The Intelligence Era · Industry · 2025
DeepSeek R1
A Chinese AI lab trained a frontier-grade reasoning model for roughly $6 million, shattering the assumption that world-class AI required hundreds of millions in compute.
DeepSeek R1 was released by DeepSeek, a research lab affiliated with the Chinese quantitative hedge fund High-Flyer, in January 2025. The model was designed specifically for complex reasoning tasks — mathematics, coding, and logical inference — and was trained using a reinforcement learning pipeline that rewarded verifiable correct answers rather than relying exclusively on supervised fine-tuning from human-labeled data. Its release was accompanied by a detailed technical report and, crucially, open weights, making it immediately accessible to researchers and practitioners worldwide.
DeepSeek R1 fundamentally disrupted the prevailing economic narrative of the AI industry. The dominant assumption entering 2025 was that the gap between frontier AI labs and all other actors would widen indefinitely because only a handful of organizations could afford the compute required to train competitive models. R1 demonstrated that algorithmic innovation — specifically in reinforcement learning efficiency and sparse activation architectures — could substitute for raw spending in ways that most analysts had not anticipated. This shifted the conversation from 'who can buy the most GPUs' toward 'who can engineer the most efficient training pipelines.'