AI and deep tech in India, explained

What counts as AI and deep tech?

AI and deep tech covers startups whose main edge is difficult technology rather than a new way to sell something. A food delivery app is a business-model idea. A company that trains a language model, designs a control chip or builds a launch rocket is a technology idea. The second kind takes longer to build, costs more before the first sale, and is harder for a rival to copy.

StopDown groups this sector into four areas. Foundation models and generative AI are the large models that write, speak and reason. AI dev tooling and agents are the products developers and companies use to build on those models. Computer vision and applied AI turn images, video and data into decisions in factories, farms, banks and hospitals. Robotics, chips and other deep tech cover physical hardware: robots, semiconductors, drones, space and quantum computing.

The word deep tech is loose. Some people use it for anything science-heavy, others only for hardware. This guide uses the wider meaning, because in India the same investors, the same government schemes and the same talent pool cut across all four areas.

How did AI and deep tech grow in India?

India had a large services industry long before it had AI product companies. Firms such as Fractal Analytics grew by helping big companies use data and analytics, and that experience trained a generation of engineers who later started their own companies. Fractal is a well-known example of an Indian analytics firm that moved toward AI.

The jump came with generative AI from 2023. Founders realised that Indian languages, Indian accents and Indian cost levels were an opening. Sarvam AI, which builds models for Indian languages, and Gnani.ai, which works on enterprise voice AI, are among the names that grew from this idea. Krutrim, which positions itself as an AI cloud infrastructure provider, and E2E Networks, which rents GPU computing power, show the infrastructure side of the same wave.

In parallel, India opened its space sector to private companies. Skyroot Aerospace, a private rocket maker, and TakeMe2Space, which works on computing in orbit, are examples. Defence and drone work also drew private builders, such as Flying Wedge Defence & Aerospace. So by October 2026 the sector has two tracks: software-led AI that can reach customers in months, and hardware-led deep tech that needs years and patient capital.

Who are the customers?

Most AI startups sell to businesses rather than to individuals. Banks, insurers, telecom firms, retailers and manufacturers buy voice agents for customer calls, document reading tools, fraud and quality checks, and analytics. A voice AI firm like Gnani.ai or Smallest.ai sells to companies that run large call centres. Bolna AI describes itself as a voice AI orchestration platform, which means it connects speech, language models and phone systems so a company can deploy a calling agent.

The government is a second big customer. Departments buy chatbots, language tools and data platforms, and defence and space agencies buy drones, sensors and launch services. A recent headline from the StopDown data says CoRover won a MeitY contract for a government agentic AI platform, a good picture of how public buyers work with startups. An agentic AI system is software that can plan and carry out several steps on its own, not just answer one question.

Developers are a third group. Companies such as Emergent, which describes itself as an AI software creation platform, and Soket AI, which released an agent harness (a framework for running AI agents) in early October 2026, sell tools to people who build software. Consumer products exist too, such as MagTapp, an AI browser for Indian readers, but consumer AI in India has found it harder to charge money.

How do these companies make money?

Software AI companies usually earn in one of three ways. They charge per use, for example per minute of voice call or per thousand words processed. They charge a subscription, often per seat or per company. Or they take a project fee to build and run a custom system, which looks more like consulting. Usage pricing is common because AI models cost real money every time they run.

Infrastructure companies rent capacity. A GPU is a chip that handles the heavy maths that AI needs. A firm like E2E Networks or Neysa rents GPU time by the hour or month, and earns the gap between what its hardware costs over its life and what customers pay. Utilisation matters most here: an idle chip earns nothing.

Hardware and deep tech companies earn from product sales, contracts and development grants. A robot maker like ANSCER Robotics sells machines or robots-as-a-service to factories. A component maker like SEDEMAC, which designs and makes electronic control systems, sells to other manufacturers. A rocket company sells launches. All of them often get money from pilots and government programmes long before steady revenue arrives.

What does it cost to build, and who pays?

