Locally Explains

Your Local AI Lab Costs More Than The GPU (Three Builds)

Every figure the video says or shows, with the source it comes from. Prices were checked on 7 October 2026 unless a source states its own date. Prices are US, before tax.

The parts and their prices

• RTX 5060 Ti 16GB, $799.99. MSI Ventus 2X 16G OC Plus, sold and shipped by Newegg. https://www.newegg.com/p/N82E16814137958
• GIGABYTE AI TOP ATOM (NVIDIA GB10), $7,999.99, 128GB LPDDR5X unified memory and a 4TB PCIe 5.0 NVMe SSD, sold by Newegg. https://www.newegg.com/p/N82E16859252047
• CyberPower CP1500PFCLCD, $239.95, 1500VA / 1000W, PFC sine wave. Newegg and Best Buy list it at $239.95; CyberPower’s specification gives 1500VA / 1000W. https://www.newegg.com/cyberpower-cp1500pfclcd-nema-5-15r/p/N82E16842102134 https://www.bestbuy.com/product/cyberpower-pfc-sinewave-series-1500va-battery-back-up-system-black/JX8P9297PT
• APC Back-UPS 650VA / 360W, about $80. Best Buy, $79.99 in the script writer’s capture of 7 October 2026 (see Not checked).
• TP-Link TL-SG108 8 port gigabit switch, $19.99, sold by Newegg. https://www.newegg.com/p/N82E16833704173
• A 2TB NVMe SSD for about $250. HUADISK 2TB PCIe 4.0 (7,100MB/s) at $259.99 and its 5,000MB/s version at $229.99 on Newegg; the Crucial P310 2TB was $308.99 there the same day. https://www.newegg.com/HUADISK-2TB-NVMe-1-4/p/0D9-0118-00006
• A 4TB NVMe SSD for about $456. The cheapest name-brand 4TB NVMe on Newegg was the Team Group T-Force G50 EVO 4TB at $455.99. (The supplied script said $326 from a 2 October deal; corrected before recording.) https://www.newegg.com/team-group-4tb-t-force-nvme/p/N82E16820985406
• A network cable, about $10. Not tied to one listing.
• The $500 and $2,000 hosts are the channel’s reference build budgets, not listings. https://locallyexplains.site/best-local-ai-machine-under-500/ https://locallyexplains.site/best-local-ai-machine-2000/

The totals (arithmetic)

• Starter: $500 + $800 + $80 + $30 = $1,410. Without the host: $910.
• Serious: $2,000 + $250 + $240 + $30 = $2,520.
• No compromise: $8,000 + $240 + $30 = $8,270 (the GB10 already has its 4TB drive).

The software

• Ollama serves models to other machines on the network. https://docs.ollama.com/faq#how-can-i-expose-ollama-on-my-network
• Open WebUI connects to Ollama. https://docs.openwebui.com/getting-started/quick-start/connect-a-provider/starting-with-ollama/
• Tailscale Serve shares a local service with devices on your tailnet (“Available within your tailnet” is the line it prints). https://tailscale.com/docs/features/tailscale-serve

Model sizes

• Qwen 3.5 122B, about 70 GB quantized. Unsloth’s Q4_K_S GGUF is 71.7 GB (Ollama’s default qwen3.5:122b tag is 81 GB). https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF https://ollama.com/library/qwen3.5/tags
• Qwen 3 235B at 4 bit, 142 GB. Ollama’s qwen3:235b tag. https://ollama.com/library/qwen3/tags
• 3 × 142 = 426 GB; 426 + 3 × 70 = 636 GB.

Power

• NVIDIA ratings: RTX 5060 Ti 180 W, RTX 5070 Ti 300 W, RTX 5080 360 W (total graphics power). StorageReview: “The NVIDIA GeForce RTX 5080 has a listed peak power draw of 360W.” https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5060-family/ https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5070-family/ https://www.nvidia.com/en-au/geforce/graphics-cards/50-series/rtx-5080/
• 584 W, an RTX 5080 system under an AI workload. StorageReview: “the system power increased from 239W as the test was preparing to 584W when the GPU was put under load.” https://www.storagereview.com/review/nvidia-geforce-rtx-5080-review-the-sweet-spot-for-ai-workloads
• The tier estimates. Loaded: card limit + a 100 W host allowance (180 + 100 = 280 W; 300 + 100 = 400 W; the no compromise tier uses StorageReview’s measured 584 W). Idle: 45 W and 51 W from published GPU idle readings (ComputerBase) for the whole system; these are the script’s estimates, not our measurements. https://www.computerbase.de/artikel/grafikkarten/nvidia-geforce-rtx-5060-ti-16-gb-test.92119/seite-8 https://www.computerbase.de/artikel/grafikkarten/nvidia-geforce-rtx-5070-ti-test.91379/seite-9
• 500 W ≈ 1,700 BTU per hour. 1 W = 3.412 BTU/h; 500 × 3.412 = 1,706.

Electricity and the monthly bill

• 18.2 cents per kWh, US residential average forecast for 2026. EIA Short-Term Energy Outlook, released 6 October 2026 (table “Electricity, Coal and Renewables”). https://www.eia.gov/outlooks/steo/report/elec_coal_renew.php
• 30 day months at 18.2c/kWh:
  • 45 W all month: 45 × 720 h = 32.4 kWh = $5.90 (the script says about $5.80).
  • Starter, 20 h idle at 45 W + 4 h at 280 W: 60.6 kWh = $11.03.
  • Serious, 20 h at 51 W + 4 h at 400 W: 78.6 kWh = $14.31.
  • No compromise, 4 h at 584 W plus ~54 W idle: ~102 kWh = $18.60.
  • Eight loaded hours instead of four add about $5, $8 and $12.

Against the cloud plans

• $20 a month = $240 a year; $200 a month = $2,400 a year (ChatGPT Plus / Pro and Claude Pro / Max price tiers).
• Starter: $20 − $11 = $9 saved a month; $1,410 / $9 = 157 months ≈ 13 years; $910 / $9 = 101 months ≈ 8.5 years; an $800 GPU / $9 ≈ 89 months ≈ 7 years.
• Serious: $2,520 / ($200 − $14.30) = 13.6 months ≈ 14; the $2,000 host alone, 10.8 ≈ 11.
• GB10: $200 − $18.60 = $181.40 saved; $8,270 / 181.4 = 45.6 ≈ 46 months.

Not checked

• The APC 650VA / 360W price: Best Buy’s search page lazy-loads prices and could not be read headlessly on 7 October; $79.99 is the script writer’s capture from the same day.
• The idle wattages (45 W, 51 W, ~54 W) are the script’s estimates from published GPU idle readings, not measurements of these exact builds. The narration says the 100 W host allowance is part of the idle values; the arithmetic shows it belongs to the loaded values.
• The cable price ($10) is not tied to one listing.