The DeepSeek-R1 model, a reasoning model, further intensifies the competition by using advanced techniques like generating its own training data. Despite the shock from DeepSeek's advancements, investments in AI infrastructure continue, with projects like Stargate in Texas. Bernstein analysts caution against overreacting, noting that DeepSeek's $5 million cost figure excludes prior research expenses. The competition for AI model supremacy remains fierce, but demand for computing power is expected to rise, driven by Jevon's paradox, where increased efficiency leads to higher demand.
Key takeaways:
- China's DeepSeek model is significantly cheaper and more efficient than US AI models, raising questions about US investments in AI infrastructure.
- DeepSeek-V3 uses a "mixture of experts" architecture, making it smaller and easier to run while maintaining impressive performance.
- DeepSeek-R1, a reasoning model, competes with OpenAI's latest models and uses innovative techniques like generating its own training data.
- Despite DeepSeek's achievements, large-scale investments in data centers continue, and demand for computing power is expected to rise due to Jevon's paradox.