Hardware Is the Access

Recognizing the Pattern

For most of 2024 I told myself I was busy. It’s an easy story to live inside, because the calendar agrees with it: the days were full, the tickets were real, the production issues weren’t imaginary. Nothing about it looks wrong — which is exactly why it took me most of a year to notice anything at all.

Near the end of the year I stopped looking at the week and looked at the pattern. Same tasks. Same production issues. Same shape of problem, arriving on the same cadence. Each one met fresh and resolved again, with nothing about the system getting better for it. The last genuinely new thing I’d learned at that job was somewhere back in mid-2024. Familiarity had been passing for competence for about a year.

Boredom would have been easier to explain away with a weekend off. What I had was quieter than boredom: nothing broken enough to complain about, the days still filling, the problems still arriving, and the year ending with me exactly as capable as I’d started it. A heavy workload looks identical from the inside.

Yep — I was stagnating.

The Answer Was Money

Mine. Spent on something that would still be there on the days nobody was paying me to use it.

I asked work to cover it, more than once. They didn’t. So I paid for it myself.

The other reason it had to be mine: I’d been using AI tooling through a subscription my employer pays for. It’s a good subscription — it just belongs to them. The quota, the model choices, the limits, plus a set of security and privacy questions that were never mine to answer. That’s what a company tool is. It was never going to be the place where I tried the things I actually wanted to try.

So I spent about three months looking — not for a laptop, exactly, but for three properties at once: enough memory to run a usable model locally, enough efficiency that I wouldn’t have to think about it, and a price I could pay myself. Memory was the one I wouldn’t trade, and more of it was simply better, which took most of the market off the table before I’d looked closely at anything: the 14-inch couldn’t get there where I live, the previous generation topped out below it, and the base chip was never in the conversation.

Almost nothing had all three.

The gaming laptops were the easiest to rule out. Video memory comes in fixed steps — 8GB, 12, 16, 24 at the top — and the models I wanted to run needed more than the bottom of that ladder. The 24GB machines started at $2,899, the price of the configuration I eventually bought, somewhere else, on a different price list. What that money bought here was 8GB and a wall socket. And the whole generation had barely reached shops — the last week of March, the same week I was deciding.

The AMD side was harder to dismiss, because that’s where the memory might have been. The chip was real, the 128GB configurations were announced with prices attached, and the machines were niche 13-inch detachables rather than anything I wanted to work on. What was actually orderable at the end of Q1 was the 32GB version, shipping in spring; the 128GB one wasn’t available in any market I could reach, and the range everyone remembers now filled in later in the year, long after I’d bought. I also already owned an old desktop, and rebuilding it into a machine that could do this meant new parts, more power pulled from the wall, and a platform frozen where it was for another few years.

So the choice made itself on a boring scorecard: it was the only machine that checked nearly every box at once — enough memory to hold a model, no wall socket required, available without a queue. And the part I have to concede: the closest competitor was, in the market where it was actually on sale, cheaper than what I paid. The configurations that matter most arrive last here, if they arrive at all. That, more than any cleverness on my part, is the honest reason none of the alternatives came close.

The Spec I Could Afford

48GB is the ceiling for that chip — Apple’s own spec sheet lists 24GB as the base for the M4 Pro and 48GB as the max — so the memory question answered itself in one line: as much as this silicon can address, at 273GB/s, shared with the GPU instead of sitting behind a PCIe slot.

Storage went the other way. 512GB is where the 16-inch starts, and where I stayed — a strange place to land on a machine whose whole job is to hold models. I’ll come back to that.

The 16-inch M4 Pro started at $2,499 when it launched in November 2024; the 48GB configuration was the $2,899 tier. I’m not turning this into a receipt for what I paid. What matters is that it was a lot of money for one item by my own standards, and it came out of my own account.

The amount wasn’t the hard part. Deciding was. I’d been reading for three months and could have kept going for another three, and I wasn’t the only person who had to be moved. What I argued — to myself first, then out loud — was that I was buying access: a machine I could run things on without asking anyone, in a field moving faster than reading can keep up with. I needed something new, and I’d gone a year without any.

