(The Weekend Insight) - The AI Wrapper Gold Rush in India

In today’s deep-dive, we will look at India’s growing crop of AI startups and ask a simple question: how many are building a real business, and how many are simply sitting on top of someone else’s model? We will look at companies that own technology, data, workflows or distribution, and also those that may struggle as OpenAI, Google, Anthropic and others keep improving. The idea is not to dismiss every wrapper, but to understand what is actually left when the underlying AI becomes cheap and widely available.

There is a useful story from Bengaluru that explains the problem better than any definition of an “AI wrapper”.

NeuroPixel was founded before the current AI rush. The company was working on a real problem in fashion ecommerce. Instead of organising an expensive photoshoot every time a brand launched a new shirt, dress or pair of jeans, NeuroPixel could generate images of models wearing those clothes.

This was not a founder putting a new screen on top of ChatGPT. The company had built its own technology and had spent years working on the problem.

Then Google and other large AI companies improved their image-generation models.

NeuroPixel founder Arvind Venugopal Nair later described what happened in unusually simple words. They had been “massively outgunned overnight”.

By April 2026, NeuroPixel had shut down.

That one company tells us a lot about what is happening in Indian AI right now. Building your own AI does not guarantee survival. Using somebody else’s AI does not automatically mean you have a bad company either. What matters is whether you own something that becomes stronger as AI improves, or whether better AI removes the reason for your company to exist.

That is the real question behind the current AI startup rush in India.

Indian AI startups raised around $676 million across 57 funding rounds in the first half of 2026, more than four times the amount raised in the same period a year earlier. This happened while overall startup funding in India was falling.

At the same time, the first batch of failures is becoming visible. Tracxn data cited by Financial Express showed that 18 Indian AI startups had shut down in roughly 16 months. NeuroPixel was one. Dotagent, Zeda.io, LEGOAI and MoneyyAI were among the others.

There are now more than 1,500 Indian startups using AI in some form. Inc42 has tracked more than 170 companies that can be called native AI startups. What is more interesting is where the money has gone. Around 86% of the funding received by Indian AI startups since 2020 has gone into the application side of the market.

That means most founders are not trying to build the next OpenAI or Gemini. They are taking existing AI models and using them for sales, marketing, coding, customer service, medical diagnosis, accounting, video creation, voice calls, presentations, recruitment and dozens of other jobs.

There is nothing wrong with that.

The trouble starts when the whole company is little more than someone else’s model, a decent interface and some marketing.

A wrapper is not automatically a bad business

The word “wrapper” is being used too casually.

If a startup uses GPT, Claude or Gemini, people often dismiss it as a wrapper and assume it has no future.

By that logic, thousands of successful software companies should never have existed. They used Amazon Web Services instead of building data centres. They used Stripe instead of building payment systems. They used databases built by other companies.

A startup does not need to build everything underneath the product.

The better question is what it has built around it.

Suppose OpenAI makes its models twice as good next month and cuts the price by half. What happens to the startup?

For some companies, that is wonderful news. Their product becomes better and cheaper. For others, it is terrible news because OpenAI has now made most of their product unnecessary.

That difference is more useful than arguing about who is or is not a wrapper.

Take Nanonets.

The company works on document and finance automation. A business may receive thousands of invoices, purchase orders, receipts and other documents every month. Nanonets helps read those documents, extract information, push it into accounting systems, run approval rules and handle exceptions.

The AI model is only one part of the job.

If a better model arrives tomorrow, Nanonets can use it. The customer still has years of documents, accounting integrations, approval processes and staff habits built around the product.

The customer is not paying merely to “use AI”. It is paying to get invoices processed. That is a much healthier place to be.

Now compare that with a startup whose main product is “upload a PDF and ask questions about it”.

ChatGPT can already do that. Claude can do it. Gemini can do it. The product can disappear almost overnight.

Qure.ai is what a strong AI company looks like

Qure.ai gives us one of the clearest Indian examples of a company that has built something difficult to copy.

The company uses AI to read medical images and help doctors detect problems such as tuberculosis, lung cancer and stroke.

During FY25, Qure said its technology reached more than 32 million people across over 4,500 healthcare facilities in more than 100 countries. The company has raised around $125 million. Reuters reported that revenue had been growing at roughly 60% to 70% a year and that the company wants to reach profitability before eventually going public.

