What hardware should you buy to run AI at home in late 2026?

Advice on buying local AI hardware goes stale quickly this year, because the price of memory keeps moving. NVIDIA's 128 GB DGX Spark launched at US$3,999 and now lists at US$6,950. Framework's 128 GB desktop, US$1,999 at its February 2025 launch, was US$3,449 by July. So every price below carries the date it was read, and you should check it again before paying.

The guide's 2024 hardware overview explains why memory matters, and the October 2026 setup guide covers installing the software. This article answers the next question: what to buy, at what budget.

Three numbers decide the purchase

Memory the processor can reach quickly. On a PC with a graphics card, that is the card's VRAM. On an Apple Mac or an AMD Ryzen AI Max machine, it is a shared pool called unified memory. This sets the largest model you can load.

Memory bandwidth. Measured in gigabytes per second, it sets how fast words appear once a model is loaded. A Mac mini M6 starts at 153 GB/s, a DGX Spark has 273 GB/s, and an M5 Ultra has 1.2 TB/s.

Software support. Tools usually arrive first for NVIDIA's CUDA platform. That matters little for chatting and a lot for training, as the Mac versus PC comparison explains.

Budget tiers, with prices read November 2, 2026

Prices exclude tax. Apple and Framework figures come from their store pages as shown to a Canadian visitor. Graphics card figures are the cheapest new US listings on the Tom's Hardware index for October 26, as read by PC Server and Parts, a used-card seller that discloses its interest. They cover the card alone, not the computer around it.

"B" means billion parameters, the numbers a model learned in training. The measured sizes are Framework's own, from its September 2026 reference tests.

Two things stand out. The 16 GB graphics cards cost almost as much as a small computer: the RTX 5060 Ti 16 GB launched at US$429. And prices disagree between sources: the same index lists the RTX 5080 at US$1,699 while a Hardware Unboxed survey found US$2,200 that month. Treat any single number as a snapshot.

A worked example: a private coding assistant

Suppose you want a local model to work on code you cannot send anywhere. The open coding models guide points to Qwen3.8-27B, an 18 GB download. Ollama recommends at least 64,000 tokens of context for coding tools, and that context needs memory beside the model.

• 16 GB: the model does not load. Skip this tier for that goal.
• 24 GB: it loads with little left over, so contexts stay modest.
• 32 GB shared: it fits with room for context, since the system and the model share the pool. Expect modest speed.
• 64 GB: it fits with plenty of context, and you can try larger models.

So the sensible spend for this person starts at the CA$1,789 tier, not at a CA$7,799 machine. The extra money buys speed and bigger models, which only help if you will really use them.

Common mistakes

• Buying the chip instead of the memory. A faster processor with too little memory cannot load the model at all.
• Trusting a parameter ceiling in marketing. Those figures assume heavy compression and little context. The memory calculator shows how to check.
• Forgetting that shared memory is soldered. Macs and Ryzen AI Max machines cannot be upgraded later, so choose the memory you will need.
• Ignoring running costs and noise. A big graphics card draws far more power than a small box. The power and noise guide has the figures.

Should you buy now, or rent until prices settle?

Nobody can tell you when prices will fall. Micron, one of the largest memory makers, says it cannot see when supply and demand will balance and expects the crunch to worsen in 2027 and 2028. That is a company's forecast rather than a certainty, so waiting may or may not pay off.

The more useful test is why you want local AI. If the reason is saving money, the numbers rarely work: hosted gpt-oss-20b was listed on OpenRouter from about US$0.066 per million output tokens, which is close to what electricity alone costs to run it at home. If the reason is privacy, offline use or learning how these systems behave, then the machine buys something a hosted service cannot sell you.

A cautious path is to start with the computer you already own and a small model. Next, test the model you want through a hosted API or a rented graphics card for a few weeks. Buy only once you know which model you will really run and how much context you need. That way the purchase is sized by your own measurements, not by a spec sheet.