Google DeepMind's India research team is playing a central role in advancing the company's flagship Gemini AI model, handling everything from multilingual understanding to voice interactions and model efficiency. The roughly 75 researchers based at Google Ananta, the company's Bengaluru campus that opened in February 2025, are now at the forefront of making Gemini smarter across multiple dimensions.

What Is Google DeepMind's India Team Actually Building?

Manish Gupta, Senior Director for India and Asia-Pacific at Google DeepMind, explained that the Bengaluru team's work spans several critical areas that directly impact how Gemini performs globally. Rather than handling peripheral tasks, these researchers are embedded in core model development, tackling problems that affect millions of users.

The team's focus areas include:

• Multilingual Understanding: Improving how Gemini comprehends and responds to languages beyond English, a crucial capability for a global AI model.
• Voice Interaction: Enhancing Gemini's ability to understand and respond through spoken language, making AI more accessible to non-text users.
• Continual Learning: Developing systems that allow Gemini to improve over time without requiring complete retraining.
• Model Efficiency: Making Gemini faster and less resource-intensive while maintaining or improving its capabilities.

This work is significant because it demonstrates that world-class AI research is no longer confined to Silicon Valley. The fact that Google DeepMind chose to station a substantial research team in India, rather than treating it as a support or engineering hub, signals confidence in local talent and the quality of work being produced there.

Why Is India's AI Talent Pool Suddenly Competitive?

Gupta noted that India's challenge is not a shortage of talent but rather a lack of ambitious, compelling research problems. Several researchers on his team chose to return to India after earning PhDs from elite institutions like MIT, Carnegie Mellon University, and the University of Texas at Austin, specifically because they found intellectually challenging work to pursue.

"The biggest thing that talent needs is exciting work. So if there is exciting work happening, talent will come back," said Manish Gupta.

Manish Gupta, Senior Director for India and APAC at Google DeepMind

This observation flips the conventional narrative about brain drain. Rather than India losing talent permanently, the issue is that researchers leave to find cutting-edge problems elsewhere. When those problems exist locally, they return. Google DeepMind's Bengaluru operation has become one such magnet, proving that India can retain and attract world-class researchers if given the right opportunities.

What Does India Need to Build More World-Class AI Companies?

Gupta argued that India needs far more AI startups with the ambition and resources of companies like Sarvam and Emergent, which are building foundational AI models rather than just applications on top of existing models. However, ambition alone is insufficient. India Inc. must dramatically increase research and development spending.

The numbers tell a stark story. Indian industry currently spends approximately 0.35 percent of revenues on research and development. By contrast, China and the United States spend upwards of 3 percent, roughly nine times more. This gap directly translates to fewer breakthrough innovations and a smaller pool of resources for ambitious research projects.

"The Indian industry spends close to 0.35 percent of its revenues on R&D. If you look at China and the United States, the numbers are much higher, like upwards of 3 percent," explained Gupta.

Manish Gupta, Senior Director for India and APAC at Google DeepMind

Gupta also emphasized the importance of tackling what Google DeepMind CEO Demis Hassabis calls "root-node problems," which are difficult foundational challenges that, once solved, unlock solutions to many downstream problems. India's AI ecosystem tends to focus on lower-hanging fruit and incremental improvements rather than these transformative research directions.

How Can India Build a Complete AI Ecosystem?

Rather than specializing in a single layer of the artificial intelligence value chain, Gupta argued that India should develop capabilities across the entire stack. This means building expertise in hardware, foundational models, fine-tuning tools, applications, and deployment infrastructure.

A key part of this strategy involves leveraging open-weight models, which are AI models whose underlying parameters are publicly released. Google's own open-weight offering, Gemma, is built using the same research and technology as its flagship Gemini models. The latest release, Gemma 4, crossed 200 million downloads in less than three months, demonstrating significant developer interest.

"Once those weights are available, there is no going back. Nobody can take those weights away. There is a lot of work that can happen by building on those open-weight models," noted Gupta.

Manish Gupta, Senior Director for India and APAC at Google DeepMind

This point carries particular weight in the context of AI sovereignty. Recent controversies surrounding the availability of Anthropic's proprietary models highlighted how countries can become dependent on foreign AI infrastructure. Open-weight models provide a hedge against such restrictions, allowing developers to continue building even if access to proprietary models is restricted.

What About the Rising Cost of Advanced AI Models?

A common concern in the AI industry is that frontier models, which are the most advanced and capable systems, are becoming prohibitively expensive. Gupta acknowledged this reality but offered important nuance. While frontier models do command higher per-token costs due to their advanced reasoning abilities, the cost of achieving equivalent capabilities has actually fallen.

For example, Gemini Flash, a more recent model, typically offers roughly the same capabilities as the previous generation Pro model but at a lower token cost. This means developers can achieve the same results for less money by using newer, more efficient models rather than paying premium prices for older frontier models.

The distinction matters for developers and enterprises evaluating AI costs. While the absolute price of the most advanced models has increased, the cost-per-capability metric, which is what actually matters for budgeting, has improved. This trend suggests that AI adoption will remain economically viable even as models become more sophisticated.

Google DeepMind's investment in India research, combined with Gupta's candid assessment of what India's AI ecosystem needs, paints a picture of an industry at an inflection point. The talent is there, the infrastructure is being built, and the problems are compelling. What remains is whether India's corporate sector will commit the resources necessary to compete at the frontier of AI research.