By S. Alex Yang, Professor of Management Science and Operations, London Business School
In December 2024, the then obscure Chinese company DeepSeek shook the artificial intelligence (AI) community by releasing its DeepSeek-v3 model, which achieved performance comparable to advanced models developed by Western AI giants, but at a fraction of model training and inference cost.
One month later, DeepSeek captured global attention with its DeepSeek-R1 model, which rivals OpenAI’s latest reasoning model.
This development has been hailed as China’s “ChatGPT moment”. While much discourse centers on the geopolitical rivalry between the U.S. and China, it fails to capture a more significant technological direction in AI – cost efficiency.
By honing programming techniques to maximize the utilization of Nvidia chips (a resource particularly scarce for Chinese firms due the US chip ban), DeepSeek has shown that high quality models can be made with much higher cost efficiency.
This focus prompts the industry to recalibrate of the direction for AI development.
Instead of focusing solely on the pursuit of artificial general intelligence (AGI), more attention should be directed toward developing cost-effective AI models and applications that can be widely adopted.
This approach represents another version of AGI – Artificial Good-enough Intelligence.
Generative AI enters the final stage of the innovation cycle
Economist Joseph Schumpeter’s theory of innovation outlines a three-stage process: invention, innovation and diffusion. In the invention phase, ground-breaking scientific discoveries emerge, such as the development of artificial neural network by the 2024 Nobel laureates John Hopfield and Geoffrey Hinton, and the transformer architecture that underpins the current Generative AI models.
The innovation stage involves the commercialization of these inventions, exemplified by OpenAI’s ChatGPT, which brought advanced language models to the public.
Finally, the diffusion stage sees widespread adoption and optimization of the technology, making it accessible and practical for a broader audience. DeepSeek’s advancements signify this diffusion phase, focusing on cost efficiency that facilitates mass deployment.
As one type of technology matures, the emphasis of innovation shifts to engineering and operational improvements that drive significant cost reductions. This philosophy is mirrored in Henry Ford’s invention of the assembly line, which revolutionized automobile manufacturing, making cars affordable to the average American family. In post-World War II Japan, manufacturers led by Toyota pioneered lean manufacturing to overcome severe resource constraints, ultimately surpassing their American competitors in both cost and quality.
As Generative AI transitions from the invention to the diffusion stage of the technology cycle, DeepSeek exemplifies the same philosophy, employing techniques akin to those used by quantitative traders to shave nanoseconds off execution speed. By open sourcing their innovations, DeepSeek has enabled the global community to collaborate and accelerate progress in this effort.
Value chain innovation for Generative AI
Yet the techniques adopted by DeepSeek are not without controversy.
When training their model, instead of using terabytes of text data generated by humans, DeepSeek utilized machine outputs from other Generative AI models, a technique known as model distillation.
While DeepSeek demonstrated that using a small set of carefully curated machine data could achieve significant cost saving without significantly sacrificing model quality, OpenAI accused DeepSeek of violating its terms of service by using its API to generate training data.
This dispute underscores a larger question: how should AI companies compensate content creators – human or other AI models – for using their data in model training?
This debate is not new. In December 2023, The New York Times sued OpenAI and Microsoft for copyright infringement, alleging that their content had been used without compensation to train AI models. This lawsuit ignited a global discussion on fair use standard in the context of Generative AI.
My research with Angela Huyue Zhang (University of Southern California Gould School of Law) highlights an emerging reality in the AI value chain: As AI generated content becomes more valuable, compensating human creation becomes crucial. After all, their original work remains essential to advancing frontier AI.
The same reasoning applies to synthetic training data, like that used by DeepSeek. If the output from advanced AI models could help develop cost-efficient alternatives, frontier AI companies should be incentivized to continue pushing technological boundaries.
If pursued, this dynamic may lead to a two-tier structure for AI development: a small set of companies will focus on developing cutting-edge frontier AI, the output of which could be provided to a larger group of mass AI model developers to build cost-efficient models and applications.
Such an ecosystem balances capability and efficiency, but it requires a well-structured market for synthetic data. Establishing pricing models for AI-generated outputs used for model training could create a new revenue stream for frontier AI developers, ensuring continued investment in cutting-edge research.
S. Alex Yang is a Professor of Management Science and Operations at London Business School. He is an expert in supply chain management, finance and technology. His recent work focuses on innovations and governance in the space of digital technologies and business models, such as AI, blockchain, and digital platforms.