She wants to learn AI.
I told her not to buy the laptop yet.

The most useful conversations I have about AI are almost never with technology people. This one started with a question I could not answer honestly without first checking a price.

The student

She is not a computer science graduate. She runs a real business — supplying raw material into hawker centres, the kind of operation that runs on WhatsApp, a calculator, and a memory for who owes what. She approached me to learn AI, and she is willing to pay for the teaching.

She also has no laptop. So the first question in the course is not what can AI do — it is what do I have to buy before I can begin?

That question deserves a more honest answer than most people in my position give, because the person answering it is usually selling something. I intend to walk her through the entire risk before she commits a ringgit. What follows is roughly what I plan to say.

What the machine costs now

I was going to recommend a MacBook Air. It is the machine I would put in most people’s hands: light, quiet, long battery, and it runs everything a business owner actually needs.

Then I checked the price. The configuration I would recommend — 16GB memory, 512GB storage — went from RM4,699 to RM5,499. The fifteen-inch equivalent went from RM5,499 to RM6,299. An extra RM800, and not at a generational launch either: a repricing of a machine already sitting on the shelf.

It is not only the laptops. The Mac mini now starts at US$899, three hundred dollars above where the previous generation began. The new Mac Studio with the top chip opens at US$5,499, and the version carrying 512GB of memory, arriving in late October, is expected to land well beyond US$10,000.

Read the stated reason, because it is the whole story: component and memory costs.

The boom is repricing the ladder into it

Sit with that for a second.

Memory is expensive right now because AI infrastructure is consuming it. Every data-centre buildout, every accelerator, every training cluster is bidding for the same supply. That demand does not politely stay in the data centre — it arrives in the shop, on the laptop a woman needs in order to learn the technology that caused the price rise.

The boom is quietly repricing the ladder people would use to climb into it. Not by policy, not by anyone’s decision — simply by demand. And an extra RM800 is felt most sharply by exactly the person with the least room: the small operator who is not yet sure any of this will work for her.

I have written about a gap that gets manufactured early and published later. This is a smaller, faster version of the same machine.

The same question, at four different sizes

Here is what struck me about her question, because I have been writing about the identical problem all week at far larger scale.

Nations are asking it. Europe is moving government systems off American software, and Malaysia now administers another country’s export-control list as the price of hosting the hardware.

Organisations are asking it. The reason companies self-host open models is almost never cost — it is that the data cannot leave: regulated industries, client material under NDA, data-residency rules. Sovereignty, not savings.

My own company asks it. We keep a fallback stack and we are extending our runway, because access can be withdrawn faster than it can be replaced.

And now a hawker-centre supplier is asking it, without using any of those words. What do I need to own before I can start?

The question is the same at every scale. The right answer is not.

What I am going to give her

Not a shopping list. A plan.

I will show her what the first month looks like, what she will be able to do at the end of it, and what each path costs. Then she decides the investment. It is her money and her business, and the person selling the teaching should not also be the person deciding how much hardware she buys.

But I will not pretend I have no preference, because I do.

I want her on a Mac. Not because it is faster, and not because Apple needs the money. Because my entire team runs on the same platform — and when someone learning hits a problem at nine at night, the difference between “I know exactly what you are looking at” and “send me a screenshot, what version are you on” is the difference between five minutes and a week.

Consistency is not a luxury when someone is learning. It is most of the support.

The practical version of that is remote help. On the same platform, screen sharing is built in — nothing to install, no third-party account to create, no licence to buy — and because both screens look identical, every instruction translates exactly. “Top right corner, the icon beside the clock” means the same thing on her machine as on mine. Across different platforms, the first ten minutes go on installing something and finding where the equivalent button lives, before anyone has looked at the actual problem.

For someone learning alone at night, that gap is not a technicality. It is the difference between getting unstuck and giving up.

And I should disclose the obvious: that preference also serves me. It is easier to teach what I use every day. So I will say it to her plainly, the way I have said it here, and let her weigh it.

Which gives a simple framework. If the machine is comfortably affordable — buy it, and I can support her properly from day one. If RM5,499 is a stretch — do not stretch. Start on what she already owns; the first month is a browser and one real problem, and nothing in it requires the machine. Buy the good one when the work has proven it needs one.

She also does not need to host AI, whichever machine she picks. Open-weight models are genuinely excellent now — Qwen, Llama, DeepSeek and the open GPT releases are real options, and the tooling has settled — but self-hosting solves a problem she does not have. It exists for organisations whose data legally cannot leave the building. Her supplier prices and customer list are valuable, but nobody is going to subpoena them.

What I will not do is let the hardware decision become the reason she never starts. That is the real risk here — not buying the wrong laptop, but spending four months deciding and never opening the first tool.

To be fair — when you should buy the machine

I would be selling her a comfortable half-truth if I stopped there.

Free tiers have a ceiling, and hitting it is precisely when you should spend — free is a weapon you wield, not a diet you live on. If she ends up processing supplier documents daily, running long sessions, or handling anything with real customer data in it, the cheap machine becomes the bottleneck and the paid tools become the cheaper option. A bad laptop is its own tax: waiting on a slow screen kills the habit before the habit forms.

And there is one scenario where sovereignty does reach her. If she ever holds data she genuinely cannot send outside — a client’s proprietary formulation, a contract that forbids it — then local models stop being an abstraction. That day may arrive sooner than people expect, because it now tends to arrive as a compliance clause rather than a technology decision.

So the answer is not never buy. It is buy second, not first — after the work has told you what it needs.

The bottom line

If you are where she is: pick one task that costs you real time this week. Attack it with free tools on whatever device you already own. Buy hardware only when a specific job demands it.

And if you are the person she is paying — the consultant, the trainer, the vendor — the test is simple. Does your advice ever cost you a sale? Mine just did, and the course has not even started.