Generative Artificial Intelligence has advanced astonishingly, and DeepSeek, a Chinese AI startup, has just shaken the industry. With its ground-breaking DeepSeek-R1 model, the company has upended traditional assumptions about AI development, challenging OpenAI, Anthropic, and Google. However, as with any significant innovation, opportunities and risks exist.
In this article, we will explore DeepSeek’s achievements, hidden dangers, and what it means for the future of generative AI. Towards the end, we will also examine other recent LLM models.
How is DeepSeek reshaping AI?
DeepSeek R1 is a state-of-the-art reasoning model similar to OpenAI’s O1 and O3 models. At least in certain areas, it is as good as OpenAI’s O1 model. The Chinese company developed it in just a couple of months, with much less hardware and a budget far below that of other leading models.
1. A New Standard for AI Reasoning at a Fraction of the Cost
DeepSeek has achieved something that many thought was impossible: building an advanced reasoning model at just $6 million, compared to the hundreds of millions spent by OpenAI and others. This is a watershed moment in AI—showing that innovation isn’t just about brute-force computing power but also about efficiency and novel approaches.
2. Open-Weight Model: AI for Everyone?
DeepSeek has chosen an “open-weight” approach, meaning developers and researchers can access and experiment with its trained model parameters. While this isn’t complete open-source freedom, it represents a significant step toward making cutting-edge AI models available outside corporate control. DeepSeek-R1 is now available as an NVIDIA NIM microservice preview to help developers build specialised AI agents. Microsoft has integrated DeepSeek-R1 into Azure AI Foundry, and Amazon has made it available through AWS, with CEO Andy Jassy encouraging users to explore its capabilities. Finally, Huggin Face has made all models from DeepSeek available on their site.
3. A Disruptor to the AI Business Model
With DeepSeek’s AI model pricing at a fraction of OpenAI’s costs, this could force a dramatic shift in AI pricing. Currently, OpenAI charges around $15 per million tokens for input processing, whereas DeepSeek offers similar capabilities at just $0.55 per million tokens. If companies move toward cheaper alternatives like DeepSeek, this could shake the dominance of OpenAI and Anthropic in the AI landscape.
4. Hardware Efficiency Without Relying on Nvidia’s top GPUs
DeepSeek’s models are also built to be efficient on lower-end hardware, meaning AI development isn’t entirely dependent on Nvidia’s high-end GPUs such as H-100s. Given the U.S. sanctions limiting China’s access to Nvidia’s best chips, DeepSeek’s ability to build competitive models using only older GPUs, such as H-800s without them, is a significant achievement—and a challenge to the Western AI giants. They have proven that it is possible to create advanced AI models without the extreme budgets Western AI companies have had.
The Negatives: Censorship, Surveillance, and Security Risks
1. Built-in Chinese Censorship
Despite its innovation, DeepSeek’s AI is subject to government-mandated censorship. If you ask about Tiananmen Square, Taiwan, Hong Kong protests, or other politically sensitive topics, the model will either refuse to respond or provide vague answers. Similar concerns about Western AI models have been raised primarily regarding political correctness, though the providers have been more willing to correct the biases. At least in some cases, these biases have originated in the training material. This raises concerns about the ethical limitations of AI models built under authoritarian regimes.
2. Deep Data Collection and Privacy Concerns
DeepSeek’s web-based AI extensively tracks user data, including IP addresses, device details, and keystroke patterns. This data is stored in China, where the government has legal access to corporate records. Security experts warn that using DeepSeek’s hosted services could expose sensitive conversations and business data to potential government surveillance.
3. Security Vulnerabilities and Potential Exploits
Recently, cybersecurity researchers at Cisco found that DeepSeek failed all their tests. When tested with Qualys TotalAI’s state-of-the-art Jailbreak and Knowledge Base (KB) attacks, it failed about half their tests. While at least some of these were reportedly taken offline since the tests were published, it raises serious concerns:
Can DeepSeek’s security infrastructure be trusted?
