hour, because it’s the thing I touch.

API keys, mostly Claude, between €1,500 and €2,000 a month. That’s not me typing. That’s Hermes Agent, running on a Mac Studio I leave in my wife’s closet, headless, always on, next to the router, silent enough that nobody has complained yet. Multiple enterprise clients, concurrent projects, agents doing passes over real repos and real decks and real research, most of it while I’m doing something else. Roughly a quarter of that line is my own personal use of the loop. The other three quarters is project work that runs daily and pays for itself. Nobody’s hobby ChatGPT bill looks like that, and the number isn’t a badge. It’s a cost of goods sold.

A fixed SaaS bundle at about $312 a month. Higgsfield at $23 for media generation, ElevenLabs at $99 for voice, Vercel at $20, Neon at $20, GitHub Enterprise at $150. Boring infra and media utilities. Real money and still noise next to the API line. There’s a related Azure note I want to be careful about, because people mash it into “free GitHub”: I get $150 of Azure credits for R&D, they offset Azure spend on that side of the ledger, and they do not erase the GitHub line. Credits are credits. The $150 still hits the card.

Wall power, about €25 a month. The Studio pulls 15 to 25W average for agent work. The gaming PC with the NVIDIA 5090 in the office pulls 600 to 850W when it’s actually running Stable Diffusion, two to four hours a day on active months, and I don’t leave it idling. At an Italy all-in orientation of roughly €0.32/kWh that’s about €5 for the Studio and €20 for the 5090. Approximate R&D maths, not a utility invoice. Honest ownership cost. Still not the thing that decides anything.

The metal itself, €0 a month (kind of…). The Studio was €11,099. The 5090 was already in the office for other reasons. I amortise them the way you’d amortise any workstation, and I’ll do that maths in a minute, because it’s the maths that actually answers the title.

Then there’s a line that isn’t on the card at all, and it’s the one this whole essay turns on.

The line that never reaches the invoice

For about eight hours a day, the Studio isn’t waiting for Claude to answer. It’s running Qwen and Gemma locally. Open weights, not frontier models, and I want to be precise about that because the difference is the whole argument. Those models do the narrow steps in the pipeline: classification, extraction, first-pass summaries, the tedious middle of an agent loop where a frontier model would be wasted. Hermes routes to them by default and escalates to Claude only where the work needs it.

That work would otherwise be metered. So I did the boring thing and priced it.

Eight hours a day at a realistic sixty percent duty cycle, at about 40 tokens a second averaged across a fast MoE and a slower dense model, comes to roughly 21 million output tokens a month. Agent passes are input-heavy, so call it 166 million tokens of input on top. Price that at mid-tier API rates and it’s about €750 a month. Price it at frontier rates and it’s about €3,700. Price it the way the work would actually route, a blend of the two, and it lands at about €1,500 a month.

That’s the number that makes the closet make sense. Not “local AI is cheaper.” It isn’t, per token, for anything that needs a frontier model. What the closet does is take a tier of work off the meter entirely. My API line is €1,750 a month because the local tier absorbs another €1,500 that never becomes an invoice. Without the Studio, the same work costs about €3,250 in keys.

Which turns the Studio from a lifestyle purchase into an investment with a payback period.

€11,099 divided by €1,500 a month is 7.4 months. At the conservative end, if everything it displaces would only ever have gone to a mid-tier model, it’s fifteen months. At the aggressive end, three. I bought it in [month], so by the time you read this it has paid for itself under any of those assumptions, and everything it absorbs from here is margin.

I want to sit on this for a second because it’s the bit I got wrong in my own head for a year. I used to say the Studio buys availability and custody, not intelligence. That’s still true. But it also buys a tier. Availability, custody, and a tier of work at zero marginal cost. The third one is the one you can put in a spreadsheet, and it only exists if you have the volume to fill it. Hold that thought; it comes back at the end.

