Data, not chips, will decide India’s AI and biotech future, say deeptech founders

The founders mean that chips, computing power, and applications can be upgraded or replaced. Data is different. It captures the local language, biological material, disease patterns, and operating conditions that determine whether technology works in practice. Without relevant data, better hardware alone cannot solve India’s problems.

Arvind Pani said companies can keep replacing compute and applications, but cannot replace the data. In language AI, models trained on scripted studio recordings may fail during everyday conversations. In biotech, Loopworm spent 12 to 13 months cleaning data for Indian silkworm strains and pathogens before its tool could support useful experiments.

This gives proprietary, context-rich data strategic value. India still imports about 73% of key biopharma materials from China, while most Indians are not comfortable using English alone. The founders therefore want India to design technology for its own needs and fund the basic data generation required for future AI and drug discovery.

Trusted, context-rich data is accurate information gathered from the environments where a technology will actually be used. It includes local speech, accents, biological strains, pathogens, and other details that raw datasets may miss. It matters because AI and biotech systems learn from those patterns. If the underlying information is poor or irrelevant, the output will also be unreliable.

Pani said smaller amounts of enriched data can outperform huge volumes of raw data. Studio recordings made from scripts may teach an Indian language model correct words, but not interruptions, accents, or natural conversation. Loopworm similarly had to clean biological data and adapt it to Indian silkworm strains and their pathogens.

Changing computers or models can improve speed and processing. It cannot supply the missing examples or correct inaccurate records. That is why Pani described proprietary data as a competitive moat, citing Harvey, which changed its models and compute while retaining domain-specific data as its advantage.

The article describes gaps in both supply and usable technology. India still imports about 73% of key materials, including active pharmaceutical ingredients and reagents, from China. It also has a language-access gap: only about 4 to 5% of the population is comfortable using English alone. These limits affect both innovation and adoption.

India supplies a large share of the world’s vaccines by volume, yet that represents only about 5% of the market by value. Pani also said India’s internet was built for English users, forcing most citizens to learn technology and another language. Even English-language AI can struggle with regional Indian speech.

The figures suggest that scale alone is not enough. India needs domestic biological capabilities and datasets covering its languages and accents. The founders also called for public and foundational funding, because traditional investors are unlikely to support basic data generation and fundamental science.

Large datasets are not automatically useful datasets. When recordings come from scripts and controlled studio settings, models may learn polished speech rather than the variations people use in daily life. They can perform well under similar test conditions but struggle when speakers change pace, accent, wording, or conversational style.

Pani gave Indian language datasets as an example. Many are recorded in studios from scripts, so models trained on them may recognise prepared sentences but fail in real-world conversations. He argued that smaller amounts of enriched, contextual data can outperform much larger volumes of raw data. The same principle applies to biological information, where Bagaria warned: garbage in, garbage out.

The consequence is unreliable technology. A model may appear accurate but fail for regional accents or ordinary users. India therefore needs datasets gathered from real contexts and checked for quality. Better data can make AI more useful without simply making the dataset larger.

Datasets must match the conditions where a system will operate. Indian speech includes many languages, regional accents, and local ways of speaking. Biological systems also vary. A silkworm strain or pathogen found in one setting may not behave like one found elsewhere. Data designed around another country can miss these differences.

Pani noted that models trained mostly on American English can struggle with Australian or regional accents, and that English spoken across Indian states varies widely. Bagaria said Loopworm’s data had to be tailored to Indian silkworm strains and the pathogens affecting them. The company spent 12 to 13 months cleaning existing information.

These examples show why local data is not a minor adjustment. It determines whether AI understands real users and whether biotech experiments reflect local biology. India must gather and characterise its own languages, organisms, and conditions instead of assuming another country’s datasets will solve its problems.

Loopworm’s work shows how better data can change biotech operations. The company built a digital tool that maps the inner workings of the silkworm. Before the tool could be useful, it spent about 12 to 13 months cleaning existing data and adapting it to Indian silkworm strains and their pathogens.

That preparation improved the way experiments were selected and run. The data-led system helped Loopworm reduce a set of experiments from 10 months to two and a half months. Their cost fell from Rs 25 lakh to about Rs 4.5 lakh. The tool supports Loopworm’s work using silkworms to produce sustainable animal feed and high-value recombinant proteins.

The result is a practical example of data reducing trial and error. Faster, lower-cost experiments can help a biotech company test more efficiently and bring products closer to market. Bagaria believes AI could have its largest impact on biotech within 10 years.

Drug discovery is the process of identifying and developing new molecules that may become medicines. It involves finding promising biological candidates and studying how they work before a product can be developed. The article presents discovery as a foundational capability, not simply the manufacturing of an existing medicine.

Biosimilars are medicines designed to be highly similar to an already established biological medicine. Bagaria said India has operated in the biosimilar space for a long time, where another party discovers the molecules. In that model, Indian companies can produce or develop comparable products without originating the underlying discovery.

The difference is strategic. Discovery creates original molecules and intellectual property, while biosimilar production builds on molecules found elsewhere. Bagaria argued that India should move into drug discovery, especially because it imports about 73% of key materials. He also called for public funding for basic data and the sequencing of Indian organisms.