What happens if an AI model trained on sensitive inputs gets breached? Could corporate users accidentally leak trade secrets into an unsecured AI model? These risks make self-hosting a safer option, but that requires expensive infrastructure and technical expertise.
Several governments have taken decisive action in response to these concerns. Australia has banned DeepSeek from all government devices, citing national security risks. Similarly, the United States is considering legislation prohibiting the app on government devices due to fears of data transmission to the Chinese government. These measures underscore the apprehension surrounding DeepSeek’s data practices and potential influence operations.
What’s Next for Generative AI?
1. The AI Arms Race Just Got More Intense
DeepSeek’s rise forces OpenAI, Google, and Meta to rethink their AI strategies. AI could become increasingly commoditised with a competent low-cost alternative, shifting the industry from exclusive, high-cost models to affordable, widespread access.
2. China’s Increasing AI Dominance
While the West has led AI research for years, DeepSeek’s breakthrough demonstrates that China is catching up fast. Its efficiency-driven approach could give Chinese AI firms a significant edge, particularly in industries that don’t need the absolute best AI but simply cost-effective, good-enough performance.
3. The Shifting AI Landscape: China’s Rise and Open-Source Disruption
Just when it seemed that DeepSeek had rewritten the rules of AI, another Chinese model, Kimi 1.5, entered the scene. Developed by MoonShot AI, Kimi 1.5 also claims to outperform OpenAI’s models.
Kimi 1.5: A New Challenger in AI
Kimi 1.5 is a multi-modal model that can process and reason across text, vision, and coding tasks. Early benchmarks suggest that it might surpass DeepSeek in some areas, further intensifying the competition in AI development.
This raises an important question: Could we enter an era where China leads AI innovation? Western AI firms may struggle to maintain their edge if Chinese companies continue pushing boundaries at competitive prices.
The Western AI Response: Strengths and Weaknesses
While this scenario is possible, DeepSeek’s open-source nature may allow Western generative AI companies to adapt quickly. With access to more powerful GPUs, they could enhance their models using the new ideas emerging from Chinese research. However, the abundance of money and computational power has arguably made Western AI firms complacent, leaving them unprepared for this sudden surge of competition.
Meanwhile, export restrictions have forced Chinese companies to find alternative ways to innovate, leading to creative breakthroughs. Ironically, DeepSeek’s open-source nature could help Western firms regain momentum by reducing their reliance on expensive GPUs or by integrating these new approaches into models that benefit from superior hardware capabilities.
More Disruptors on the Horizon
As if DeepSeek and Kimi 1.5 weren’t enough, Stanford and University of Washington researchers have introduced an open-source alternative to OpenAI’s O1 reasoning model. Built on Google’s Gemini 2.0 Flash Thinking Experimental, this model offers advanced reasoning capabilities at a fraction of the cost—under $50—making high-level AI reasoning more accessible.
The Battle is Far From Over
While these developments mark significant shifts, ruling out OpenAI or Google would be premature. As I explored OpenAI’s latest O3 models while writing this, it became evident that the competition remains fierce. The AI race is far from settled, and the coming months may bring even more surprises.
Final Thoughts: Innovation vs. Ethics vs. Power
DeepSeek’s AI revolution presents a complicated reality:
✔ It is technologically impressive
✔ It is cost-efficient and widely available
✖ But it comes with built-in censorship
✖ Serious data privacy concerns (Not a problem when using the models via AWS, Amazon, Hugging Face or when downloading and running them yourself).
Organisations and individuals considering adopting DeepSeek’s technology must conduct thorough due diligence. Understanding the inherent risks related to data privacy, security vulnerabilities, and content censorship is crucial. Implementing robust security measures and staying informed about regulatory developments can help mitigate potential risks using DeepSeek’s AI models.
For businesses and developers, the key question is how to balance access to cutting-edge AI while protecting intellectual property, security, and ethical standards.
One thing is certain: the Gen-AI revolution is far from over—and the next breakthrough might not come from Silicon Valley but from unexpected challengers like DeepSeek, Kimi 1.5 or whatever may come next.
Please get in touch with us today to discuss how we can help you use AI to give your company an edge, contact us today!