It took more than one attempt to land. I told my wife, and she wasn’t convinced — she handed me a book about being decisive instead, which is the most useful thing anyone’s given me in a while. I read it and tried again, and it still wasn’t settled.

What closed it was older, and had no bibliography at all: a sentence from my mother, years back, which I passed to her the way it had been passed to me. Don’t be cheap on yourself. She came around after that, reluctantly.

She wasn’t talking about computers. But I’d spent a year being cheap with myself, and the price was twelve months of work I wasn’t learning anything from.

What the Other Options Cost

If you’d told me in March 2025 that I was overpaying and a better answer was on the way, I would mostly have agreed with you. It just wasn’t buyable.

The chip that made this category interesting was AMD’s Ryzen AI Max+ 395 — up to 128GB of unified LPDDR5X. The machines built around it were announced in February 2025 with prices attached, and they were pre-orders: first shipments in early Q3 2025, with later batches slipping past that. AMD’s own developer platform in this class didn’t reach the US until the middle of 2026, at $3,999.

So the timeline isn’t that I bought a month before the alternative arrived. The alternative spent most of 2025 as a web page I could read, and by the time it was buyable I had a year of work on the machine I already had.

Which leaves the version of the question that actually matters — the one I asked myself at the time: why not wait a year and buy the better machine when it arrives?

For one thing, it didn’t arrive for me. The companies building those machines sell direct, into a list of countries mine isn’t on, and a warranty outside them is theoretical. For another, waiting isn’t free. What I was short of was time, and the diagnosis I’d already given myself was that a year had passed without me getting better at anything. Delaying the purchase would have delayed the growth by the same amount, and I didn’t have a second year to hand over. Hardware keeps improving on its own. The months don’t.

What the Memory Market Did Next

I’ll get this out of the way, because the honest version of it is luck rather than judgment. While I was using the laptop, the components inside it were turning into the scarcest parts of the industry, and memory went through the roof.

I’m not claiming foresight. I bought a fixed amount of capacity at a moment when capacity was briefly, accidentally reasonable, and everything around it has moved in one direction since. The memory was luck. The storage — the one component I left at the base — was the actual guess, and the market went after it: two months later I bought a 1TB Samsung NVMe to patch the hole. Here, 1TB meant an M4 Max, which was past what I could spend. Two purchases in two months, for one machine. It didn’t quietly turn into a mistake while I was busy with something else. That’s all I’m claiming.

The counterfactual has a price too. The generation I would have waited for arrived in March 2026 — better chips, a higher memory ceiling, twice the base storage — and it cost more than the one it replaced. Three months later Apple raised prices across the Mac line and said why: memory and storage costs. Waiting wouldn’t have made this cheaper. It would have bought a better machine for more money. I was making a different purchase. The thing I was short of had a price, and it wasn’t going down.

What the Money Bought

The honest accounting is that a purchase buys capability, and capability isn’t progress. What the machine actually got used for later — the part that matters — has its own article coming: tooling I wrote myself, which hands a model my production data and the business knowledge I’ve built up around it, and a small model running locally to reason over them. A frontier model in the cloud would reason better and answer faster. It would also never be allowed near my production data, so the local one is the one that runs. All of that cost me work, not money.

Paying for the machine was the cheap half of this. I put my own money on the line, which was the entry fee and the commitment. Nobody was going to keep it alive for me. The half that produced anything took months of unpaid attention.

But the thing I keep coming back to isn’t the machine. It’s that between mid-2024 and March 2025 I didn’t learn anything new at work, and the reason wasn’t that I’d run out of things to learn. Every experiment I could have run had a price attached to it — someone else’s quota, a bill, a security review, a reason to ask permission first — and when the cost of trying something is administrative, you mostly don’t try it.

That’s the fee I opened with: paid in motivation, invisible, never weighed against the price of removing the friction. I paid it for about a year without seeing the total. The money and the motivation were both costs, and the money turned out to be the cheaper one. It was paid once, and what it bought was the access.

Access to tools is what makes learning cheap. I’ve believed that for years — this is the first time I spent real money testing it on myself. What I bought was the removal of an excuse, and I’d rather pay for that than keep calling it being busy.