Deep discount: DeepSeek can alter the rules of artificial intelligence game

The recent release of two open source artificial intelligence (AI) programmes and chatbots from Chinese company DeepSeek has caused a tectonic shift in perceptions. DeepSeek achieved several large technical breakthroughs in training its R-1 and V-3 algorithms for a claimed cost of around $5.6 million, which is a fraction of the hundreds of millions required to develop ChatGPT, Grok, Claude, Llama and other large language models (LLMs). The new AI has performance standards rivalling the above products. But it was developed on relatively low-end chips, which cost a lot less than the premium graphic processing units (GPUs) used for ChatGPT. Moreover, it is claimed the training was on a configuration of 2,000 GPUs — a fraction of the size of the chip-clusters used to develop ChatGPT.

 

Since the code is open-source with an application programming interface (API) made easily available, the code can be checked and modified by anyone and any programmer can write applications around it. It may even be possible to install and run the programme on off-the-shelf computers, and certainly possible to do so in big hyperscaling data centres. OpenAI, Google, Meta, and other AI developers, are likely to incorporate some of the concepts demonstrated by DeepSeek to improve the next generation of generative AI models. The code indicates DeepSeek has developed a more efficient memory management system. It also introduced a couple of other technical wrinkles, which help optimise use of computer resources. This enables DeepSeek to train and work off less powerful hardware while delivering comparable performance to its rivals.

 

This is a classic example of finding workarounds to cope with constrained resources. It was forced upon DeepSeek due to the fact that the US has successively cut off supplies of high-end GPUs to China. There are quite a few implications. As such, DeepSeek’s success calls the entire concept of US protectionism and denial of high-end computing resources into question. This triggered innovation that produces equivalent performances with fewer resources. Moreover, by going open-source and allowing easy access to an API, DeepSeek has enabled easy proliferation of the algorithms. Despite concerns about data privacy, given Chinese parentage, the chatbot has overtaken ChatGPT in popularity where downloads are concerned. Programmers everywhere will be looking for potential applications riding these models. The dramatic reduction in development costs will also spark more investment in AI R&D.

 

In geopolitical terms, the United States will have to rethink its strategy of trying to maintain an edge in AI by denying access to computing resources outside the US and a few chosen allies. Now that these new concepts have been released globally via an open source code, it is likely that the next generation of generative AI iterations out of Silicon Valley will be more powerful, but those will also be matched by similar models released elsewhere. In economic terms, DeepSeek has caused an upheaval in the markets. Chipmaker Nvidia, which is at the cutting-edge of AI development, has seen 20 per cent wiped off its market value in a few days (even though DeepSeek was developed on Nvidia chips). Other listed companies and unicorns working in AI have also seen valuation downgrades. India’s policymakers should take note. Indian engineers are good at this sort of adaptation. What they may need is incentives. Incentivising AI-related research in India by offering targeted schemes could serve as a powerful accelerator. The rapid scaling up of India’s data centre capacity, alongside the substantial domestic market and large digital economy, could make India an excellent test bed.