The RTX 3090 Trap That Costs Local AI Buyers $1,600
Every figure the video puts on screen, with its primary source. Prices are US prices checked on 9 October 2026.
Prices
| figure |
what |
source |
| about $1,600 |
Used RTX 3090 24GB, going rate (median of 84 screened eBay asking prices, range $1,550 to $1,799; $1,623 on 9 Oct 2026). Asking prices, not completed sales. |
RigPrice RTX 3090 |
| about $3,200 |
Two used RTX 3090s at about $1,600 each (1,600 × 2) |
derived |
| $8,422 |
MSI RTX 5090 32G Ventus 3X OC on Newegg ($8,419.00 main offer and $8,420.00 from a second seller on 9 Oct 2026; $8,422 on 7 Oct) |
Newegg listing |
| $3,499.99 |
GMKtec EVO-X2, Ryzen AI Max+ 395 (Strix Halo), 128GB RAM + 1TB SSD, US |
GMKtec EVO-X2 |
| $5,499 |
Mac Studio with M5 Ultra (30-core CPU, 64-core GPU), 96GB unified memory, 1TB, the 96GB starting configuration |
Apple Store, Mac Studio |
Our calculations
| figure |
formula |
| about $38 per token per second, used RTX 3090 |
$1,600 ÷ 41.93 tok/s = $38.16 |
| about $111 per token per second, RTX 5090 listing |
$8,422 ÷ 75.74 tok/s = $111.20 |
| about $92 a year, RTX 3090 |
0.350 kW × 4 h × 365 days × $0.18 per kWh = $91.98 |
| about $151 a year, RTX 5090 |
0.575 kW × 4 h × 365 days × $0.18 per kWh = $151.11 |
| about 1.8 times faster |
75.74 ÷ 41.93 = 1.81 |
The power estimate uses each card’s rated board power at full load for four hours a day; real inference draw varies, and the rest of the system adds to it.
Memory capacity
- RTX 3090: 24GB GDDR6X. RTX 5090: 32GB GDDR7. NVIDIA RTX 3090, NVIDIA RTX 5090
- Mac Studio: up to 512GB unified memory with M5 Ultra; M5 Max up to 128GB. Apple Mac Studio specs
- DGX Spark: 128GB of coherent unified system memory. NVIDIA developer blog
- Ryzen AI Max+ 395 (Strix Halo): up to 128GB unified LPDDR5X. AMD
- A 70 billion parameter model at about 4 bits per weight needs roughly 35 to 40GB for the weights alone (70e9 × ~4 to 4.5 bits ÷ 8), before the context cache and the runtime.
Memory bandwidth
Published speed results
- Qwen 3.8 27B Q4_K_M, llama.cpp llama-bench, 512 token prompt, 128 generated tokens, full GPU offload, three runs after a warmup: RTX 3090 41.93 tok/s, RTX 5090 75.74 tok/s. GPU Battle (updated October 2026)
- DGX Spark: Qwen3 235B in NVFP4 on TRT-LLM, two DGX Spark systems linked through ConnectX-7, input 2,048 tokens, output 128, batch 1: 11.73 tok/s decode. NVIDIA developer blog
- Strix Halo, GPT-OSS 120B MXFP4, llama.cpp tg128: 51.0 to 56.6 tok/s across ROCm and Vulkan builds in the AMD Strix Halo Toolboxes project’s results. The video quotes 53.7 to 53.8 tok/s from published results, which sits inside that range. kyuz0 AMD Strix Halo Toolboxes
- Prompt processing (reading the input) and token generation (writing the reply) are reported separately by llama-bench (pp and tg); prompt rates are many times the generation rates on the same hardware.
Power
- RTX 5090: 575 W total graphics power; 1000 W required system power. NVIDIA RTX 5090
- RTX 3090: 350 W graphics card power. NVIDIA RTX 3090
- Ryzen AI Max+ 395: configurable TDP of 45 to 120 W; mini PCs built on it run the chip above 100 W under sustained load. AMD
Software backends
- llama.cpp documents CUDA for NVIDIA GPUs, HIP for AMD GPUs, SYCL for Intel GPUs, Metal for Apple Silicon, plus Vulkan and others. llama.cpp README, build docs
- OpenVINO is Intel’s inference toolkit for Intel CPUs, GPUs and NPUs. OpenVINO
- MLX is Apple’s array framework designed for Apple Silicon and its unified memory. MLX
Cloud rental
Not checked
- The exact 53.7 to 53.8 tok/s Strix Halo figure for GPT-OSS 120B MXFP4 was not located; the closest primary results (kyuz0 toolboxes, tg128) span 51.0 to 56.6 tok/s by backend.
- “Over one hundred watts in published testing” for Strix Halo rests on reviews of specific mini PCs, not on one cited test.
- Mac Studio M5 Ultra is listed by Apple as “coming late October”; the $5,499 price is Apple’s listed price, not a shipping unit.
- Resale prices next year are not knowable; RigPrice reports asks, not completed sales.