DeepSeek, AI sovereignty, and India

Along came DeepSeek-R1^[1] last week, an open-source large language model (LLM) reportedly rivaling OpenAI’s top offerings, sending shockwaves through the industry and generating much excitement in the tech world. It apparently started as a side project at a Chinese hedge fund before being spun out. Its efficacy, combined with claims of being built at a fraction of the cost and hardware requirements, has seriously challenged BigAI’s notion that “foundation models” demand astronomical investments. I have personally been playing around with R1 and have found it to be excellent at writing code. Speaking of foundation models, one rarely hears that term anymore; unsurprising, given that foundation is now commodity. Building a foundation-level LLM was once touted as the cornerstone of AI sovereignty, but that rhetoric has also waned. Much has changed regarding the idea of AI sovereignty.

When OpenAI’s and MidJourney’s generative AI models came out, they changed not just the mainstream psyche, but the techno-geopolitical landscape. Soon thereafter, Meta’s and StabilityAI’s open-weight models broke the floodgates to decentralised development and iteration in the tech community. More than that, the number of AI breakthroughs that have been coming out of the global open-source realm has been nothing short of astounding. AI capabilities thought to be impossible can now be downloaded and run on commodity hardware. Freely accessible AI models along with the vast ecosystem of open-source tooling around them have become commodities. Meanwhile, large AI companies continue to burn massive amounts of cash offering AI software-as-a-service with no pathways to profitability in sight, thanks to intense competition and the relentless race toward commoditisation. Where does India’s idea of AI sovereignty fit in?

The past two roller-coaster years have offered ample evidence for some informed speculation: cutting-edge generative AI models obsolesce rapidly and get replaced by newer iterations out of nowhere; major AI technologies and tooling are open-source and major breakthroughs increasingly emerge from open-source development; competition is ferocious, and commercial AI firms continue to bleed money with no clear path to direct revenue; the concept of a “moat” has grown increasingly murky, with thin wrappers atop commoditised models offering none; meanwhile, serious R&D efforts are directed at reducing hardware and resource requirements—no one wants to bankroll GPUs forever.

If DeepSeek’s cost claims are true (~$5 million to train the model as opposed to hundreds of millions elsewhere), then hardware and resource demands have already dropped by orders of magnitude, posing significant ramifications for a lot of players. Given that, in India’s national perspective, does anchoring the idea of AI sovereignty on GPUs and foundation models matter? GPUs are a means to an end tied to specific architectures that are in vogue right now. Today’s LLMs are milestones in a decades-long R&D trajectory; tomorrow’s models will likely rely on entirely different architectures. Is AI sovereignty then about developing a generative model with national interests baked in? Or is it about government-backed GPU clusters for industry? If foundation-level open-source models of ever-increasing efficacy are freely available, is model creation even a sovereign priority? Could such attempts anywhere keep up with co-operative, global, open-source innovation?

As Carl Sagan famously said “If you wish to make an apple pie from scratch, you must first invent the universe.” Without the universe of collective capacity—skills, understanding, and ecosystems capable of navigating AI’s evolution—be it LLMs today, or unknown breakthroughs tomorrow—no strategy for AI sovereignty can be logically sound. Mountains of evidence at this point, and the dissipation of chest-thumping and posturing from the Indian industry, point to this inescapable reality. Everyone is going to use these innovations in all kinds of ways and derive value from them regardless. Consumption and usage of these technologies do not require a strategy, and production and breakthroughs in the open-source AI world will continue unabated irrespective of sovereign policies or goals. And of course, a new open-source model will beat R1 soon enough.

India’s AI sovereignty and future thus lies not in a narrow focus on LLMs or GPUs, which are transient artifacts, but the societal and academic foundation required to enable conditions and ecosystems that lead to the creations of breakthroughs like LLMs—a deep-rooted fabric of scientific, social, mathematical, philosophical, and engineering expertise spanning academia, industry, and civil society. Any AI sovereignty focus must thus direct resources to fostering high quality research capacity across disciplines, aiming explicitly for a fundamental shift in conditions that naturally disincentivise skilled, analytical, critical-thinking, passionate brains from draining out of the country. In fact, the bulk of any long-term AI sovereignty strategy must be a holistic education and research strategy. Without the overall quality and standard of higher education and research being upped significantly, it is going to be a perpetual game of second-guessing and catch-up. Realistically, the horizon for that is ten, if not twenty years, and that is okay, as long as we collectively accept this reality and strive to address it.

The secret sauce lies not in model weights, but in biological brains.

An abridged version of this piece was originally published in the Economic Times on 29 Jan 2025.