AI memory per dollar: Mac Studio, Strix Halo and Nvidia compared

Strix Halo is the cheapest local AI memory in this comparison at $27 per gigabyte. Used Mac Ultras cost $34-$41 per gigabyte but provide roughly three times the memory bandwidth. Nvidia’s 32GB and 96GB cards lead on bandwidth and cost $147-$167 per gigabyte.

What to buy for the model you want to run

Two numbers decide whether a model runs: the weights have to fit, and your context needs room on top. A third, memory bandwidth, decides how fast. The tool takes your budget and what you already own and answers keep, add a card or buy a machine, with the dated price and the retailer in the same view.

Model	Cheapest upgrade to a PC you own	Cheapest whole machine
The machine you already own — $0. Check this first. If it already has the “to work” figure free, keep it and spend nothing. The tool asks what you own before it shops and names what you would otherwise have paid.
Llama 3.1 8B4.5GB weights · 7.5GB to work	RTX 5060 Ti 16GB, new$790 · 16GB · 448GB/s	Mac mini M4, 16GB / 256GB, used$749 · 10.7GB · 120GB/s
Qwen 3.5 27B15.5GB weights · 25.9GB to work	Two used RTX 3090s$2,064 · 48GB · 936GB/s per card, not pooled	Mac Studio M5 Max, 36GB, new$2,499 · 27GB · 460GB/s
Llama 3.3 70B39.1GB weights · 65.2GB to work	No route we can source	GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · 128GB · 256GB/s
gpt-oss-120b65.4GB weights · 109GB to work	No route we can source	GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · 128GB · 256GB/s

Cheapest route that clears the recommended working figure, ranked the way the tool ranks it: on the total you pay — the dated price plus anything the price leaves out — with anything we found out of stock ranked below what you can order today. A card needs a desktop PC, a supply and the clearance to take it; a whole machine does not. The first answer is always to check what you already own — the tool asks that first, and says keep it when it fits.

The exact products behind those two figures

• RTX 5060 Ti 16GB, new$790 · lowest new listing we quote (Newegg), Sept. 15, 2026

  • ASUS Dual RTX 5060 Ti 16GB GDDR7 OC — 2.5 slotsAmazon · exact productOne specific board, not the listing the dated price above came from. No price or stock is shown here for it; check both on the listing.

• GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB$3,499 · maker's own store price, 128GB + 1TB, Sept. 16, 2026

  • GMKtec's own store — the $3,499 128GB + 1TB configurationGMKtec · exact product · where the price above was readOpens on 128GB RAM + 1TB SSD / USThis is the price and stock state we read ourselves.
  • GMKtec EVO-X2, 128GB LPDDR5X, 2TB SSDAmazon · exact productThe 2TB configuration, not the 1TB priced above: the difference is storage, not memory for a model. No price or stock is shown here for this listing.

Other configurations and the option where you buy nothing are on the tool.

Answer it for my model and budget

A machine for running models locally is mostly a memory purchase: the model must fit in memory the processor can address, and token speed is bounded by how fast that memory moves. Price every machine on those two axes and the market sorts into three groups that overlap much less than their stickers suggest. For measured Mac speeds and which Mac fits a given model, see how many tokens per second a Mac actually gets.

Choose by the constraint that decides

Find the first row that describes you. That constraint picks the platform before price per gigabyte does.

Deciding constraint	Buy	Memory, bandwidth	Price (dated)	What you give up	Where to check
You already own 32GB or more of unified memory or VRAM	Nothing yet	What you own	$0	Nothing, if the model’s minimum fits with room to spare	canirun.ai minimums; Mac fit by model
CUDA tools, models up to about 27B at 4-bit	RTX 5090	32GB, 1.79TB/s	~$4,700 street as reported in August; not rechecked Sept. 15	A 4-bit 70B model won’t fit; $147/GB; needs a PC to install it in	Nvidia’s RTX 5090 page, which showed no Founders Edition price Sept. 15; Amazon search results
CUDA with a 70B model on one card	RTX Pro 6000	96GB, 1.79TB/s	$16,000 on Nvidia’s marketplace in August; not rechecked Sept. 15	$167/GB	TechSpot’s Aug. 13 report
CUDA with 128GB in one small box	DGX Spark	128GB, 273GB/s	$4,699 as reported in August; not rechecked Sept. 15	Speed: about a third of a used M1 Ultra’s bandwidth	Nvidia’s DGX Spark page
Most memory per dollar, with speed secondary	Framework Desktop or another Ryzen AI Max+ 395 (Strix Halo) mini PC	128GB, 256GB/s	$3,449 at Framework, listed out of stock Sept. 15	70B at 4-bit estimated at single-digit to low-double-digit t/s; no measured figure cited	Framework Desktop; Amazon search results
Large models at usable speed in one box	Used Mac Studio M1 Ultra	128GB, 800GB/s	$5,249 median ask (Aug. 18-25)	No CUDA; $41/GB	M1 Ultra price tracker
The same, with 256GB or 512GB	Used Mac Studio M3 Ultra	256GB or 512GB, 819GB/s	$9,999 or $17,448 median ask (Aug. 18-25)	No CUDA; $39/GB or $34/GB, at high absolute prices	M3 Ultra price tracker; Apple’s Mac Studio store for new models
More than 512GB	No single machine here	—	—	The simplicity of one box	The cluster math

Fit follows canirun.ai’s minimums at roughly 4-bit quantization: Qwen 3.5 27B needs 15.5GB and fits on a 32GB card; Llama 3.3 70B needs 39.1GB and doesn’t. Longer prompts need memory beyond the minimum.

