Locally Explains

The Best Local AI Machine For $2,000 Has A Catch

Every figure shown on screen, with where it comes from. Prices and listings were checked on 4 October 2026 unless a different date is given. Prices move daily; treat every price here as a snapshot.

NVIDIA RTX 5080 and RTX 5070 Ti

• RTX 5080: 16 GB GDDR7, “Starting at $999”. NVIDIA, GeForce RTX 5080 product page. https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/
• RTX 5070 Ti: 16 GB GDDR7, “Starting at $749”. NVIDIA, GeForce RTX 50 Series 5070 family page. https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/
• Memory bandwidth: RTX 5080 960 GB/s, RTX 5070 Ti 896 GB/s. NVIDIA specification pages for each card.
• NVIDIA’s US marketplace showed no Founders Edition RTX 5080 on 4 October 2026. Every RTX 5080 listed was a partner card, out of stock, at $1,699.99 to $2,099.99. https://marketplace.nvidia.com/en-us/consumer/graphics-cards/
• RTX 5070 Ti street price: GPU Poet reports a lowest average price of $1,096 in September 2026, ranging from $968 to $1,200, with tracked deals starting at $1,112. https://gpupoet.com/gpu/learn/price/september-2026/nvidia-geforce-rtx-5070-ti

System memory

• “The cheapest 32GB DDR5 RAM you can buy is $375 … $374.97 to be precise”, per PCPartPicker tracking. Tom’s Hardware, 3 June 2026. https://www.tomshardware.com/pc-components/ddr5/32gb-of-ddr5-now-costs-usd375-minimum-ai-shortage-continues-to-squeeze-pc-building

Published speed results

• RTX 5080, deepseek-r1:14b (the R1 Distill Qwen 14B, 9 GB in Ollama): “about 70 tokens per second performance with a context window up to 16k”, using the prompt “tell me a story”. Above 16k the output drops to 19 tokens per second. Windows Central, Richard Devine, 25 August 2025. https://www.windowscentral.com/artificial-intelligence/just-what-sort-of-gpu-do-you-need-to-run-local-ai-with-ollama-the-answer-isnt-as-expensive-as-you-might-think
• RTX 3090, compiled community results. SpecPicks, “RTX 3060 12GB vs RTX 3090 for Local LLMs (2026)”, updated 1 October 2026:
  • Qwen3 35B-A3B (sparse mixture of experts, 3B active parameters): 112 to 135.7 tokens per second.
  • 12B to 14B dense class (Qwen3 14B among them): 52.1 to 55.8 tokens per second. https://specpicks.com/reviews/rtx-3060-12gb-vs-rtx-3090-local-llm-2026
• Strix Halo (Ryzen AI Max+ 395, 128 GB), one benchmark repository:
  • gpt-oss-120b, MXFP4 weights, llama.cpp with Vulkan: 53.7 tokens per second.
  • llama3.1:70b (dense, 43 GB) in Ollama: 5.6 tokens per second. https://github.com/huppiflupp/strix-halo-llm-speeds
• Framework Mainboard with Ryzen AI Max+ 395, 128 GB, Llama 3.2 3B: 88.14 tokens per second. Jeff Geerling, ai-benchmarks. https://github.com/geerlingguy/ai-benchmarks

Model file sizes

• Qwen3 14B and DeepSeek R1 Distill Qwen 14B at Q4_K_M: about 9.0 GB each (Hugging Face GGUF listings).
• Qwen2.5 32B Instruct at Q4_K_M: about 19.9 GB (Hugging Face GGUF listing).
• Qwen3.6 35B A3B at Q4_K_S: about 20.9 GB (unsloth GGUF listing on Hugging Face).
• gpt-oss-120b MXFP4 weights: 63.4 GB as listed in the Strix Halo benchmark repository above.

Used RTX 3090

• 24 GB GDDR6X, 350 W graphics card power, 313 mm long, three slots. NVIDIA, GeForce RTX 3090 Founders Edition specifications.
• Used prices: “$1,200-1,400 as of August 2026”. Of 52 completed used eBay sales the median was $1,275 and the middle half sold at $1,200 to $1,300; “Nothing sold below $1,099”. InsiderLLM, used RTX 3090 buying guide, updated 2 October 2026. https://insiderllm.com/guides/used-rtx-3090-buying-guide/
• Capacity per dollar, our arithmetic: 24 GB at the $1,275 median sale is about 19 GB per $1,000; 16 GB at $749 is about 21 GB per $1,000; 16 GB at $999 is 16 GB per $1,000.

AMD Ryzen AI Max+ 395

• Radeon 8060S graphics, up to 128 GB of memory. AMD product page. https://www.amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-max-plus-395.html
• Cheapest 128 GB system in the comparison: Bosgame M5 at $3,324.04 with a 2 TB SSD, on 28 September. Cheapest 395 mini PC in stock: GMKtec EVO-X2 with 64 GB and 1 TB at $2,199.99. ComputingForGeeks, 2 October 2026. https://computingforgeeks.com/ryzen-ai-max-395-mini-pc-comparison/

Apple

• Mac Studio: “Buy from $2499” (M5 Max). Apple Store. https://www.apple.com/shop/buy-mac/mac-studio
• Mac mini M6: from $899, 16 GB unified memory, configurable to 24 or 32 GB; 153 GB/s on the 16 GB models and 170 GB/s on the 24 GB model. Mac mini M5 Pro: from $1,699, 24 GB, configurable to 48 or 64 GB, 307 GB/s. Apple, Mac mini tech specs and store. https://www.apple.com/mac-mini/specs/ https://www.apple.com/shop/buy-mac/mac-mini

Not checked

• Used RTX 3090 for $700 to $1,000: current sources put used sales at $1,200 to $1,400.
• Qwen3 14B at about 70 tokens per second on an RTX 3090: SpecPicks gives 52.1 to 55.8.
• The Qwen3 35B model “35 billion active”: it is a 35B model with 3B active parameters, measured at about 22.4 GB rather than 21 GB.
• Qwen3.6 35B A3B at 16.4 tokens per second on an M5 with 32 GB: no primary source found.
• Llama 3.3 70B at about 5 tokens per second and GPT OSS 120B at 54 from one author: the closest single source measures Llama 3.1 70B at 5.6 and gpt-oss-120b at 53.7.
• A 3B Llama at 93 and GPT OSS 20B at 77 tokens per second from one developer: not found; the nearest published figure is Llama 3.2 3B at 88.14.