Should your business buy its own AI hardware? A plain buyer's note

It's a question I had to answer for my own lab this month, and it's one any Hunter business looking at AI will run into. "Should we buy our own AI box and run it here, or keep paying for ChatGPT, Claude, Copilot, Gemini, Grok and the like?"

It's a fair question, and running AI locally gets talked up a lot online. Here's my take, with Australian prices I checked in October 2026.

What running AI locally means

When you use a cloud AI tool, your question goes over the internet to a big data centre, gets answered, and comes back. Running it locally means the AI model (the part that does the work) sits on a computer in your own office, and your data stays in the building.

The catch is memory. The larger open models, which are free to download and the closest thing to the big cloud tools, usually need a machine with about 128GB of memory. Smaller models run in 64GB, and NVIDIA now argues they are good enough for a lot of everyday work (The Register). Memory is where most of the money goes.

What it costs in Australia right now

These are new prices including GST, as listed on 12 October 2026.

• Apple Mac Studio with M5 Max, 128GB memory and 1TB storage. A$8,649 at the Apple AU store, on a four to five week wait.
• AMD Ryzen AI Max+ 395 mini PCs with 128GB. About A$5,100 to A$5,300. The MINIX ER939-AI is A$5,079 at PB Tech but out of stock, and the GMKtec EVO-X2 is A$5,256 on Amazon AU.
• NVIDIA DGX Spark with 128GB and 4TB. About A$8,700 to A$9,200, for example A$8,699 at Radium PCs and A$9,199 at Adelaide Online Computers.

Used high-memory Mac Studios come up on eBay in Australia too, and the few I found are in the resale section below.

So you're looking at about A$5,000 to A$9,000 for one machine, before any setup time.

What's happening to prices

New hardware is going up. On 2 October NVIDIA raised the US price of its 128GB DGX Spark to US$6,950, nearly 75% more than a year earlier (The Register). The Australian shop prices above hadn't caught up when I checked. Apple lifted Australian Mac starting prices by 17 to 23% in June (Information Age), and the Mac Studio still starts at the same A$4,299.

The cause is memory chips, which AI data centres are buying up. TrendForce expects DRAM (the main type of computer memory) contract prices to rise another 10 to 15% in the last quarter of 2026, which is still a rise but a slower one (TrendForce). Micron, one of the biggest memory makers, expects supply to be much tighter in 2027 and 2028 than it was this year (TechPowerUp).

Used hardware is getting cheaper. Hard Bargains tracks US second-hand listings. After Apple released new models on 25 August, asking prices on the same used M3 Ultra Mac Studio listings fell about 16.5% by 3 October, and M4 Max Mac Studios about 12.8% (Hard Bargains). These figures are US asking prices.

So buy when the sums work for your business, whatever prices do next, and bargain hard on used gear.

The real cost is price less resale

The real cost of AI hardware is what you pay for it, less what you get back when you sell it, plus the power and time to run it. That net figure is the one to budget around, and for a machine that holds its value it can be a lot smaller than the price tag.

On the little I found, high-memory Macs seem to hold their value in Australia. In September a 2022 Mac Studio with an M1 Ultra and 64GB sold on eBay for about A$4,750 before eBay's buyer fee, a bit under 80% of its A$6,099 launch price. A 2023 M2 Ultra with 128GB is listed at A$8,211, about three quarters of its A$10,799 launch price, as an asking price. It's a small sample, and US used prices have eased since August, so plan on a cautious resale figure. As a rough guide, if a Mac Studio kept about two thirds of its price over three or four years, an A$8,649 machine would cost about A$2,900 net, before selling fees. I haven't found Australian resale figures for the AMD or NVIDIA boxes.

When local makes sense

Local AI can be the right call in three cases.

• Your data can't leave the building. Contracts, client confidentiality, health or legal records, or rules you have to follow.
• You'd use it heavily every day, so the machine stays busy.
• You need it to work offline, such as sites with poor or no internet.

When cloud is the better choice

For most Hunter small businesses, cloud AI is still the better choice today. You pay monthly, get the strongest models, someone else handles updates and security, and you can stop any time. A local box only pays off if it replaces a lot of cloud spending or solves a problem the cloud can't.

What I decided for myself

On 5 October I decided against buying a 128GB machine for my own lab, for now.

For Pairti, paying monthly for cloud AI wins over buying hardware right now. A local box would take too long to pay for itself, because none of the three cases above apply to us: our own data can go to cloud services, AI included; we wouldn't keep a machine busy every day; and we always have internet. The cloud models are stable and good. Getting good at using AI on real problems matters more than owning hardware.

I expect that to change. As cloud prices start to reflect what these models really cost to run, and as free open models keep getting better, owning the hardware will look better for more businesses, including mine. So I'm keeping an eye on it and on the prices, and I'll run the numbers again.

A quick self-check

1. Do we have data that must stay in the building? Think contracts, client confidentiality, or rules you have to follow.
2. Would we use it hard, every working day? Or would it mostly sit there?
3. Do we need AI to keep working with no internet?
4. Is cloud AI already doing real work for us? If not, start there. A monthly plan costs little, and it shows you which jobs AI handles well before you spend on hardware.

If you answered no to the first three, stick with cloud for now. If you said yes to one or more, local may be worth a closer look, with proper sums done first.

Want to talk it through?

If you're weighing this up, use the chat button on this page and I'd be glad to talk. I'll look at what your business actually does, and we'll work out together whether local AI makes sense for you or whether it can wait.