Two ways to read the same desk

Scenario A, what I actually pay now. €200 for Max, €1,750 mid for API keys, about €290 for the SaaS bundle in euros, €25 of power. €2,265 a month, €27,180 a year. Plus the €1,500 of absorbed work that isn’t on the card.

Scenario B, after I move the personal share of the loop to Grok Bot. I’m migrating the part of Hermes that serves me, not clients, onto a flat €100 a month Grok Bot seat. That’s roughly a quarter of the API line, about €440, replaced by €100. The project work that runs daily for clients stays on metered keys because it needs to. €1,925 a month, €23,100 a year.

That’s a fifteen percent saving for an afternoon of migration work. Worth doing. Not the lever. I’m flagging it this way on purpose, because the temptation with any migration story is to make it the hero, and the honest arithmetic says the hero was already in the closet.

Show Image

The other thing this migration taught me, which applies even if you never see a four-figure API bill: migrating a model is not migrating a harness. The prompts move in an afternoon. What doesn’t move is everything the harness has been implicitly compensating for, the failure modes you’ve tuned around, the places where your agents got reliable because you learned exactly how this model goes wrong. Swap the model, inherit a new set of wrong. So the migration is running as a parallel track, same tasks, same evidence gates, compared on output I’d actually hand to a client. If the cheaper loop produces work that needs a human rescue pass, it isn’t cheaper. My time is the most expensive token in the stack.

Who this bill is actually for

Most people reading this should not pattern-match to my desk. Here’s the field, in illustrative mid-2026 monthly cash bands, so you can find yourself before you copy anyone.

Everyday AI enthusiast. Laptop only. One product seat, maybe an image tool. €20 to €40 a month.

Freelance or indie builder. A heavy product seat if you live in the chat or the IDE all day, Cursor, a little API, maybe a tiny VPS so something stays awake. €100 to €400.

Small studio, two to five people. Shared Max or team seats, Cursor team, a modest shared API pool, hobby-to-pro hosting. €500 to €1,500.

Studio or agency, fifteen to thirty. Seats at scale, an IDE per head, a real shared API pool, media and voice SaaS. €3,000 to €8,000.

This desk. €2,265 in cash and another €1,500 absorbed. One person doing enterprise AI consultancy costs more than a five-person studio, before and after the Grok Bot migration. What would move this desk down a band isn’t a cheaper seat, it’s routing discipline, and I’m still working on that.

There’s a test hiding in that last band, and it’s the one worth borrowing whatever size you are.

Spread across the concurrent enterprise engagements it serves, my €1,750 token line is about €290 per engagement per month. Attached to billable work, a four-figure token line is a cost of goods sold. Attached to nothing, it isn’t a receipt. It’s a leak. If you can’t divide your API bill by something that pays, the bill is too big, whatever the number is.

The things you’re buying, and why they don’t substitute

Here’s the separation that took me an embarrassingly long time to make.

Product is what you pay for when you want the interface, the polish, the thing that keeps working when the underlying model changes. Claude Max is product. I’m not buying inference, I’m buying a desk I can sit at. No local box replaces it, because no local box ships me an updated Claude Code every fortnight.

Tokens are what you pay for when a machine is doing work on your behalf, at volume, without you watching. That API invoice is tokens. Tokens are pure variable cost, which is both the problem and the opportunity. You can’t engineer a subscription down. You absolutely can engineer a token loop down, and that’s most of what I’m doing this quarter.

A local tier is what owned metal gives you, and I’d never have called it that until I priced it. It’s the slice of work that a good open-weight model does to a standard I’d put my name on, at zero marginal cost. That slice is real. It’s also smaller than the internet suggests, and it only pays if you have enough of it.

Media and small SaaS is Higgsfield, ElevenLabs, Vercel, Neon, GitHub. Fixed lines, not heroic, about $312 together.

Unmetered iteration is the 5090. What it gives me isn’t “free AI.” It’s the absence of a meter in my peripheral vision while I’m exploring. There’s a specific kind of bad decision you make when every experiment has a visible price: you stop trying the dumb idea. The dumb idea is frequently where the taste comes from.

