•    DeepSeek’s R1 and V3 Large language Models (LLMs) are open-source and substantially cheaper to use than ChatGPT, with API costs up to 96% lower than ChatGPT’s o1.

•    DeepSeek’s compute-efficient AI models rival the best models on performance, challenging the long-standing AI industry notion that achieving cutting-edge large-scale AI models requires enormous financial investments and brute-force compute power.

•    Although it is still early to predict DeepSeek’s future success, as its models are under intense scrutiny from existing players, investors are now awakening to the possibility of efficient use of resources to enhance performance.

•    It may also prompt AI players to reconsider the current scaling laws and the assumption that significant compute power is essential for training advanced models.

•    The potential reduced demand for compute resources led to an immediate stock market reaction that saw across-the-board selling of stocks related to data centers from chip to power companies.

•    DeepSeek's innovations may be limiting the growth of some chipset players, NVIDIA in particular. However, it is too early to quantify the impact, given that the cost reductions may boost the development of higher quality and costlier LLMs, and many companies are aiming for AGI and Super-Intelligence. Moreover, most players have reiterated their intention of continuing with their current CAPEX plans.

•    A big trend playing out is the closing of the gap between open-source and closed-source AI. DeepSeek and Meta with open-source AI are giving some tough competition to proprietary models such as OpenAI. The rise of open-source LLMs performing at a similar level to proprietary counterparts represents a transformative shift in the AI landscape.

•    As a Chinese company, DeepSeek has added a geopolitical angle to the debate. The timing of key model releases with US announcements of the $500 billion "Stargate Project" is likely not just a coincidence.

•    There are question marks around DeepSeek’s real cost of model training – likely the $6 million cost is not the full picture. However, DeepSeek’s achievements in training efficiency, and the low API cost, are both clear.

Access full report here.
If you are not a client, reach out to us at info@counterpointresearch.com