English 
A New DeepSeek Paper Highlights the Growing Push for More Efficient AI Training
The Chinese startup DeepSeek recently published a technical paper introducing a new training method for large artificial intelligence models, called Manifold-Constrained Hyper-Connections (mHC). At first glance, the work appears to be an academic contribution focused on neural network architecture. In practice, however, it goes further, directly addressing one of the biggest bottlenecks in modern AI: the economic and computational cost of training increasingly large models.
The study gained international attention after being highlighted in a Bloomberg article, which places the research within China’s broader push to improve efficiency in AI development—particularly in a context marked by restricted access to advanced chips and direct competition with companies in Silicon Valley.
The problem behind giant models
Over the past few years, the advancement of language models has followed a clear logic: more parameters, more data, and more computational power. This approach has delivered impressive gains, but it has also created an important side effect: training state-of-the-art models has become extremely expensive, both financially and in terms of energy consumption.
From a technical perspective, modern architectures such as Transformers rely heavily on residual connections, which help maintain training stability in deep networks. However, attempts to expand these connections — such as so-called Hyper-Connections — tend to introduce numerical instabilities, making training more difficult as models grow larger.
This is where DeepSeek’s proposal comes in.
What is mHC and why it matters
The Manifold-Constrained Hyper-Connections method proposes a mathematical reformulation of these expanded connections. Instead of allowing connections to grow freely and potentially become unstable, mHC constrains them to a specific mathematical manifold, preserving important properties such as identity mapping, essential for training stability.
In practice, the paper shows that this constraint makes it possible to scale larger models with greater predictability, reducing issues like gradient explosion or degradation. The experiments presented indicate that the method adds a relatively small computational overhead to training while significantly improving stability in models with billions of parameters.
Although the paper is technical, its central message is clear: it is possible to improve model scalability without simply doubling computational costs.
Economic and strategic impact
This is precisely where the work gains relevance beyond the academic sphere. Training large AI models can cost tens or even hundreds of millions of dollars, in addition to requiring access to cutting-edge hardware such as the latest-generation GPUs.
For Chinese companies, this challenge is even greater due to chip export restrictions imposed by the United States. In this context, the pursuit of efficiency is not merely a technical advantage—it is a strategic necessity.
DeepSeek had previously drawn attention with its R1 model, which was developed at significantly lower costs than those incurred by major Western laboratories. mHC reinforces this narrative: rather than competing solely on brute scale, the company is betting on architectural innovation and mathematical efficiency as a competitive differentiator.
A signal to the AI industry
Although the paper does not directly announce a new commercial model, it hints at the foundations of a broader strategy. Methods such as mHC could influence future generations of models—not only in China, but across the entire AI ecosystem—as the industry increasingly questions the sustainability of the “bigger is better” approach.
More than an isolated technique, DeepSeek’s work points to a shift in focus: optimizing architecture, stability, and cost rather than relying exclusively on more hardware. In a global landscape shaped by energy constraints, regulatory pressure, and rising costs, this approach is likely to become increasingly relevant.
Conclusion
The new method presented by DeepSeek shows that meaningful advances in artificial intelligence do not depend solely on more data or more GPUs. By addressing fundamental architectural and stability challenges, mHC reinforces the idea that efficiency can be just as strategic as scale.
If this line of research gains traction, it could redefine not only how we train large models, but also who is able to compete in this market—a crucial factor in the ongoing global race for AI leadership.