Buyers comparing 16GB, 24GB and 32GB cards, including a used RTX 3090, can use the GPU chooser to match a model to a card with measured speeds and power-supply needs.

Both Amazon links are links to search results, not specific products. Ryzen AI Max+ 395 mini PCs are sold in several memory configurations, so confirm 128GB on the exact listing. RTX 5090 results can include prebuilt PCs and other cards, so confirm the card itself and its 32GB. The prices on this page come from manufacturers, published reports and our used-Mac crawler, not from Amazon.

Local AI memory prices compared

Prices as reported in August 2026; used Mac Studio prices are median asks from Aug. 18-25. The Sept. 15 checks are in the chooser table.

Machine	Memory	Bandwidth	Price	$/GB
Framework Desktop (Strix Halo)	128GB	256GB/s	$3,449 new	$27
Nvidia DGX Spark	128GB	273GB/s	$4,699 new	$37
Mac Studio M1 Ultra, used	128GB	800GB/s	$5,249	$41
Mac Studio M3 Ultra, used	512GB	819GB/s	$17,448	$34
Mac Studio M3 Ultra, used	256GB	819GB/s	$9,999	$39
Mac Studio M5 Ultra	96GB	1.2TB/s	$5,499 new	$57
Mac Studio M5 Max	36GB	460GB/s	$2,499 new	$69
RTX 5090	32GB	1.79TB/s	~$4,700 street	$147
RTX Pro 6000	96GB	1.79TB/s	$16,000	$167

Which machine has the cheapest local AI memory?

If the question is purely “how many gigabytes can I point a model at per dollar,” the Strix Halo machines win: AMD’s Ryzen AI Max+ 395 with 128GB of unified LPDDR5X, sold as the Framework Desktop and several mini-PCs, at $27 a gigabyte. It is also x86 Linux, repairable and ugly in the way a person who distrusts aluminum jewel boxes might prefer. Which of those boxes to actually buy — with maker and retailer prices checked Sept. 15 and how much of the 128GB a model can address — is in the 128GB mini PC comparison.

The catch is the second axis. That memory moves at 256GB/s — a third of a used M1 Ultra’s bandwidth, a fifth of the M5 Ultra’s — and token generation is bandwidth-bound. Compute Market’s comparison expects 4-bit 70B generation to run at single-digit to low-double-digit tokens per second on both Strix Halo and Nvidia’s DGX Spark, because generation is bandwidth-bound; it cites no measured 70B generation figure. The models fit; they don’t hurry. The DGX Spark makes the same trade at a worse price, and sells CUDA and Nvidia’s stack for the difference.

Which machine offers the most memory bandwidth for the price?

Among machines that hold serious models, the Macs own the bandwidth column: 800GB/s on Ultras Apple sold in 2022, 1.2TB/s on the new M5 Ultra. Nvidia’s cards move memory faster still — but at $147 to $167 per gigabyte, capacity is what you give up. A 32GB RTX 5090 cannot hold a 4-bit 70B model at all, and the 96GB card that can costs $16,000 after its August repricing, more than two used 96GB M3 Ultras ask.

Which is why the used Ultras are the quiet value rows. A used M1 Ultra with 128GB — the same memory count as the Strix Halo box, at three times its bandwidth — ran a $5,249 median ask in our Aug. 18-25 observations. The used 256GB and 512GB M3 Ultras price memory at $34-39 a gigabyte with 819GB/s behind it: no other path to a half-terabyte of model-addressable memory in one box comes close. Their absolute prices are high because the shortage repriced them; their per-gigabyte prices are mid-table anyway.

Above 512GB, the question changes from choosing one machine to pricing a Mac Studio cluster.

Why local AI hardware got more expensive in 2026

The comparison is a snapshot of a market that has repriced violently in one direction. The Framework Desktop’s 128GB configuration sold for $1,999 in mid-2026 and was out of stock at $3,449 in early August. Nvidia raised the DGX Spark from $3,999 to $4,699 citing the LPDDR5X shortage, and took the RTX Pro 6000 from $8,565 at launch to $16,000 in about 16 months. Against that, Apple’s June repricing — 14-33% across the Mac line — was, strange to write, the restrained one. Memory is what all of these machines are made of, and memory is what the AI buildout is consuming; nobody sells an exit from that.

For what these memory sizes actually run, model by model, see the fit table; for what local capacity is worth against a subscription, see the subscription comparison. And if the numbers here age — they will — the data note carries the dates.