# The AI Development That’s Stopping Memory Costs To Go Down

Source: https://www.youtube.com/watch?v=qFpps5Ur-qs
Recap page: https://rapidrecap.app/video/qFpps5Ur-qs
Generated: 2026-02-14T16:33:27.908+00:00

---
## Quick Overview

The high cost of generating AI videos is primarily driven by the massive computational demands of world models, which require significantly more memory and processing power than training large language models (LLMs) or traditional video compression techniques, leading to a massive surge in demand for AI memory and hardware.

**Key Points:**
- Generating one minute of 720p video at 10 FPS using current world models requires about 40 GB of VRAM, which is far more costly than text generation.
- NVIDIA's Cosmos platform is purpose-built for physical AI, featuring generative world foundation models (WFMs) that can be post-trained for specific applications like autonomous driving and robotics.
- The training cost disparity is highlighted by the fact that AI-generated videos are far more expensive to create ($$$) than world models ($), which are comparatively cheaper.
- The video generation process requires complex tokenization, including 3D Patchify, and relies on both diffusion and autoregressive transformer models for high-quality, consistent video output.
- World models, like Cosmos, focus on maximizing physical consistency (geometry, trajectory following, object permanence) rather than just perceptual realism, making them superior for real-world applications.
- NVIDIA's Cosmos Cookbook is available to help developers learn and apply these techniques, potentially leading to an economic shift where robotics offers unlimited labor and capacity, unlike the human labor market.
- The exponential growth in AI compute demand is evidenced by NVIDIA's soaring revenue post-ChatGPT launch, primarily driven by Data Center segment growth.

![Screenshot at 00:04: The graphic comparing affordable PC parts \(ship in a bottle\) constrained by the supply chain, visually representing bottlenecks in hardware availability, which directly impacts AI model training costs.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-00-04.jpg)

**Context:** The video explains the high computational cost associated with training and running advanced AI models, particularly focusing on the difference between text generation, general AI video generation, and specialized 'World Foundation Models' (WFMs) designed for physical AI tasks like robotics and autonomous driving, using NVIDIA's Cosmos ecosystem as a primary example. The video contrasts the cost and data requirements of these different AI modalities.

## Detailed Analysis

The video argues that the cost barrier for AI advancement is shifting from language models to physical AI, specifically world models, due to the immense computational expense of video generation. A short 3.6-second segment of 720p video at 10 FPS requires 128k text tokens, translating to approximately 40 GB of VRAM usage, making video generation significantly more expensive ($$$) than text generation ($). NVIDIA's Cosmos platform is presented as a solution for physical AI, designed to generate realistic, consistent simulations for robotics and autonomous vehicles via its workflow involving Cosmos Predict, Transfer, and Reason/RL modules. This approach allows for training specialized models on curated data derived from the pre-trained foundation model, enabling simulation-based learning that is far more cost-effective than real-world data collection for robotics. The comparison between the automotive industry (limited by population and road capacity) and the robotics industry (unlimited labor and capacity via simulation) suggests a major economic shift driven by these simulation capabilities. The video references key concepts like teacher forcing in training to improve model accuracy, and the efficiency gains from using highly compressed token representations rather than raw video data, although the memory consumption for video tokens remains high. Ultimately, the video concludes that while AI video generation is expensive, world models offer a more sustainable path for training embodied AI agents.

### RAM Costs & Video Generation

- 2 sticks of RAM costing $900 due to high memory demands
- DDR5-6000 2x32GB average price spike
- 128k text tokens equate to 3.6 seconds of 720p video at 10 FPS (40 GB memory usage)

### NVIDIA Cosmos Overview

- A platform for physical AI featuring generative world foundation models (WFMs)
- Includes Cosmos Predict, Cosmos Transfer, and Cosmos Reason/RL modules
- Designed specifically for real-world systems like AVs and robots

### World Model vs. Video Generation Cost

- AI generated videos ($$$) are far more expensive than World Models ($) due to computational demands and memory usage

### Training Methodology

- World models prioritize physical consistency (geometry, trajectory following, object permanence) over perceptual realism
- Uses a reinforcement learning loop (Observe -> Imagine Future -> Pick Action -> Execute -> Policy Update)
- Post-training allows specialization for different robots/sensors using custom datasets

### Robotics vs. Automotive Economics

- Robotics industry offers unlimited labor and capacity via simulation, unlike the human labor market limited by population and road capacity
- Robots (approx. $20k one-time cost) are financially superior to human workers (approx. $64k annual cost)

### Cosmos Architecture & Training

- Utilizes a diffusion transformer architecture processing corrupted tokens from encoded video input, guided by text prompts
- Post-training involves fine-tuning the pre-trained WFM using custom data and RL frameworks (like Cosmos RL) to adapt to specific tasks and embodiments.

![Screenshot at 00:04: Graphic illustrating that affordable PC parts are constrained within a bottle labeled 'Supply chain,' symbolizing hardware bottlenecks.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-00-04.jpg)
![Screenshot at 00:11: Chart showing the dramatic price increase of DDR5-6000 2x32GB RAM over 18 months, reaching nearly $1000.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-00-11.jpg)
![Screenshot at 00:20: News clipping showing OpenAI and NVIDIA partnership to deploy 10 Gigawatts of NVIDIA systems, highlighting massive infrastructure investment for AI.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-00-20.jpg)
![Screenshot at 01:22: Table comparing different visual tokenizers, showing Cosmos-Tokenize1 achieving high marks across Causal, Image, Video, Joint, Discrete, and Continuous capabilities.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-01-22.jpg)
![Screenshot at 04:44: Image of scales balancing 'Cons' against 'Pros,' with 'Pros' heavily outweighing 'Cons' in the context of robotics vs. human labor costs.](https://ss.rapidrecap.app/screens/qFpps5Ur-qs/00-04-44.jpg)