India can leverage DeepSeek’s AI moment to build its own breakthroughs: Accel’s Prayank Swaroop

And while the final results clearly need more refinement, DeepSeek's biggest achievement has been in raising conversations around reducing the cost of building AI models, he said.

Excerpts from the interview:

Q: Are you convinced or somewhat sceptical about claims made by Deepseek and Qwen given limited information at this point? A: The likes of DeepSeek and Qwen have showcased how innovation in fundamental AI research is indeed possible – contrary to OpenAI chief Sam Altman’s words to India to “not even try” building foundational models. When a venture claims to bring down the cost of a pioneering innovation by 100x and promises to open it for the world to see, you cannot help but take notice.

Think of this as a moment similar to India’s frugal space missions – as against Perseverance’s cost of over USD 2.5 billion, India completed its first Mars mission successfully in just USD 74 million, a meagre three per cent of NASA’s.

That said, deeper scrutiny of DeepSeek has revealed clear shortcomings, such as in simple reasoning, contextual generation of content, and political bias. This shows that innovation is not a one-size-fits-all glove for all nations to wear. While I’m fond of its frugality premise, the final results clearly need work.

Q: What does DeepSeek's rise mean for the competitive landscape of AI startups globally, and in India? A: DeepSeek’s biggest achievement was in raising conversations around reducing the cost of building LLM. As we iterate upon its open-source R1 LLM, we’re sure to find ways to further reduce the cost of AI, and bring down the dollars-per-token metric in generative AI. This will accelerate the pace of innovation in AI, and can democratise AI from the current strongholds of the US, to the rest of the world.

For India, I’m conflicted about what this truly means, even though cabinet minister Ashwini Vaishnaw has spoken about letting developers in India access its open-source framework to build upon it. This is because DeepSeek’s claims, since rising to the fore on January 20, have now been widely questioned through increasing discrepancies.

But the conversation that DeepSeek brought to the fore can now be capitalised upon by Indian AI engineers and researchers to build a foundational AI model with a clearly auditable code base, and a transparent dataset repository. In fact, this can even lead us to make new, fundamental breakthroughs in AI.

Q. What implications does this hold for Accel's strategy to invest in AI technology? A: DeepSeek’s phase of hype has made us at Accel even more bullish on AI. We expect more generational jump phases to come by in AI, and we’re excited about the opportunities these innovations will bring in order for us to make a significant impact in shaping the future of every aspect of our lives.

Our immediate interest in AI is in deep-tech fields that will bring core innovation to the forefront. These include startups working on agentic AI, synthetic data generation, inference optimisation, enterprise and consumer apps built on multimodal AI models, data stack innovation, AI semiconductor, and cyber security in AI.

It is not just about replicating DeepSeek's success – with the right strategic funding, India can go well beyond and create new headlining discoveries.

Q. Is the time right for India and Indian startups to invest in foundation models, or are we a bit late? A: That India should be investing in building its own foundational AI models is not even a question. Think of this as owning our own technology, versus licensing it from companies governed by the law of another land – which can leave us at their mercy if geopolitical equations change.

India has missed the bus twice – first, when it came to the semiconductor boom starting in the 1960s, and second, when it came to setting-up a circular economy for electronics manufacturing.

Today, we’re racing against time to skill engineers in chip fabrication, and subsidising global players to set-up component shops and chip fabs in the country. We can’t afford the same with AI, for which foundational models are key.

To do this, we need substantial government support to bring India up to speed with the rest of the world. We already have the brainpower, but access to capital has often been an issue. Supporting initiatives such as the Central government’s USD 1.2 billion AI Mission, an upcoming second phase of the Indian Semiconductor Mission, and the ongoing National Quantum Mission, are good places to start.

We also need more R&D investments in core tech – the FY26 union budget’s announcements on fellowships programmes for IITs can help us start building up research abilities.

Q. In light of DeepSeek's success, and other new challengers being touted (like Qwen) what opportunities do you see for Indian startups looking to develop AI solutions? A: Solutions and applications that leverage AI will form the backbone of how much AI contributes to India’s economy, as a part of its GDP, in future.

It’s important to know that foundational models will not be built by everyone – they do, after all, take considerably larger amounts of time, money and effort in R&D. Being a layer-one level of solutions integrator for foundational models, or further onward, will offer the biggest business opportunities for hundreds or even thousands of startups in future. These startups, built successfully, will scale up and take India’s AI innovation to the world.

Owning the patent rights to a fundamental technology is one key part of India’s AI journey, but an even greater economic role will be played by those that build applications that appeal to users at scale.

Look at Big Tech today – there are only five or six companies that build the foundational technologies, but millions of others are the ones that impact the biggest chunks of the global tech economy. AI will go the same way too, and India will also see numerous such AI startups in the long run.