Capacity · GB
A model and its KV cache have to fit in memory. 24 GB runs a tuned 27B model, 128 GB of unified memory runs much larger ones, and 512 GB fits frontier-class models.
Local AI hardware, compared.
Ways to run AI models locally, compared on memory bandwidth, memory capacity and today's price, new or used: GPUs, Apple Silicon, unified-memory mini-PCs and datacenter accelerators, in one chart.
Prices updated 2026-09-29Indicative US market figures61 devices
Which bottleneck am I buying?
Ahmad Osman puts it this way: local AI is capacity × bandwidth × software stack. Capacity is what fits, bandwidth is how fast it can decode, and the stack decides how much of the spec sheet you get to use.
Bandwidth · GB/s
Decode is memory-bound, so tokens per second follows bandwidth. GPUs have the most; unified memory trades speed for size.
Software stack
The same chip can run several times faster on a tuned stack than at stock, and a model that fits does not yet serve well. The chart shows capacity and bandwidth; the stack is up to you.
How to read this chart.
Two questions decide a local AI box: does the model fit, and how much speed do you get for the money. Memory runs up the side and bandwidth per dollar runs across, with bubble size showing raw bandwidth, so up and to the right is best. Cheap gaming cards sit bottom right: good value, too little memory. Big-memory Macs and datacenter cards sit higher but further left: more room, less speed per dollar. Lucebox (yellow) sits high, with up to 192 GB of unified memory, real GPU bandwidth and a tuned stack, at a fixed price. Turn Datacenter off to see only what you can buy, and click Lucebox for the build.
Show
up = more memory (what fits) · right = more bandwidth per dollar (speed for the money) · bubble = raw bandwidth · both axes log · tap Lucebox for the build
Filter and compare.
Sort any column, or tick rows to compare a few. Value is memory × bandwidth ÷ price, so lots of fast memory for the money scores high. Price is the new price while a part is sold and the used price once it is discontinued. For Apple, mini-PCs and Lucebox, memory is the full unified pool.
| Device ⇅ | Class ⇅ | Memory GB ⇅ | Bandwidth GB/s ⇅ | Price ⇅ | Value ⇅ | |
|---|---|---|---|---|---|---|
| ★ Luceboxdetails → Lucebox Zero 495: Radeon AI PRO R9700 32GB + up to 192GB unified, tuned Lucebox engine | Lucebox | 160 u | 640 | $5,999 new | 17 | |
| MI400 / MI455X 432GB HBM4, first Helios rack shipments late Q3 2026; per-GPU price estimated | Datacenter | 432 | 19,600 | $50,000 new | 169 | |
| B200 192GB | Datacenter | 192 | 8,000 | $48,000 new | 32 | |
| B300 288GB Blackwell Ultra | Datacenter | 288 | 8,000 | $55,000 new | 42 | |
| MI325X 256GB | Datacenter | 256 | 6,000 | $26,000 new | 59 | |
| MI300X 192GB | Datacenter | 192 | 5,300 | $16,000 new | 64 | |
| H200 141GB | Datacenter | 141 | 4,800 | $37,000 new | 18 | |
| H100 80GB SXM | Datacenter | 80 | 3,350 | $32,000 new | 8 | |
| A100 80GB | Datacenter | 80 | 2,039 | $6,500 used | 25 | |
| RTX 5090 32GB Out of stock at US retailers; marketplace price vs $1,999 MSRP (GDDR7 shortage) | Consumer GPU | 32 | 1,792 | $6,450 new | 9 | |
| RTX PRO 6000 Blackwell 96GB 96GB workstation, 1792 GB/s | Workstation | 96 | 1,792 | $15,499 new | 11 | |
| RTX PRO 5000 Blackwell 72GB | Workstation | 72 | 1,344 | $4,900 new | 20 | |
| Mac Studio M5 Ultra 96GB Shipping since September 2026, 1.2 TB/s | Apple | 96 u | 1,200 | $5,499 new | 21 | |
| Mac Studio M5 Ultra 256GB 512GB ships late October | Apple | 256 u | 1,200 | $11,999 new | 26 | |
| AMD Instinct MI50 32GB Cheap-VRAM favorite, used only | Datacenter | 32 | 1,024 | $320 used | 102 | |
| RTX 3090 Ti 24GB | Consumer GPU | 24 | 1,008 | $1,300 used | 19 | |
| RTX 4090 24GB | Consumer GPU | 24 | 1,008 | $2,500 used | 10 | |
| RTX 5080 16GB | Consumer GPU | 16 | 960 | $1,650 new | 9 | |
| RX 7900 XTX 24GB | Consumer GPU | 24 | 960 | $900 used | 26 | |
| RTX 6000 Ada 48GB | Workstation | 48 | 960 | $6,800 new | 7 | |
| RTX 3090 24GB 24GB used-market favorite for local AI | Consumer GPU | 24 | 936 | $1,350 used | 17 | |
| RTX 3080 12GB | Consumer GPU | 12 | 912 | $340 used | 32 | |
| Intel Arc Pro B60 Dual 48GB Two B60 GPUs on one card | Workstation | 48 | 912 | $1,200 new | 36 | |
