China’s DeepSeek trained its open source R1 AI model for just $294K, shaking Nvidia, challenging Western giants, and proving efficiency can outpace raw spending.
China’s DeepSeek has stunned the AI world by training its new open source, reasoning-centric R1 model for only $294,000, a fraction of the hundreds of millions spent by OpenAI and Google. The breakthrough, published this week in a peer-reviewed Nature paper, positions open-source efficiency as a major disruptor in the global AI race.
R1 was trained on 512 Nvidia H800 GPUs, chips approved for export to China under US restrictions. DeepSeek maximised performance using distillation techniques, where the model learns from outputs of existing AIs, while also adopting methods from open models such as Meta’s Llama. This efficiency-first approach not only cut costs but also pushed boundaries on the ethics of distilling knowledge from other systems.
At its core, R1 is open source, designed for accessibility and scrutiny, amplifying debates over intellectual property while enabling global adoption. By delivering results comparable to costlier Western LLMs, the model signals that AI progress is no longer about brute-force compute and budgets, but about ingenuity, efficiency, and openness.
The market impact was immediate: Nvidia lost hundreds of billions in value after the announcement, highlighting how low-cost, open AI can shift competitive dynamics overnight.
The geopolitical backdrop adds weight to the achievement. With China banning domestic firms from Nvidia’s most advanced hardware, local champions like Huawei (Ascend chips) and Cambricon (record profits) are filling the gap, moving the country closer to chip self-sufficiency.
DeepSeek’s founder Liang Wenfeng has kept a low profile, but the disclosure is widely seen as a strategic signal: open-source innovation, coupled with efficiency, could redefine the global balance of AI power.
