The year 2025 started with a bang with the emergence of DeepSeek as a foundational Large Language Model (LLM) from China that surpassed leading LLMs from OpenAI, Google and Anthropic on many benchmarks. The multiple whammies for the market-leading models were that the DeepSeek performed better, had been developed at 1/20th the cost ($5 million vs $100 million for Open GPT models assuming $2 per GPU hour), used far fewer graphics processing units or GPUs (2000 H800s vs 100,000 H100s for OpenAI), and was trained for only two months before becoming operational. A lot has already been written about the capabilities of DeepSeek. The questions that now come to the fore are: What does the emergence of DeepSeek portend for the AI race between the US and China, and where does India stand in this race?
Why China can win the AI race
Story continues below this ad
The Chinese quant hedge-fund High-Flyer that developed DeepSeek has proven beyond doubt that developing foundational GenAI models need not be an expensive venture. The cost numbers mentioned above did not consider the salary and R&D costs. Given that the salary costs in China are lower than the US, the former has a competitive edge in terms of overall cost. If we are mesmerised by what DeepSeek has been able to achieve, wait for Qwen from Alibaba and DuoBao from ByteDance to come out. At the time of writing this article, Alibaba announced the release of Qwen 2.5-Max, claiming to outperform DeepSeek-V3. There are over 10 labs in China that are off the scale of OpenAI and tens of tier-two labs.
So how do these labs flourish? There is tremendous support from the Chinese government both at the central and provincial levels. There is no paucity of funds (hint: China has a trade surplus of $1 trillion). Talent is not scarce; Tsinghua University and Peking University have admission standards that are tougher than those of our IITs. There is an embargo on exporting Nvidia chips to China. China develops its own GPU chips called the Ascend 910B, which are at par with the Nvidia series A100 chips. Soon, China will have its own version of GPUs with 80 billion transistors to compete with Nvidia H100 chips and will become an exporter of GPUs.
Training and running foundational models require that large amounts of power be available to the data centres. This is not an issue with China, which has an energy surplus and continues to invest in solar, hydro and nuclear power. Access to DeepSeek is free of charge, unlike OpenAI models which are charged at $200 per month for unlimited use. This is going to further propel DeepSeek to the top.
Story continues below this ad
China has learned to improvise and innovate, especially when it is pushed to a corner. If China can develop sixth-generation fighters ahead of the US, develop an “artificial sun” and make it almost practical and send humans to space, then there is nothing that can stop China from leading the AI race. Only China can undo China.
As mentioned earlier, all innovation in China is heavily funded by the government. Any shift in the government policy will have an impact on the area of research affected. China has still not emerged from the impact of Covid. Industrial growth is shrinking, sustainable growth recovery is yet to be seen. While the US has been at the forefront of innovation for the last few decades, the arrival of China as a pioneer and as a country that can excel under constraints has put it in an enviable position. US President Donald Trump is correct to say that DeepSeek is a wake-up call.
Where does India stand
India has no dearth of talent and brain power. India does, however, lack a risk appetite. Venture capital and investment funds are leery of investing in startups developing foundational models. And their fears are justified. Startups in AI tend to have very high valuations at minimal revenue streams. The mistakes of the 1990s when startups that barely had an “html” website were funded have taught a valuable lesson. Furthermore, there is not much support for R&D in terms of funds and other resources from the Indian government. There is also a general chatter that Indian startups should build on what is already available.
Instead of developing foundational models, Indian companies should take existing LLM templates and train them with their specific data. When the likes of GPT, Gemini, Llama, and Claude among others are already available, then building another foundation model would be a purely academic exercise. Even if startups were to venture out and build LLMs, they would not be able to win the price wars against the market leaders. The only way this would be possible is if India were to foster a protective culture just like China and not allow foreign players to operate on Indian soil. But can India, a democracy, afford to do that? Not really.
most read
India, thus, should first invest in building infrastructure such as data centres to foster the growth of AI. At the same time, investment should be made in the power industry so that these centres could be made operational. India has made significant advances in hydrogen fuel cell technology which could potentially be a source of power. The US has put India in a category that can receive only a limited number of H100 GPU chips. Though companies like Reliance and Adani Power dream of building the largest data centres in the world and may have the means for the needed gigawatts of power, their ambitions would be restricted by the availability of GPUs. Even foreign enterprises may find this to be a deterrence when contemplating setting up computational facilities in India.
Under the IndiaAI program, the government wants to procure 10,000 GPU chips and distribute them among private companies. The number of chips, however, would fall short of the required quantity. One way to be a serious player in the AI race is to form a consortium where countries pool their limited resources and benefit from the advancements in AI. Why not form a Quad for AI? The other option is to do what India did in the 1960s when the US would not export wheat to India or later in the 1990s when it would prevent the transfer of GSLV engine technology. India invested in R&D and became self-reliant. Maybe that is what is needed in the AI space to win the race.
The writer is the Chief Data Scientist at Yatra Online with 25 years of experience in product, data and technology. He is a graduate of Massachusetts Institute of Technology and IIT, Delhi
