Local AI Hardware for Small Business: What to Buy in 2026

The right local AI box depends on the job. For a private office chatbot, a Mac mini is the simplest buy: the M6 model with 32GB of memory is $1,299 and the M5 Pro model with 64GB is $2,699. For cameras on a production line, an NVIDIA Jetson Orin Nano Super Developer Kit is $399, and a Raspberry Pi 5 with an AI HAT+ handles lighter vision work. Memory costs pushed most of these prices up in 2026, so quote from the day's price.

This page is part of the edge AI series. It covers what to buy. The software that runs on it is in the local LLM stack.

Which box fits which job?

Start from the client's job, then pick the cheapest box that does it with room to spare. Prices are US list prices checked in October 2026. The user counts are our estimates for planning; test with the client's real load.

Why did local AI hardware get more expensive in 2026?

Memory. Every box on this page is mostly a memory purchase, and memory prices climbed through 2025 and 2026.

• Raspberry Pi raised Pi 5 prices four times between December 2025 and October 2026, citing "a seven-fold increase over the last year in the price of the LPDDR4". The 16GB Pi 5 is now $305, up from $120 at launch.
• NVIDIA raised Jetson prices in July 2026 by 33% to 101%. The Orin Nano Super kit went from $249 to $399, the AGX Orin 64GB kit from $1,999 to $3,499 and the AGX Thor kit from $3,499 to $5,499. CNX Software has the full old and new table.
• DGX Spark, NVIDIA's 128GB desktop AI box, launched at $3,999. The Register reported in October 2026 that NVIDIA raised it to $6,950 and added a 64GB model at about $4,999.
• Graphics cards are the worst hit. Tom's Hardware's tracker, updated 22 October 2026, shows the RTX 5060 Ti 16GB at $765 against a $429 launch price, and the cheapest RTX 5090 listing at $8,599 against $1,999.
• Mac mini went up less. The previous base model rose from $599 to $799 in June 2026, according to MacRumors, and the new M6 model starts at $899.

For an installer this has two effects. Quote hardware at the day's price and hold the quote for a week at most. And do not stock boxes on spec: a price change can wipe out your margin on stock you have not sold.

Is a Mac mini good for running a local LLM?

For office chat it is the easiest box to support. Apple silicon shares one pool of memory between the processor and the graphics cores, so a 64GB Mac mini can load a model that would need a 64GB graphics card on a PC. Ollama and LM Studio both run on it natively.

The August 2026 models are the M6 and the M5 Pro. The M6 starts at $899 with 16GB and goes up to 32GB. The M5 Pro starts at $1,699 with 24GB and goes up to 64GB, with 307GB/s of memory bandwidth. Bandwidth sets how fast a model writes its answer, so the Pro answers noticeably faster on the same model.

AppleInsider's price guide lists these US prices:

Power is the Mac mini's other strength. Apple's power consumption page rates the M6 model at 4 watts idle and 70 watts at most, and the M5 Pro with 64GB at 6 watts idle and 145 watts at most.

The weak spot is concurrency. One Mac answers one long request at a time well. If 20 people ask at once they queue, so set expectations in the demo.

Buy the memory you need on day one. Memory on Apple silicon cannot be upgraded later, and EveryMac reports Apple charges $400 to go from 16GB to 32GB on the M6.

When does an NVIDIA Jetson make sense?

When the job is cameras. Jetson boards are built for vision on the factory floor: they run NVIDIA's CUDA software, which most vision models are trained on, and they draw little power.

The Jetson Orin Nano Super Developer Kit gives 67 INT8 TOPS (up from 40 TOPS before a software update), 8GB of memory at 102GB/s and a 7 to 25 watt power range, for $399. It handles one or two camera streams running a small YOLO detection model.

Above it, NVIDIA no longer sells its own Orin NX developer kit. Partner boxes such as Seeed's reComputer J4012 with an Orin NX 16GB list at around $1,399 to $1,599 from retailers. The Jetson AGX Orin 64GB Developer Kit is $3,499 and the AGX Thor kit is $5,499. Those are for many cameras or heavy models, and most small factory pilots do not need them.

