Does Deepseek mark a historic turning point or simply a logical evolution?
I probably don’t need to remind you of the stir caused by Deepseek’s arrival. $600 billion evaporated in a single session, Nvidia dropping to third place in market capitalization, dragging Broadcom, Cisco, and Oracle down with it...
Beyond the stock market aspects, Deepseek’s arrival seems to mark the end of the arms race in the field of LLMs. While major players like OpenAI and Google have surpassed 1 trillion parameters, Deepseek has been released with 30% fewer parameters, a 95% reduction in training time, and a 90% decrease in power consumption—according to them.
Although impressive, these figures remain quite abstract. So, is Deepseek a turning point, a disruption, or even a challenge to the historical players in the industry? Or is it simply a logical evolution of the market?
Deepseek stands out primarily due to its energy efficiency and resource optimization. While OpenAI and Google have focused on ever-larger models that are more demanding in terms of computing power, Deepseek has adopted a more pragmatic approach. It is designed to be more economical while maintaining performance levels on par with its competitors.
To achieve these results, Deepseek relies on three main factors:
Deepseek uses a mixture of experts architecture, which differs from what has recently been offered to the public. Instead of relying on a single generalist network, it runs multiple specialized neural networks in parallel, activating only a portion of its parameters per query. Therefore, instead of activating all 671 billion parameters each time, Deepseek only activates 37 billion. This not only reduces consumption but also makes the model more scalable. However, it’s important to note that this is not an innovation per se; OpenAI and Google already use this architecture for some of their models.
Another key feature that has garnered attention is Deepseek’s training process. The public mainly remembers that Deepseek used ChatGPT for training. This technique, known as knowledge distillation, involves leveraging the correct labeling and probability distributions from ChatGPT's initial training by comparing its responses to Chinese text. But this is not the only distinctive feature of Deepseek’s training process.
After the initial training, Deepseek underwent fine-tuning using a Chain of Thought (a multi-step reasoning method), validated by a reinforcement learning system, using rewards and penalties. Again, this is not groundbreaking in itself, but the innovation lies in the clever combination of established techniques.
Lastly, Deepseek reduced its power requirements by minimizing the number of decimal places in its calculations. For those wondering why a company like Nvidia, which initially made products for gamers, became the main supplier for AI companies, it’s all about precision. CPUs are specialized in whole number calculations, while GPUs, designed for game graphics, use floating-point numbers, making them more efficient for such tasks. Deepseek’s reduction in decimal precision allows it to use older, less powerful, and supposedly more energy-efficient hardware. Additionally, Deepseek developed its own language to access chip resources (instead of CUDA) for greater efficiency. This allows it to bypass chips like the A100 and H100 used by OpenAI, opting for the more economical H800 chips. This is one of Deepseek's major innovations: achieving better results with fewer resources, rather than blindly following Moore’s Law.
This efficient approach is a response to American restrictions and the evolving geopolitical landscape. For several years, the United States has imposed strict export controls on essential components, particularly the advanced chips needed for AI training. These sanctions aimed at limiting China’s technological capabilities, but they have had the opposite effect—forcing China to innovate differently.
Instead of relying on brute force, which has become increasingly inaccessible, the Chinese company optimized the use of available resources. These hardware constraints led to a focus on efficiency, a factor that Western companies had largely neglected until now.
It is ironic that these economic sanctions had a more immediate effect on AI energy efficiency than decades of discussions about climate change. While environmental regulations have struggled to impose limits on big tech, economic and geopolitical constraints have proven to be powerful drivers of innovation.
With Deepseek, we are likely witnessing the end of the arms race in generative AI. For two years, the dominant trend has been escalation: more computing power, more data. Today, however, we are witnessing a new dynamic, where efficiency and optimization take precedence over raw power. And I believe this approach is relevant for tackling the challenges we face today.
But does this mean we are witnessing the end of generative AI? If raw computing power is no longer a differentiating factor, how will LLMs evolve?
If raw computing power is no longer the primary differentiator, data sources are becoming the new frontier. Deepseek reveals that generative AI is not only defined by its technical capabilities, but also by the information it relies on.
For example, Deepseek is clearly aligned with the Chinese government’s political stance. Topics such as Tiananmen Square or Taiwan are treated with a strong ideological bias. Similarly, Elon Musk’s AI, Grok, reflects Twitter's perspectives. Its responses are influenced by the data on which it has been trained, leading to specific biases.
By making generative AI more economically and technically accessible, Deepseek opens the door to the proliferation of specialized AI models based on specific data corpora. This meets a clear need, as shown by the many articles and proofs of concept on Retrieval Augmented Generation (RAG).
Deepseek makes this possible for a wide range of organizations thanks to its open-source, locally hosted version, which offers:
Thus, Deepseek enables the emergence of viable business models for generative AI. Unfortunately, the advertising model seems to be taking precedence.
Perplexity AI pioneered this by integrating sponsored links into its responses, compensating content publishers who host the generative AI. This model will soon be coming to France with a partnership signed with Numérama. This approach mirrors Google's advertising model, even though Google itself was not a pioneer in this segment, which is a surprising paradox for a company that defined the online advertising model.
These alliances between platforms and AI signal a market consolidation around proprietary data. Deepseek is therefore likely more a marker of the maturity of generative AI than a breakthrough innovation like the arrival of ChatGPT.
In just two and a half years, and almost defying Gartner’s hype cycle, generative AI has moved from an emerging technology to a daily tool.
The market has rapidly structured itself around:
The AI industry is also being shaped by exogenous constraints, such as upcoming regulatory requirements, like the EU AI Act (effective February 2, 2025). This act, among other things, encourages the explainability of algorithms, which ensures transparency and accountability. This allows us to understand the reasons behind decisions made by AI, identify potential biases, detect errors, and ensure responsible usage.
The pioneering phase is already behind us. Companies and users now know what they expect from AI. The models now have to adapt accordingly. But if generative AI is maturing, does that mean we’ve hit a ceiling? Not at all! A new rupture is already on the horizon: Agentic AI.
Generative AI is reactive: it responds to user requests, generates text, creates code or images, but it doesn’t act on its own.
Agentic AI, however, goes beyond simply generating content:
This means that instead of having AI simply assist users passively, we will have intelligent agents capable of anticipating, organizing, and autonomously executing tasks.
Agentic AI will have a direct and profound impact on the world of work, especially in IT, although like generative AI, it may take time to become widespread in companies.
Today, DevOps and RPA already allow for automatic actions. With Agentic AI, a system could detect an anomaly without having encountered it before or without prior configuration, analyze its cause, propose a solution, and apply it autonomously. Unlike traditional automation systems that follow strict pre-programmed rules, these AIs can adapt to unexpected or novel situations. And with Agentic AI, these systems will no longer just respond—they will also take action: handle a ticket or reset a service without human intervention.
Agentic AI therefore paves the way for systems capable of prioritizing tasks, allocating resources, and making decisions autonomously.
But how far will this evolution go?
For now, these agents remain supervised and operate within clearly defined frameworks. In the future, entire ecosystems managed by AI, capable of collaborating with each other, may take charge of complex processes with little or no human intervention. We are already starting to see platforms for multi-agent AI systems.
While generative AI has changed the way we work and access information, Agentic AI will likely transform the way we think about tasks and how we go about achieving our missions.