Notice that none of these substitute cleanly for each other. Buying a Mac Studio does not reduce my Claude Max bill. It does not reduce Higgsfield. It reduces the token line only for the subset of work a local model can actually do, and that subset is a tier, not a discount.

Why there’s a Mac Studio in a closet

The Mac Studio isn’t in my life because I benchmarked something. It’s there because Hermes Agent needs to be on.

An agent harness that only runs when my laptop is open and awake isn’t a harness, it’s a demo. The moment you have long-running jobs, scheduled passes, an agent that should wake up, check something, produce an artefact and go quiet again, you need metal that doesn’t sleep, doesn’t close, doesn’t go to a client meeting in a bag. My laptop is a terrible server because it’s a great laptop.

So: a Mac Studio, headless, on a shelf in my wife’s closet, because that’s where the router is and because it’s silent enough that nobody in my house has complained yet. I ssh into it. I occasionally screen-share into it when something GUI-shaped needs eyes. It is the least glamorous piece of infrastructure I own and it’s the one I’d replace first if it died.

What it is not is a frontier model host. When Hermes needs frontier reasoning, the call still goes out to Claude on a metered key. What the box does with its other eight hours is run open weights for the work that doesn’t need frontier reasoning, and that’s the €1,500 above. Availability, custody, a tier. Credentials that live under my control instead of in somebody’s cloud tenancy. Client files that don’t leave the building. It’s boring. Boring is the point.

The 5090, meanwhile, is a completely different animal. That’s where generative image and video experiments happen, where I test whether a small model is good enough for a narrow step before I let Hermes route to it, where I burn an afternoon on Stable Diffusion for a moodboard nobody’s paying for yet. It’s not always-on and it doesn’t need to be.

Two machines, two jobs. Always-on harness with a local tier. Unmetered sandbox. If I’d tried to make one box do both I’d have bought the wrong box, and I nearly did.

The thing I actually want local metal for

If you asked me what would justify a serious local machine tomorrow, even without the token maths, it’d be computer use I can watch.

A lot of what I do, and a lot of what the designers and PMs reading this do, happens inside GUI tools that have no meaningful API surface. Blender. CAD. Video editing timelines. The kind of software where the state of the work lives in a viewport, not in a file you can diff.

Agents that drive those tools are genuinely useful and genuinely alarming, and the alarming part is the reason local matters. When an agent is clicking around inside a modelling app, I want to be looking at the screen. Not reading a log afterwards. Looking at it, in the moment, ready to take the mouse back. That’s the same instinct that makes you sit next to a junior the first week. Variance is the UX.

Running that through a remote sandbox is technically possible and practically miserable. Latency between intent and reaction is the difference between supervising and receiving a report. On a machine in my house, with a display I can throw up on a second monitor, supervision is cheap. Remote, supervision is a video call with a robot.

Second reason, less romantic: client files. Some of these assets don’t leave my custody, full stop. A CAD file for an enterprise client isn’t going through a third-party computer-use sandbox because I felt like trying something on a Sunday.

The data doesn’t leave the building

That last point deserves its own section, because it’s the part of the closet that has nothing to do with token maths and everything to do with why I could sign the contracts I sign.

Enterprise clients come with paper. Data processing agreements, residency clauses, lists of approved sub-processors, the occasional line that says in plain language that a category of material may not be uploaded to any third-party AI service. Not “should be careful with.” May not. For that work, a frontier API isn’t expensive. It’s unavailable. The only tier that can touch it is the one running on metal I own, on a network I control, with weights that don’t phone home.

So the local tier is doing two jobs at once, and I priced only one of them above. The €1,500 a month is the work it absorbs that could have gone to an API. There’s a second slice, smaller but non-negotiable, that could never have gone anywhere else. I can’t put a euro figure on it honestly, because the alternative isn’t a more expensive invoice, it’s not doing the engagement. That’s a different kind of number.