| Tesla V100 32GB | Datacenter | 32 | 900 | $700 used | 41 | |
| RTX 5070 Ti 16GB | Consumer GPU | 16 | 896 | $1,200 new | 12 | |
| Radeon PRO W7900 48GB | Workstation | 48 | 864 | $3,897 new | 11 | |
| Radeon PRO W7800 48GB Gigabyte AI TOP 48GB variant | Workstation | 48 | 864 | $2,700 new | 15 | |
| L40S 48GB | Datacenter | 48 | 864 | $8,500 new | 5 | |
| Mac Studio M3 Ultra 96GB Replaced by the M5 Ultra Mac Studio | Apple | 96 u | 819 | $3,300 used | 24 | |
| Mac Studio M3 Ultra 512GB 512GB retired; the M5 Ultra 512GB ships late October | Apple | 512 u | 819 | $30,000 used | 14 | |
| RX 7900 XT 20GB | Consumer GPU | 20 | 800 | $700 used | 23 | |
| Mac Studio M2 Ultra 192GB | Apple | 192 u | 800 | $5,500 used | 28 | |
| RTX A6000 48GB | Workstation | 48 | 768 | $3,500 used | 11 | |
| RTX 3080 10GB | Consumer GPU | 10 | 760 | $370 used | 21 | |
| RTX 4080 Super 16GB | Consumer GPU | 16 | 736 | $1,100 used | 11 | |
| RTX 4080 16GB | Consumer GPU | 16 | 717 | $1,080 used | 11 | |
| RTX 4070 Ti Super 16GB | Consumer GPU | 16 | 672 | $770 used | 14 | |
| RTX 5070 12GB | Consumer GPU | 12 | 672 | $1,000 new | 8 | |
| RX 9070 XT 16GB | Consumer GPU | 16 | 640 | $800 new | 13 | |
| RX 9070 16GB | Consumer GPU | 16 | 640 | $690 new | 15 | |
| Radeon AI PRO R9700 32GB | Workstation | 32 | 640 | $1,800 new | 11 | |
| RX 7800 XT 16GB | Consumer GPU | 16 | 624 | $500 used | 20 | |
| RTX 2080 Ti 22GB (modded) China-modded 22GB refurb | Consumer GPU | 22 | 616 | $550 used | 25 | |
| MacBook Pro M5 Max 128GB | Apple | 128 u | 614 | $6,749 new | 12 | |
| Tenstorrent Wormhole n300 24GB RISC-V Tensix, fully OSS stack | Workstation | 24 | 576 | $1,399 new | 10 | |
| Mac Studio M4 Max 128GB 40-core; 128GB pulled, max now 96GB | Apple | 128 u | 546 | $3,200 used | 22 | |
| RX 6800 16GB | Consumer GPU | 16 | 512 | $380 used | 22 | |
| Tenstorrent Blackhole p150 32GB RISC-V Blackhole, fully OSS stack | Workstation | 32 | 512 | $1,399 new | 12 | |
| RTX 4070 12GB | Consumer GPU | 12 | 504 | $500 used | 12 | |
| Intel Arc Pro B60 24GB | Workstation | 24 | 456 | $674 new | 16 | |
| RTX 3060 Ti 8GB | Consumer GPU | 8 | 448 | $240 used | 15 | |
| Mac Studio M4 Max 36GB 32-core GPU, ~410 GB/s; replaced by the M5 Max Mac Studio | Apple | 36 u | 410 | $2,100 used | 7 | |
| RTX 3060 12GB | Consumer GPU | 12 | 360 | $300 used | 14 | |
| Tesla P40 24GB Cheap 24GB, no video out, needs blower | Datacenter | 24 | 347 | $270 used | 31 | |
| MacBook Pro M5 Pro 64GB | Apple | 64 u | 307 | $3,499 new | 6 | |
| Mac mini M4 Pro 64GB 64GB config pulled (DRAM shortage) | Apple | 64 u | 273 | $2,200 used | 8 | |
| DGX Spark 128GB (GB10) Coherent memory + CUDA | Mini-PC | 128 u | 273 | $6,500 new | 5 | |
| Ryzen AI Max+ 395 128GB Strix Halo, ~96GB allocatable to GPU | Mini-PC | 128 u | 256 | $3,499 new | 9 | |
| Jetson AGX Orin 64GB 275 TOPS edge dev kit; was $1,999 before July 2026 | Mini-PC | 64 u | 204 | $3,499 new | 4 | |
| MacBook Air M5 32GB | Apple | 32 u | 153 | $1,499 new | 3 | |
| Mac mini M4 16GB Replaced by the M6 Mac mini in September 2026 | Apple | 16 u | 120 | $850 used | 2 |
Method. Memory bandwidth is peak theoretical. Each device shows the price that actually applies: the current new price while it is still sold, the current used price once it is discontinued. Figures are indicative US-market numbers checked across retailers (Newegg, Amazon, B&H, Apple) and used marketplaces (eBay sold listings, resellers) as of 2026-09-29; the 2026 memory shortage has pushed new prices up and pulled several high-memory configs from sale. Value = memory × bandwidth ÷ price (capability per dollar). Datacenter parts are quote-based and sold in 8-GPU boards, so their per-unit figures are approximate. "u" marks unified memory. Numbers move fast; treat them as a snapshot, not a quote.
Full disclosure: we make Lucebox.
A Radeon AI PRO R9700 with 32 GB of GDDR6 at 640 GB/s, paired with 128 GB of unified memory (192 GB as an option), a tuned stack and a warranty. The numbers above let you judge it against the rest.
Cost, setup, speed and support, side by side.
The comparison page puts Lucebox next to cloud APIs, a DGX Spark and a Mac Studio, row by row.