Google DeepMind’s India chiefs on the race to make AI cheaper, safer — and finally good at code

Alphabet’s CEO Sundar Pichai’s admission in a New York Times podcast that Google is “falling a little behind” in AI coding tools has become one of the more pointed lines to emerge from the search giant’s leadership this year. It is also, it turns out, a fair question to put to the two executives who run Google DeepMind’s operations in India — a market the company increasingly treats as a proving ground for making its Gemini models cheaper, faster and more useful to build with.

Manish Gupta, who leads research for Google DeepMind India, and Seshu Ajjarapu, who heads applied AI for the unit, agree that closing the coding gap is now a top-tier priority internally, ranked alongside the company’s other most urgent work streams.

“Code is a top priority — P0, P1 and P2,” Mr. Ajjarapu said, using Google’s internal shorthand for its highest-stakes projects. The reasoning goes beyond product parity as coding tasks offer verifiable, checkable rewards and demand structured logical reasoning. Improving a model’s ability to code has knock-on benefits for its performance more broadly.

A local lab with a global brief

The pair’s remarks sketch a research operation that is export-oriented as Mr. Gupta shared about Matryoshka-inspired transformer — a technique developed by the Bengaluru team that nests smaller models inside a larger one — much like the Russian nesting dolls that fit one inside another.

First deployed on the Nano 3 model on Pixel phones, the approach lets an application call on only as much model as a task actually requires, extending battery life on-device. Google is now working to bring the same nesting principle to server-side workloads, where the payoff would be lower compute costs rather than battery savings.

That efficiency obsession, both executives said, is rooted in India’s market conditions rather than being a side project. “There is an inherent interest in making models more efficient” given the country’s population size and price sensitivity, Mr. Gupta said, and techniques born from that pressure have fed back into making Gemini models among the more efficient in the industry.

On the research side, Mr. Gupta describes work on finding “the right amount of thinking” a model should apply to a given problem — enough to avoid underperforming on hard tasks, but not so much that it wastes compute overthinking simple ones.

Token economics for enterprise AI

Much of the conversation returned to a theme Mr. Pichai himself raised recently on the sheer scale of token consumption in the current AI race, and what it means for companies trying to control costs. Mr. Ajjarapu framed the issue in terms of “economic value” — Google’s job, he said, is to lower the cost per token while raising the quality of what each token produces, particularly for agentic use cases, which are harder to price because their outputs are non-deterministic.

He suggested the industry’s pricing model may eventually shift away from tokens altogether and toward charging for completed tasks.

On enterprise trust and safeguarding IP, Mr. Ajjarapu offers a distinction he says is increasingly central to how customers think about deploying foundation models: a line between “public data”, which models are pre-trained on, and “private data”, which the model itself never sees.

In his view, the competitive moat for enterprises no longer sits with the underlying model but with a company’s own private context, workflows, tools and domain expertise. Mr. Gupta added that Google’s contractual position is unambiguous: customer data remains the customer’s, models do not learn from it, and all training takes place on the original pre-training corpus.

On agriculture, health and multilingual search

The India team’s remit extends well past model architecture. Mr. Gupta described an agricultural landscape model, built using satellite imagery, that can identify farm boundaries and crop types at the level of an individual field — data now being made available via API to Indian startups building products for crop insurance and credit assessment.

On healthcare, Mr. Ajjarapu points to applications built on MedGemma for leprosy detection and reproductive health, which the team intends to open-source for use across India.

Gemini’s language work, meanwhile, has been pushed into 25 Indian languages including Sanskrit, with Mr. Gupta citing adoption ranging from merchants in Surat to Tata Steel’s use of the technology for customer care and shop-floor safety.

In Search, the pair say AI Overviews and AI Mode have driven double-digit growth globally, though Mr. Ajjarapu is careful to frame advertising as a downstream concern: “If we create user value, we will figure out the rest,” he said, adding that the company is not trying to fast-track monetisation ahead of proving that value.

Asked whether Google DeepMind’s robotics ambitions extend to India, Mr. Gupta said that the company’s robotics research remains centred elsewhere. The India lab’s mandate, he says, is best summed up as “India first, but not India only.”