Which is where the NAS comes in, and why there are two boxes on that closet shelf, not one.

The Studio runs the models. A UGREEN DXP6800 Pro next to it holds everything the models touch: six bays and two NVMe slots, up to 196TB, two 10GbE ports straight into the Studio, snapshots on a schedule, and an encrypted off-site copy of the parts that matter. It’s the system of record for client projects, for the media archive that ten years of studio work produces, and for the datasets I train on. It pulls about 30W idle and 70W when the drives are busy, call it €10 to €15 a month at the wall, and I’ve left it out of the payback maths above because it makes the Studio look better and I’d rather the maths were conservative.

Three things live there that could not live anywhere else.

Client material under a residency clause. The obvious one. Files, transcripts, decks, repos. When Hermes does a pass over a client’s research corpus, the corpus is on the NAS, the model is on the Studio, and nothing in that loop has an internet connection it needs. That’s not a privacy feature I bolted on. It’s the reason the engagement exists.

LoRA training sets. When I fine-tune a small image model on a brand’s own visual language, the training set is that brand’s product photography, their archive, their unreleased work. It goes from the NAS to the 5090 over 10GbE, the adapter comes back, and the source images never leave the building. There is no version of that workflow that goes through a hosted fine-tuning API, at any price, because the price isn’t the constraint. The paper is.

The archive itself. Ten years of projects and media is a lot of terabytes, and at that scale cloud object storage is a meter too, priced per terabyte per month with egress on top. Owning the disks is the same argument as owning the inference tier: a fixed cost of production for the one line that would otherwise reprice on somebody else’s schedule.

I want to be careful not to oversell this. Most people reading this don’t have residency clauses. If your client work is a Notion doc and a Figma file, you don’t need 196TB in a closet and you’d be mad to buy it. The custody argument only exists at the point where a contract says it does. But at that point it stops being an argument and becomes the whole design of the desk, and every “just rent it” comparison quietly assumes you’re allowed to.

Renting the harness instead of owning it

Which raises the obvious counter: if the Studio is an always-on process with a local tier and custody, why is it in a closet and not in a datacentre?

Fair. Hermes, or OpenClaw, same logic, can run on a rented box. The architecture doesn’t care that the metal is mine. What it does care about is whether the rented box can run weights, because that’s where the €1,500 lives.

The useful split, once you look at real catalogues, is two different purchases people mash into one.

A CPU desk. Orchestration, schedules, files, API calls out to Claude. High-RAM CPU and boring uptime. On Hetzner Cloud in the EU you’re roughly in the €15 to €50 a month band for a practical 16 to 32GB always-on desk. Hostinger tops out around 32GB on the CPU side, with promo pricing friendlier than renew. Fine for keeping Hermes awake. It cannot run the local tier. Rent this instead of the closet and the absorbed €1,500 goes straight back on the meter.

A mid GPU. Hetzner doesn’t sell GPUs on the cloud SKUs; the dedicated path is something like a GEX45 with a 24GB RTX PRO 4000 at about €214 a month plus setup. Hostinger’s GPU hosting is hourly, you destroy the instance to stop the meter. Scaleway’s L4 is about €0.79 an hour, which is roughly €575 if you leave it on all month, which is the point: don’t. A careful week of hourly mid-GPU is tens of euros. A careless “I’ll leave it on” becomes the API bill with a different logo.

Prices are mid-September 2026, excluding VAT, orientation not quote.

Put the three side by side at my workload, with the Studio amortised over four years so the comparison isn’t flattering to me, and you get this. Owning: €2,496 a month all in. Renting a mid GPU that keeps a local tier: €2,454. Renting a CPU desk and losing the tier: €3,772.

Rent versus own is a €42 a month question. Whether a local tier exists at all is a €1,300 a month question. Most “just rent a VPS” advice is quietly recommending the expensive column, because it’s thinking about uptime and forgetting about weights.

