Every figure quoted in this video, where it came from, and what kind of number it is.
Three kinds appear here and they are not equally solid. A listing price is one shop on one day and moves without notice. A published spec is what the maker states and is stable. A vendor claim is a benchmark the vendor ran on hardware it chose, and is the weakest of the three even when it is honest. Each entry below says which it is.
Prices are US dollars, exclude tax, and were read on 13 September 2026. The model and driver pages were re-checked on 15 September 2026.
Intel Arc Pro B60, 24 GB — about $650. Listing price, newegg.com, 13 September 2026.
Nvidia RTX 5060 Ti, 16 GB — about $780. Listing price, newegg.com, 13 September 2026. This is one MSI listing rather than the lowest price anywhere, which is why the video says so on screen and in the narration. The $130 gap between the two cards is arithmetic on those two listings and moves whenever either of them does.
AMD Radeon RX 9060 XT, 16 GB — about $520. Listing price, newegg.com, ASRock Challenger RX 9060 XT 16 GB, 13 September 2026.
Used Nvidia RTX 3090, 24 GB. No price is quoted for this one on purpose. The used market has no list price, and the video’s point about it is about return terms rather than a figure.
RTX 3090: 350 W board power, 750 W system supply recommended. Published spec, nvidia.com, RTX 3090 reference specification. Board partner cards differ from the reference design, which is why the video marks this one “versions differ” rather than stating it flat.
Intel Arc Pro B60: 200 W. Published spec, Intel, for the card in this comparison.
Qwen3 14B, four bit — 9.3 GB. Read from the model’s own page,
ollama.com/library/qwen3, which lists qwen3:14b at
9.3 GB. Shown on screen as a capture of that page.
The same model as GGUF files. Read from
huggingface.co/Qwen/Qwen3-14B-GGUF, shown on
screen as a capture of the file listing: Q4_K_M 9 GB, Q5_0 10.3 GB, Q5_K_M 10.5 GB,
Q6_K 12.1 GB, Q8_0 15.7 GB. The four bit build is the smallest one published there and
is still larger than 8 GB, which is the claim the video makes with it.
Qwen3 Coder 30B — about 19 GB. Read from
ollama.com/library/qwen3-coder, which lists
qwen3-coder:30b at 19 GB. Shown on screen as a capture of that page.
The KV cache growing with the conversation is described qualitatively in the video and no figure is put on it, because how fast it grows depends on the model, the context length and the runtime.
CUDA is Nvidia only. This is what CUDA is, not a measurement.
ROCm on AMD: a long list on Linux, a much shorter one on Windows. Read from docs.ollama.com/gpu, shown on screen as a capture of that page. The Linux section lists Radeon RX, Radeon AI PRO, Radeon PRO, Ryzen AI and Instinct families; the Windows section lists two families. The video does not put a count on either list, because the right comparison is which families are covered rather than how many model numbers are printed.
Intel via llama.cpp. The video says a supported language model can run on an Intel card without CUDA, and that this does not make it a drop in replacement for every program written for CUDA. Support depends on the specific card and build, which is why the video says to confirm the path for your own card rather than giving a number.
M6 Mac mini, 16 GB shared memory — from $899. Announced price, apple.com. This is a preorder figure before tax, for a machine that had not shipped when the video was made, which is why the video treats waiting for it as a condition rather than a recommendation.
Older M4 Mac mini, 24 GB — $849 listing, sold out. Listing price and availability, apple.com, 13 September 2026.
Up to 4.8× faster prompt processing than the M4 mini. Vendor claim, apple.com/mac-mini, Apple’s own testing on configurations Apple chose. The video labels it on screen with the conditions attached to it (8K token prompt, 14B model, four bit, Apple tested) and says explicitly that it is a prompt processing figure and not a promise about how fast the whole answer is written.
These are not sourced because they are subtraction, and they are shown being done:
The $1000 budget is the video’s own premise rather than a figure from anywhere.
Prices move. Everything above was read on one day from one shop, and a video is watched for months. Check the current listing before spending anything, and check the specification for the exact board partner card you are looking at rather than the reference design.