Generational

Interests

AI, Computing & Frontier Technologies

Applied AI, compute, robotics and the deep technologies India has to build or buy.

How we see it

India will meet AI mainly as a user of models and only secondarily as a maker of them. The largest value over the next decade lies in applying capable models to Indian problems: a farmer asking about a scheme in Marathi, a clerk reconciling a land record, a hospital triaging scans. That is a deployment challenge, and it depends on data, language coverage, integration with existing systems and trust. We therefore ask whether an AI business holds a workflow that a customer cannot easily give up, and care less about which model sits underneath it.

Physical AI is software that perceives, decides and acts through a machine: a robot arm that picks mixed items, a drone that inspects a power line, a sprayer that treats only the weeds, a camera system that stops a defective part on a line. Recent advances in vision and learning have made robots far more adaptable than the programmed machines of the past, and humanoid designs now draw attention and capital. India's angle differs from that of the United States, Europe or China, because the settings are messier and the budgets tighter.

Deep technology differs from ordinary software in its timelines and its risks. A chip design, a quantum device, a new alloy or a biologic can take many years to reach a customer, and the technical question is only the first one. The second is whether anyone will buy it at a price that supports the business. India has strong research institutions and a deep pool of engineers, but the path from laboratory to product has too few steps, and too little money for the middle ones. We care most about that gap.

Why now

Compute, models and public rails have arrived at the same time. The IndiaAI Mission is putting subsidised compute within reach of start-ups and researchers, Bhashini is building Indian-language translation and speech services as public infrastructure, and Aadhaar, UPI and DigiLocker give developers identity, payment and document layers to build on, and each lowers the cost of a new service. Few countries have all three layers in one stack.

Capable sensors, compute and batteries have become cheaper while model quality has climbed, which moves many tasks from laboratory demonstration to commercial pilot. Indian demand is arriving at the same time: e-commerce and quick-commerce fulfilment, electronics and auto plants with rising quality requirements, power and rail networks that need inspection at scale, and agriculture short of labour in peak seasons.

Trends we follow

Indian-language models move from demos to services

Speech and text models for Hindi and other Indian languages are now good enough for customer support, scheme enquiries and document handling, though dialects, code-mixing and low-resource languages remain weak. Public datasets and language services lower the cost of building. We look at who owns the voice interface in banking, health and government, and whether quality holds up outside metro accents.

Retrofit automation in existing factories

Most Indian plants are brownfield, with mixed equipment and little spare floor. Vision-guided inspection, collaborative arms and mobile robots that can be added to an existing line are easier to buy than a new automated plant. Buyers want payback they can see within a budget cycle, which favours simple, narrow tools over general-purpose humanoids.

Semiconductor activity broadens from assembly to design and materials

Under the India Semiconductor Mission, packaging and test plants and a first set of fabrication projects are being developed, and the design-linked incentive scheme backs chip design firms. Skilled people, specialty gases and chemicals, and reliable utilities determine whether plants succeed. We follow design talent, equipment supply and which products customers will qualify.

Inference, not training, drives the compute build-out

Shared compute under the IndiaAI Mission was designed with model training in mind, but most enterprise and public demand will be for inference: running models in production at predictable latency and cost. That favours data centres near users, efficient serving software and smaller models tuned for a task. We follow who controls cost per query, since that decides which services a state government or a small bank branch can afford.

Warehouses and dark stores as the first scale market

E-commerce and quick-commerce operators run dense, repetitive operations with good data and a willingness to experiment. Sorting, picking assistance, autonomous carts and inventory scanning are being tried at scale. The question is whether Indian-built systems can match imported ones on reliability while undercutting them on price and servicing.

Quantum work moves from national mission to applied devices

The National Quantum Mission sets goals in computing, secure communication, sensing and materials, with hubs at leading institutions. Quantum sensing and communication may reach users before large computers do. We look at which hardware groups have credible roadmaps, how they work with industry and defence users, and what supply chains for components they depend on.

What we look for

  • We look for companies that own a workflow in a regulated or high-volume sector and can show that customers renew because the product is embedded, not because the model is novel.
  • Teams that have deployed machines in real Indian plants, farms or warehouses and can show uptime and cost data from the field.
  • We look for founders with a technical edge that is hard to copy, such as a patented process, a proprietary dataset or an unusual manufacturing capability.
  • We look for teams that treat Indian languages, dialects and voice as the primary interface and test on real speech and real documents.
  • Companies that own the perception, control or data software that makes a machine useful and can run it on commodity hardware.
  • We look for companies that already have a named pilot customer who has paid for a trial, rather than a market estimate.

Questions we are working on

  • Where should India insist on domestic capacity, where should it buy the best available model, and who decides case by case?
  • Where does automation pay back in India when labour is cheap, and where does it fail?
  • Which deep-tech sectors can India reasonably lead, and which are better pursued through partnerships and selective capability?
  • How can public datasets be opened for training under the DPDP Act, 2023 without weakening consent?

Writing on ai, computing & frontier technologies

Part of 7 cross-cutting trends on the interests page.

All interests