A few caveats I want stuck to the page, because they protect the thesis. Spot instances and ephemeral pods are a bad home for always-on agents; your harness wants a boring lease, not a machine that can vanish mid-cron. Rented boxes don’t give you a monitor, so no eyes-on computer use without bringing your own desktop stack. A 24GB card is not 512GB of unified memory; the rented tier is a smaller tier. And if your clients care where the bits live, read the residency page before the pricing page.

What this means for my desk: I already own the Studio, it’s paid for itself, and the €42 buys custody, a monitor cable, and a bigger local tier than any rented card. If I didn’t own it, a €214 GPU box would be a reasonable move. A €32 CPU box would not, whatever the uptime.

Seven levers, in the order I’d pull them

This is the part I wish someone had handed me a year ago. It’s ranked by how much money it moves, which is not the order people usually try them in.

1. Narrow the agents. The reason my token line got that big wasn’t model pricing. It was agents doing more work than the job required. Re-reading context they already had, taking a wide pass where a narrow one would do, running a full research sweep because nobody told them what “enough” looks like. Most token bills are a harness design problem wearing a pricing costume. Add gates so work stops when it’s done instead of when it runs out of ideas. Cut the catch-all agent that’s been quietly reading your whole repo every run. This is the biggest lever and it costs nothing.

2. Route by tier. Frontier models for the steps that need frontier reasoning, cheap models for everything else. The cheap models are good now. GLM, Qwen, Gemma and their peers do classification, extraction and first-pass summaries at a standard that’s fine for the middle of a pipeline, at a fraction of frontier cost. This is what the Studio does for me at zero marginal cost. You don’t need the Studio to do it. You need a router.

3. Put personal loops on a flat seat. The Hermes work that serves me, not clients, is exactly the shape a flat subscription is built for: bursty, unattended, not billable. Grok Bot at €100 replaces about €440 of metered personal use on my desk. If you’re running a self-hosted harness for yourself and not for clients, a flat seat probably replaces most of your API line, not a quarter of it.

4. One key, many models. OpenRouter or an equivalent gives you every model behind one key, per-request routing, spend caps and fallbacks. It’s how you make lever two operational without rewriting your harness every time a cheaper model ships. It’s also how you find out, with receipts, which model your pipeline actually needs.

5. Cache and batch the unattended volume. Prompt caching for the context your agents re-read every pass. Batch APIs for anything that can wait an hour. Neither is glamorous. Both are meaningful on a four-figure line and invisible on a €20 one.

6. Right-size the SaaS. Once a quarter, list the fixed lines and ask which ones are project-shaped subscriptions that outlived the project. On my desk this is about $312 and I’ve mostly done it. On a small studio’s desk it’s often the fastest €100 a month you’ll find.

7. Buy metal last, and only for a tier you can fill. This is the one everybody reaches for first. It’s the one that pays back slowest for almost everyone.

Each hour a day of sustained local inference is worth about €188 a month of avoided tokens at blended rates. To pay back a €11,099 box inside two years you need about two and a half hours a day, every day, of a local model actually running weights, not idling, not “available.” My desk does eight. A designer poking at Stable Diffusion for thirty minutes a day is looking at a ten-year payback. That’s the number to test against your own week before you spend five figures on silicon. If you can’t measure it, you don’t have it.

So: buy the machine or pay the subscription?

Let me answer it once, properly, for the five desks I actually see.

Everyday AI enthusiast, €20 to €40 a month. No. Not on a two-year horizon, not on a five-year one, not if you value your setup time at more than zero. A Mac Studio is not a cheaper ChatGPT. Buy hardware when you want something a subscription structurally cannot give you: unmetered iteration, custody of files, an agent you can supervise on your own screen. If none of those describes a real problem you have this month, keep your money.

Freelance or indie builder, €100 to €400. No, not this machine. Pull levers one to four first. Route cheap models through a router, put your personal loop on a flat seat, keep the GPU you already own for the unmetered sandbox. If a token bill starts to sting, the fix is almost never hardware first. Hardware is a lever for a problem you’ve already characterised. It’s a terrible instrument for discovering one.