A developer kit is fine for a pilot. For a box that will sit on a line for years, quote an industrial enclosure from a Jetson partner, with a fanless case and wide temperature range. It costs more, and dust kills fans.

Can a Raspberry Pi 5 run AI?

Yes, for counting and simple detection, and for very small language models.

The Pi 5 itself now costs $45 for 1GB, $77.50 for 2GB and $305 for 16GB, after four rounds of rises. Adding the December 2025, February 2026 and April 2026 rises to the launch prices puts the 8GB board at about $175. That is our arithmetic from Raspberry Pi's posts, not a published list.

The accelerators:

• AI HAT+, from $70: 13 TOPS (Hailo-8L) or 26 TOPS (Hailo-8). Good for people counting and object detection on one camera.
• AI HAT+ 2, $200: a Hailo-10H with 40 TOPS (INT4) and 8GB of its own memory, which Raspberry Pi describes as "ideal for running large language models (LLMs) and vision-language models (VLMs) locally". Read "large" as small: 8GB holds a model of a few billion parameters.
• AI Camera, $70: a Sony IMX500 sensor that runs a detection model on the camera itself.

The older Raspberry Pi AI Kit is out of production.

A Pi is the cheapest honest demo of camera AI. It is also the box most likely to be unplugged by a cleaner. For a paying shop, put it in a proper case with a locked power supply.

Should you use Coral or an Intel NPU?

Skip Coral for new installs. Google's Coral site now points to a new RISC-V Coral NPU platform, and XDA reported in July 2026 that "most of the libraries ... haven't been updated since 2022" and that the Frigate camera project stopped recommending Coral from version 0.17. Stock is still around, but you would be installing something nobody updates.

Intel NUC boxes now come from ASUS, which took over the NUC line from Intel in 2023. The ASUS NUC 15 Pro advertises "up to 99 Platform TOPS". Read the small print: that adds the processor, graphics and NPU together. Intel's own sheet for the Core Ultra 7 255H puts the NPU alone at 13 TOPS. A NUC makes a fine small office server, but it is not a faster AI box than a Mac mini at the same price.

Which box runs the biggest models?

For 70B-class models you need around 64GB to 128GB of fast memory. Three routes, all expensive after 2026:

• AMD Ryzen AI Max+ 395 mini PCs. The GMKtec EVO-X2 lists at $2,199.99 for 64GB, and VideoCardz reported $3,499 for 128GB in July 2026. GMKtec rates it at 120 watts sustained.
• Mac Studio. Apple's M5 Max model starts at $2,499 and goes to 128GB; the M5 Ultra starts at $5,499 and goes to 512GB.
• NVIDIA DGX Spark. 128GB for $6,950 after October's rise, or the new 64GB version at about $4,999.

Big models answer slowly on all of these compared with a cloud service, and most office tasks run fine on a 20B to 30B model. Sell the big box only when a client has tested the smaller one and found it lacking.

What does a local AI box cost to run?

Less than most clients expect. The US average commercial electricity price was 14.47 cents per kWh in August 2026, according to the EIA. A box that averages 30 watts around the clock uses about 22 kWh a month, which is about $3.20 at that price.

Elsewhere the price per kWh is higher. UK non-domestic users paid an average of 24.51 pence per kWh in the second quarter of 2026, and Singapore's household tariff for October to December 2026 is 28.59 cents per kWh before GST. Even at those prices a Mac mini or Jetson costs a few dollars a month to run. A tower with a big graphics card under constant load is a different bill, so measure it with a plug meter during the pilot.

Put your own hardware price, wattage and local electricity price into the local AI vs cloud calculator to see the payback for a client.

How should you buy hardware for a client?

1. Scope first, then buy. Run the client's real documents or camera feed on your own demo box before you quote. The pilot tells you how much memory you need.
2. Quote hardware separately and date it. List the box, its price on the day and the date the quote expires. Prices moved several times in 2026.
3. Buy in the client's name where you can. The warranty sits with the owner, and you avoid carrying stock.
4. Buy one size up on memory. Models get bigger, and memory is the one part that cannot be added later on a Mac or a Jetson module.
5. Keep a spare for the demo. Your demo box is your sales tool. Do not sell it to the first client who asks.

Pricing the box, the setup and the monthly retainer is covered in how to price local AI installs.