Costs differ sharply by area. A small team can build an AI application on top of an existing model with modest capital. Training a large model from scratch needs many GPUs and months of runs, so it needs far more. Hardware needs labs, prototypes, certification and often a factory. That is why deep tech rounds tend to be larger and slower than a typical app startup round.

In the last 12 months the most active backers of this sector in the StopDown data include Peak XV Partners, Accel, Lightspeed, Bessemer Venture Partners, Khosla Ventures and Prosus, along with corporate and strategic names such as Nvidia and HCLTech. India Accelerator also appears. A venture capital fund (VC) invests other people's money in young companies for a share of ownership. A corporate investor like HCLTech often wants a link to its own business as well as a return.

Recent headlines show the range of cheque sizes. Quanfluence raised $10M for a photonic quantum computer, a machine that uses light instead of electricity to compute. Desible.ai raised ₹32 Cr in a round led by Prime Venture. Founders in hardware often also use government grants, defence contracts and bank loans, because a pure equity route is expensive in dilution.

What is the government doing?

The IndiaAI Mission was approved in 2024 with a budget of about ₹10,300 crore over five years. A central piece is shared computing power: the government pays for GPUs from private providers and offers them to startups, researchers and academia at a subsidised rate through an AI compute portal. A March 2026 government statement said more than 38,000 GPUs had been onboarded to the portal. The mission also backs the building of Indian foundation models, and Sarvam AI, Soket AI and Gnani.ai are among the companies that have been named for this work.

For chips, the India Semiconductor Mission (ISM) offers financial support for fabs (factories that make chips) and for packaging and testing units. As of July 2026, twelve units had been approved under the first phase, spread over six states, and the Union Cabinet approved a second phase, Semicon 2.0, in July 2026. Details such as the exact outlay and rates for the second phase were still being reported differently by different outlets, so check the official ISM notices before relying on a number.

Space and drones are opened through separate bodies. IN-SPACe is the agency that authorises and supports private space activity, and drone use follows the Drone Rules, 2021, with a production-linked incentive scheme for drone makers. Rules and schemes change often, so treat every figure here as a summary, not a quote.

Which laws and regulators matter?

The Digital Personal Data Protection (DPDP) Act, 2023 is the main data law. Its Rules were notified on 13 November 2025, and the duties come in phases. As of October 2026, the Data Protection Board exists, the rules for consent managers are due around November 2026, and most company duties, such as notice, consent, security safeguards and breach reporting, are due by about May 2027. An AI company that trains on or processes personal data of Indians needs a plan for this well before that date.

There is no single AI law in India as of October 2026. Instead, existing laws apply: the DPDP Act for personal data, the Information Technology Act for online harms and intermediaries, copyright law for training data questions, and sector regulators such as the RBI for finance and the IRDAI for insurance. MeitY, the electronics and IT ministry, issues guidance. Founders should read the rules of the regulator that governs their customer, because a bank's rules flow down to its AI vendors.

Hardware has its own regimes: wireless and telecom equipment need approvals, drones need type certification and operator rules, and rocket launches need space authorisation. This is a reminder, not advice: check the current rules with a qualified professional before you build or launch.

What are the main risks?

The first risk is that big global labs move fast. A startup built on one feature of a large model can lose its edge when that feature is released free inside a bigger product. Founders try to defend with data, distribution, local languages or deep integration with a customer's systems.

The second is cost. Running models is expensive, and a product that is popular but unprofitable can burn cash faster than it grows. Usage-based pricing helps, but customers push back on price when many vendors offer similar tools.

The third is long sales cycles. Banks, governments and factories test slowly and pay slowly. Hardware companies face a further risk: one failed test, such as a launch or a prototype, can set a company back by a year. Add talent risk, because senior AI researchers are scarce and global firms recruit the same people, and the risk that a fashionable theme attracts more money than it can usefully absorb.

A final point for readers: announcements are not revenue. A partnership, a pilot or a memorandum of understanding is a start, and the signal to look for is repeat, paid use by customers.

What should you watch next?