Small studio, two to five, €500 to €1,500. Not yet. A shared API pool with routing discipline and spend caps will do more than any box. The moment you have a genuinely always-on, unattended job that runs weights more than two and a half hours a day, and files with custody requirements, the maths flips and you’re in the next answer.

Studio or agency, fifteen to thirty, €3,000 to €8,000. Rent. A closet doesn’t scale to a team, and a team’s inference is an ops decision, not a desk decision. Rent the GPU tier, own the custody policy, meter everything, and make somebody accountable for the token line the way somebody is accountable for cloud spend.

AI practitioner or consultant running client work at volume. Yes, if. If you have sustained local inference above the threshold, if you have client files that cannot leave the building, if you have agents driving GUIs that you need to watch. Then the machine pays back inside a year and everything after is margin. You’ll still end up with a split, because the five cost categories don’t collapse: product subscription for the desk you sit at, metered tokens for the frontier work, a local tier for the rest, rented burst for media, and owned metal for the two things rented metal is bad at, being always-on under your own custody and letting you watch an agent drive a screen in real time.

That’s the frame. The machine in the closet isn’t there because I ran the maths and hardware won. It’s there because it does three jobs a subscription doesn’t do, and one of those jobs happens to have a payback period. The subscriptions are there because they do jobs the hardware doesn’t do. Anyone selling you a single clean crossover point is selling you a spreadsheet, not a desk.

The line you don’t control

There’s one more thing the spreadsheet can’t show you, and it’s the reason I’d buy the machine again even if every number above were half as good.

Everything in this essay is priced at September 2026 rates. Token prices have fallen for three years straight and it has become very easy to build a desk, or a studio, or a company, on the assumption that they keep falling. Nothing guarantees that. The frontier labs are burning capital at a rate that has to turn into margin at some point, and margin comes from exactly one place: the meter. If your biggest line is metered keys, you have a cost structure somebody else controls. A thirty percent price move at one vendor is a thirty percent move on your largest expense, overnight, announced in a blog post, with no notice period and no negotiation. If you’ve optimised your whole loop for one model, you can’t even walk away without inheriting a new set of wrong.

Look at my receipt again with that in mind. Claude Max, someone else’s price. API keys, someone else’s price. The SaaS bundle, five other people’s prices. The local tier is the only line where the cost of production is mine. Electricity, a machine I already own, weights nobody can meter. It’s also the only line that structurally cannot go up. That’s not a discount. That’s a hedge, and hedges look expensive right up until the day they don’t.

The same logic is why the Grok Bot move is worth making even at fifteen percent. A flat seat is a hedge against metered pricing for the part of the loop it fits. A router with fallbacks is a hedge against any single vendor. A local tier is a hedge against the meter itself. The point isn’t to pick one. The point is to not have a desk where every line reprices on somebody else’s schedule.

And this, finally, is where the five answers above stop being about money. At €20 a month there’s nothing to hedge; the worst case is a slightly more expensive subscription and you’ll shrug. At €500 you should have a router. At €2,000 with client work riding on the meter, the question isn’t whether owned metal pays back in seven months or fifteen. It’s whether you want the biggest line of your business to be a variable you don’t own. I didn’t. That’s the whole reason there’s a Mac Studio in a closet.

One last thing, and I mean this as encouragement rather than a caveat. Everything above is me spending an R&D budget I’m fortunate to have, on a stack that serves multiple enterprise clients at once. If your setup is a laptop, a subscription, and a stubborn side project, you are not behind. You’re solving a different problem, and quite possibly a healthier one. The gear didn’t make the work good. It made the boring parts survivable so there was time left over for the part where taste happens. That part still hasn’t found a way to run on anybody’s silicon but mine.

Thanks for reading Sorted Pixels! This post is public so feel free to share it.