Every figure, price, capacity and benchmark this video puts on screen, chased to a primary source. Checked 6 September 2026.
RunPod publishes per hour Pod pricing on its own pricing page, split between Community Cloud and Secure Cloud.
| GPU | VRAM | Community Cloud | Secure Cloud |
|---|---|---|---|
| RTX 4090 | 24 GB | $0.34 / hr | $0.74 / hr |
| RTX 5090 | 32 GB | $0.69 / hr | $0.99 / hr |
Source: RunPod pricing, https://www.runpod.io/pricing
The three monthly figures the video draws are arithmetic on those rates, at a full 730 hour month (365 days / 12 months x 24 hours = 730.5 hours):
| Hourly | Per month at 730 hours |
|---|---|
| $0.99 | $722.70 |
| $0.74 | $540.20 |
| $0.70 | $511.00 |
| $0.34 | $248.20 |
The narration rounds these to “roughly seven hundred thirty”, “roughly five hundred” and “about two hundred fifty”. A dollar an hour is exactly $730.50 a month, which is where the first figure comes from. Storage and idle volume charges are additional and are not in these numbers.
| Card | Architecture | Memory | Bus |
|---|---|---|---|
| GeForce RTX 3090 | Ampere | 24 GB GDDR6X | 384 bit |
| GeForce RTX 4090 | Ada Lovelace | 24 GB GDDR6X | 384 bit |
| GeForce RTX 5090 | Blackwell | 32 GB GDDR7 | 512 bit |
Sources:
The 4090 did not increase memory capacity over the 3090; both are 24 GB. The 5090 is the first of the three to move, to 32 GB.
The llama.cpp project keeps a community CUDA performance thread. The RTX 5090 entry for Llama 2 7B Q4_0, text generation of 128 tokens:
| Configuration | Tokens per second |
|---|---|
| Without flash attention | 290.02 +/- 1.10 |
| With flash attention | 300.40 +/- 0.28 |
Source: llama.cpp CUDA performance discussion, https://github.com/ggml-org/llama.cpp/discussions/15013
That is a small four bit model on a single card, which is what makes it fast; it is not a figure that carries over to a larger model or a longer context.
Mac Studio with M3 Ultra. Configurable up to 512 GB of unified memory, with 800 GB/s memory bandwidth, up to a 32 core CPU and an 80 core GPU. Source: Apple Newsroom, “Apple reveals M3 Ultra, taking Apple silicon to a new extreme”, https://www.apple.com/newsroom/2025/03/apple-reveals-m3-ultra-taking-apple-silicon-to-a-new-extreme/ and Mac Studio (2025) tech specs, https://support.apple.com/en-us/122211
NVIDIA DGX Spark. 128 GB of unified LPDDR5x at 273 GB/s on a GB10 Grace Blackwell Superchip. NVIDIA states it can fine tune models up to 70 billion parameters and run inference on models up to 200 billion parameters, with two units connected reaching 405 billion. Source: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
Qwen3 Coder. Qwen3-Coder-480B-A35B-Instruct is a 480 billion parameter mixture of experts model with 35 billion active parameters, 160 experts with 8 activated per forward pass, and a native context window of 262,144 tokens, extendable to about 1 million with YaRN. Sources: https://qwenlm.github.io/blog/qwen3-coder/ and https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct
Kimi K2. A mixture of experts model with 1 trillion total parameters and 32 billion activated, pre trained on 15.5 trillion tokens, released under a modified MIT licence and built for tool use and agentic work. It reports 65.8 on SWE-bench Verified. Sources: https://moonshotai.github.io/Kimi-K2/ and https://huggingface.co/moonshotai/Kimi-K2-Instruct
DeepSeek V3 and V3.1. DeepSeek-V3.1 is 671 billion total parameters with 37 billion active, extended through a two phase long context training process to a 128,000 token context window, with post training aimed at tool calling and agent behaviour. Weights are published on Hugging Face. Sources: https://api-docs.deepseek.com/news/news250821/ and https://huggingface.co/deepseek-ai/DeepSeek-V3.1
Ollama, LM Studio, vLLM, llama.cpp, Open WebUI, GGUF and MLX are all named in the narration as pieces a first local setup has to choose between. GGUF is llama.cpp’s model container format; MLX is Apple’s array framework for Apple silicon.