How to Actually Decide What Mac to Buy for AI Work

Last spring I argued that a $2,500 Mac Mini would not save you from the cloud. The most common response was predictable: okay, so what should I buy?

That question sounds like it should end with a model number. It does not.

I had just gone through the decision myself, and the way my old machine failed taught me more than any benchmark comparison did.

My MacBook Air M3 had 16GB of memory and a 256GB SSD. Memory was the first source of friction. Docker took a meaningful share. Chrome took more. I found myself choosing between Topaz AI and Photoshop because running both at once was a bad idea.

Annoying, but manageable. I closed apps and kept working.

Then local project work became heavier. Repositories, node modules, containers, mounted project folders, generated assets, and caches turned the SSD into the real wall. Within weeks I was deleting software to make room for the files the software needed.

Memory was “I want to upgrade.” Storage was “I need to upgrade now.”

I landed on a 16-inch MacBook Pro with an M5 Pro, 64GB of unified memory, and a 2TB SSD. At last check, I had already used 1.45TB with Optimize Storage turned off, and I still had not reinstalled every creative app I removed.

The headroom is real. It is not infinite.

This is Mac-shaped because that is what I use, but the framework applies to any hardware decision where AI tools are changing the workflow faster than the normal upgrade cycle.

Stop solving for the wrong bottleneck

Most hardware threads follow the same script. Someone asks what to buy. Fifty people answer with their setup. Almost none of them know how the person asking actually works.

The debate usually centers on chip tier and local inference. Both matter. Neither was the first constraint that broke for me.

In a cloud-first AI workflow, the laptop is often the coordination layer. It manages files, terminals, containers, browsers, project assets, and several tools running at once. The expensive reasoning happens elsewhere.

That changes the buying logic.

I assumed the M3 chip would be the constraint. Memory and storage failed months before compute became the issue. A more powerful chip inside the same cramped workflow would have been the wrong upgrade.

The honest answer to “what should I buy?” is not a SKU. It is a system for deciding where your workflow is already asking for more room.

Score the work, not the spec sheet

Before you shop, score yourself from one to five across six dimensions. Use your actual Tuesday afternoon state, not the clean desktop you see after a reboot.

Portability

Do you work from multiple locations?

I scored a five. I work from the couch, an office, coffee shops, softball fields, and wherever travel puts me. A desktop was never a serious option.

If you are a four or five, stop pretending the cheapest path to maximum compute is relevant. You are buying a laptop.

Heavy multitasking

How many meaningful applications are open at once?

My normal state includes a browser with too many tabs, an editor, Docker, terminals, communication tools, and multiple AI surfaces. The old 16GB disappeared quickly.

This is the most commonly underspecified dimension because every app works fine alone. The friction appears in the combination.

Do not size memory for the one task you are doing. Size it for the overlap between tasks you refuse to close.

Containers and local services

Do you run Docker, local databases, emulators, or several development servers?

Containers accumulate quietly. Images, volumes, and caches grow while each individual project still looks small. A high score here pushes both memory and storage upward.

The important signal is not whether you can clear old images. It is how often maintenance interrupts the work you meant to do.

Creative and media work

Video editing, local image generation, Photoshop, and AI upscaling change the equation. They create sustained loads, large assets, and more pressure on graphics resources.

I am a two or three here. The work arrives in bursts rather than every day. That made additional graphics capacity useful, but not important enough to justify buying the highest chip tier.

If media work is the product, your answer will be different.

Local inference

Are you running local models because they solve a real recurring problem, or because the idea is appealing?

I scored a three. I experiment locally and value the privacy and control. But the highest-value reasoning in my workflow still happens in cloud models, and model memory requirements grow fast as capability rises.

Local inference is a valid driver when it is central to the job. It is an expensive hobby when it only appears in your imagined future workflow.

Storage-heavy work

How many repositories, media libraries, containers, project folders, and generated artifacts need to remain local?

This was my five. It was the constraint that forced the purchase.

The wrong question is whether today’s files fit. The right question is how often your tools create another copy, cache, build directory, export, or local dependency tree.

AI-assisted building increases the number of things one person can keep moving. That also increases the amount of local state one person is responsible for.

Budget is a constraint, not a dimension

Budget does not describe how you work. It describes which compromises you have to make.

Once the six scores are visible, the trade becomes clearer: which high-scoring dimensions get the money, and which lower-scoring dimensions can stay constrained?

My scores looked like this:

Dimension	Score	Buying implication
Portability	5	Laptop only
Multitasking	4	Memory needed real headroom
Containers	4–5	Memory and SSD both mattered
Creative work	2–3	Better graphics were useful, not decisive
Local inference	3	Worth supporting, not worth optimizing around
Storage	5	2TB became the floor for my workflow

The scores did not hand me a product page. They told me where to spend.

Memory and storage were the priorities. Creative work and local inference were not strong enough to pull me into the Max tier. The M5 Pro configuration covered the bottlenecks I had evidence for without charging me for an aspirational workflow.

The upgrade priority stack

If you can only improve one part of a base configuration, use this order as a starting point:

Memory first. It affects every simultaneous workload, and you cannot add more later.

Storage second. Pressure compounds over time. Month-one comfort says little about month-twelve reality.

Chip tier third. Move it higher only when sustained creative work, local inference, or another compute-heavy workload is central to what you do.

That order is not universal. It is deliberately biased toward people using cloud AI to build products, run containers, and coordinate several forms of work at once.

Buy for the first failure

Two years earlier, 16GB and a modest SSD felt adequate. The tools changed my habits faster than the hardware cycle expected.

That is the part benchmark advice misses. You are not buying for an abstract category called “AI work.” You are buying for the first resource your own workflow consumes faster than you can manage it.

Track the friction for a week. What do you close? What do you delete? What makes you wait? What forces you to change the work instead of finishing it?

Buy for that failure first.