Watch whether the Indian foundation model projects ship products that people and companies pay for, not only open models. Watch how GPU supply and subsidised access change the cost of training in India. Watch the second phase of the semiconductor mission and whether new fabs and packaging units start production.

Also watch the DPDP deadlines in November 2026 and May 2027, which will test how AI firms handle consent and data. Finally, watch the mix of investors: the arrival of strategic names such as Nvidia and large Indian IT firms alongside venture funds suggests that this sector is now seen as a long-term infrastructure bet, not a short-term app trend. Readers who want to follow the sector can track new rounds in the AI and deep tech space on StopDown.

The breakdown

Value chain: who does what, who earns

Business models

The numbers that matter

• Gross margin: AI products pay for computing every time they run, so margins are often lower than for ordinary software. Falling model costs help, but pricing pressure follows.
• GPU utilisation: for a compute provider, the share of time chips are actually rented decides profit, because the hardware is expensive and loses value fast.
• Sales cycle: enterprise and government deals can take many months from pilot to payment, which stretches the cash a startup needs.
• Revenue concentration: many AI firms depend on a few large customers, so losing one can change the year.
• Time to first revenue: software can earn within months, while chips, rockets and quantum machines may need years of spending first.
• Talent cost: senior researchers and engineers are the biggest expense for many labs and are paid close to global levels.

Rules and regulators

Risks

• Global labs releasing similar features for free inside larger products.
• High computing costs that keep margins thin.
• Slow enterprise and government buying cycles.
• Dependence on imported chips and GPUs.
• Scarce senior talent and poaching by global firms.
• Technical failure in hardware, such as a failed test or launch.
• Data protection duties and unsettled rules on training data and copyright.

What to watch

• Whether Indian foundation models turn into paid products
• GPU availability and the price of training in India
• Second-phase semiconductor projects and first production runs
• DPDP compliance deadlines in November 2026 and May 2027
• Large strategic investors entering AI rounds
• Government contracts for agentic AI platforms and sovereign AI

Inside the sector

• Foundation models & generative AI: A foundation model is a large AI system trained on huge amounts of data that can be adapted to many tasks, such as writing, translating or answering questions. Generative AI is the use of such models to create text, speech, images or code. In India the focus is on local languages, voice and low cost.
• AI dev tooling & agents: AI dev tooling and agents are the products that sit on top of AI models: tools developers use to build with them, and software agents that carry out tasks on their own, such as answering a phone call or writing an app. This is the fastest-moving layer of Indian AI.
• Computer vision & applied AI: Computer vision is AI that understands images and video, and applied AI is the wider practice of using AI on a specific business problem. In India these companies sell to factories, banks, retailers and governments, and they usually earn by saving a customer money or time.
• Robotics, chips & other deep tech: Robotics, chips and other deep tech are startups that build physical or scientific products: robots, semiconductors, drones, rockets, satellites and quantum computers. They need more money and time than software, and government schemes play a large role.

AI & Deep Tech: latest on StopDown

• CoRover wins MeitY government agentic AI platform contract 10 October 2026
• Astrogate Labs ties up with Spacebeam on laser links 9 October 2026
• Inner Sky Labs launches physical AI foundation models 8 October 2026
• Desible.ai raises ₹32 Cr led by Prime Venture 7 October 2026
• Discovr AI launches creator advertising platform 7 October 2026
• Quanfluence raises $10M for photonic quantum computer 7 October 2026
• Sarvam AI names ex-Google Cloud head as president 7 October 2026
• Soket AI releases LOOP agent harness 7 October 2026

Every AI & Deep Tech story →

Most active investors here

1. Peak XV Partners (12 rounds)
2. Accel (8 rounds)
3. Bessemer Venture Partners (8 rounds)
4. HCLTech (8 rounds)
5. Finvolve (7 rounds)
6. Khosla Ventures (7 rounds)
7. Nvidia (7 rounds)
8. India Accelerator (6 rounds)

Rounds StopDown covered in the last 12 months. Activity is not a measure of quality.