DeepSeek or DeepFake?
The AI Arms Race and the Open-Source Dilemma
In the high-stakes world of artificial intelligence, geopolitical competition between the United States and China has intensified. The debate over whether AI development should be open or closed is now central to broader discussions about national security, innovation, and economic dominance. The U.S. has seen increasing calls for export controls and restrictions on AI-related technologies, aiming to maintain its edge while limiting China’s access to cutting-edge models and semiconductor chips. Meanwhile, China has pushed forward with its own AI ambitions, exemplified by the recent release of DeepSeek-V3 and DeepSeek-R1, models that have not only made waves in the research community but also sent ripples through the stock market. As AI models grow more powerful, the global AI arms race continues to escalate, and with it, the debate over intellectual property rights, open-source AI, and the ethical boundaries of model training.
DeepSeek’s latest AI models have attracted considerable attention, both for their technical achievements and the suspicions surrounding their development. Announced with much fanfare, DeepSeek-V3 boasts high performance at a fraction of the cost typically associated with training large-scale AI models. Technical reports claim that DeepSeek-V3 was trained on a mere $5.6 million budget—a strikingly low figure given the scale of its capabilities. The model reportedly achieves efficiency through a combination of optimized training techniques and a well-structured token processing pipeline, which allows it to perform at levels comparable to models trained on significantly larger budgets. This increase in efficiency is not only an impressive engineering feat but also a step towards more environmentally sustainable AI, as reduced training costs correlate with lower energy consumption and carbon emissions. While this efficiency has led to excitement in some circles, it has also raised skepticism. The stock market reaction to DeepSeek’s announcement saw shares of AI-related firms fluctuate, including NVIDIA, whose GPUs are at the heart of most modern AI training processes. While NVIDIA’s stock experienced temporary volatility due to speculation on market shifts, demand for their chips continues to significantly outstrip supply. This is especially true given that NVIDIA’s GPUs are not only essential for training AI models but also play a crucial role at inference time. As Jevons Paradox teaches us, when a technology becomes more efficient in using a resource, it can paradoxically lead to increased overall consumption of that resource.
However, beneath the surface, a deeper controversy is brewing.
A first debate focused on the reported $5.6 million training cost. Some experts pointed out that this figure only accounts for direct GPU rental costs and fails to consider essential expenditures such as infrastructure, data acquisition, personnel salaries, and the cost of training earlier iterations of the model. The total investment behind DeepSeek’s AI efforts could be much higher—potentially in the hundreds of millions. The second debate was even more contentious, raising allegations that DeepSeek may have used OpenAI’s proprietary technology in ways that violate terms of service. Specifically, some speculate that DeepSeek could have leveraged OpenAI’s APIs to exfiltrate training data, which was then used to fine-tune its models via distillation. This claim, if true, would mean DeepSeek is benefiting from OpenAI’s work without permission and in potential violation of intellectual property protections. While there is no hard evidence proving this yet, the lack of clarity around DeepSeek’s training process has only fueled speculation. What goes around comes around, if you think that OpenAI has built their business consuming content generated by others, many times under copyright protection.
At the heart of this controversy is a broader debate about the role of open-source AI. Distillation, as a machine learning technique, is neither new nor controversial. It is a well-established approach used to make models more efficient while preserving their capabilities. The issue is not with distillation itself but with whether DeepSeek trained its models using unauthorized access to proprietary technology. Had DeepSeek employed the same distillation techniques using an open-source model, such as Meta’s LLaMA, there would likely be far less scrutiny. This raises a crucial question: Should AI models be open for the sake of global scientific progress, or does open-source AI present a national security risk? The reality is that it is both. Open-source AI fosters collaboration and accelerates innovation, allowing a broader community of researchers to contribute to advancements. At the same time, the unrestricted availability of powerful models could enable adversarial actors to weaponize AI in ways that pose significant security threats. Striking the right balance between these competing priorities is a challenge that policymakers, researchers, and industry leaders must navigate carefully.
For now, the full story of DeepSeek remains unresolved. We still do not have definitive answers about whether its training methods adhered to ethical and legal standards or what the true cost of its development was. As the AI race between the U.S. and China heats up, caution is necessary to avoid overreactions driven by geopolitical tensions. It is easy to fall into narratives that either overhype or prematurely dismiss new developments. Instead, it is more important than ever to approach these issues with critical thinking, independent research, and a clear understanding of the commercial and national interests at play. In an era where AI is as much a tool for economic and military power as it is for technological advancement, recognizing the nuances of these discussions is essential.
The truth about DeepSeek, and the broader implications of its development, will become clearer in time. Until then, we should remain vigilant, skeptical, and open to informed debate.