Top Data Insights and Gradient Updates of 2025 (epoch.ai)

In 2025, a comprehensive review of AI trends culminated in the release of 70 short-form investigations, highlighting the most impactful findings on their website. Among these, the significant drop in LLM inference prices—over 10x since early 2023—underscores a competitive market landscape, making AI more accessible and affordable for consumers. Additionally, the availability of frontier AI performance on personal hardware demonstrates that cutting-edge models will soon be widely accessible, challenging companies to innovate continuously and complicating regulatory frameworks. Other insights revealed OpenAI’s heavy investment in experimental compute resources, which signals the growing capital intensity of AI development, while NVIDIA's doubling of AI compute power every ten months emphasizes the industry's escalating demand for advanced computational resources. Innovations in the Transformer architecture, such as those from DeepSeek, are paving the way for more efficient AI models, potentially reducing development costs significantly. Lastly, the idea of a large-scale national AI project akin to the Manhattan Project raises questions about future investments in AGI, positing that the value of AI will increasingly stem from broad automation rather than exclusively from R&D advancements.