The important part is what happens inside a hospital.

A hospital cannot simply wake up on Monday morning, cancel Qure.ai and tell its doctors to start using a general chatbot to diagnose chest X-rays.

Medical software has to be tested. It needs approvals. Doctors need confidence in it. Hospital systems need to be connected. Someone has to be responsible when the software makes a mistake.

Qure has spent years building those things.

That is much harder to replace than a prompt.

If Gemini becomes dramatically smarter, Qure may actually benefit because better models can improve parts of its own product. The company does not disappear simply because Google has built a better general AI model.

That is why we would put Qure.ai firmly in the group of Indian AI companies with a strong chance of surviving and growing.

Gnani.ai has something even the large companies cannot recreate quickly

Voice AI is becoming crowded.

Every few weeks there seems to be another startup promising an AI agent that can call customers, answer questions, recover loans, sell insurance or handle support.

A lot of these companies will probably struggle because voice models from OpenAI, Google and others are becoming very good.

Gnani.ai is in a different position.

The company has been working on speech since 2016. It says its systems have been trained on millions of hours of real telephone conversations and that it now handles roughly 30 million interactions a day for more than 200 enterprise customers.

The valuable part is the data.

Indian phone calls are not clean recordings made in a studio. A customer may start in Hindi and move into English halfway through the sentence. Someone from Tamil Nadu may speak English very differently from someone in Punjab. There may be traffic noise, bad network quality, people speaking over each other and all kinds of informal language.

Years of these calls create an asset.

A new startup can buy an API tomorrow. It cannot buy ten years of difficult Indian phone conversations tomorrow.

That gives Gnani a much stronger position than the average voice agent startup.

Sarvam is real AI, but the business still has to catch up

Sarvam AI is another company that clearly belongs on the serious side of the market.

It is building Indian-language models, speech systems and other parts of the AI stack rather than simply relying on another company's model.

In June 2026, HCLTech agreed to invest around $151 million for a 10.5% stake. Sarvam's Series B first close was around $234 million and valued the company at roughly $1.5 billion.

Those are very large numbers. The revenue is still much smaller.

Sarvam reported around ₹29 crore of revenue in FY25 and a loss of roughly ₹102 crore.

This does not mean the company is in trouble. Training models and building this kind of technology requires heavy spending before revenue catches up.

But it does tell us not to confuse “real AI” with “proven business”.

Sarvam now has to show where the money comes from.

Saying that India needs its own AI models is a good national argument. It is not enough by itself to build a large private company.

Google, Meta and OpenAI will keep improving Indian languages. Open models will improve too. Running these models will become cheaper.

Sarvam therefore needs government contracts, enterprise customers, regulated uses and long-term deployments where Indian data and Indian control really matter.

HCLTech may turn out to be more important than another technical improvement because HCL already has access to large enterprise customers.

Sarvam has built real technology. Now it has to build a real business around it.

Krutrim shows how quickly the original plan can change

Krutrim began with a huge ambition.

It wanted to build Indian AI models, a consumer assistant, cloud infrastructure and eventually chips.

It became India's first GenAI unicorn after raising $50 million at a $1 billion valuation. Then the plan changed.

Its consumer assistant Kruti was removed from app stores. Chip plans were paused. The company began putting much more attention on AI cloud services.

That shift makes sense.

A general consumer assistant has to compete with ChatGPT, Gemini and Claude. Those companies have larger research budgets, much larger global user bases and the ability to improve their products every few weeks.

Cloud infrastructure is different.

Indian companies need computing capacity whether they use OpenAI, Sarvam, Meta or another model.

Krutrim has said FY26 revenue reached around ₹300 crore and that it became profitable. The important question is how much of that revenue comes from outside customers rather than other Ola companies. Earlier reporting suggested a large share of revenue came from within the group.

That matters because internal revenue can help a business start, but it does not prove that outside customers are choosing the product.

Still, the change in direction is telling. The less glamorous business may be the better business.

Emergent gives us the opposite lesson

Emergent is important because it shows why calling every company a wrapper can become lazy analysis.

The company lets people describe the software they want and then builds it.

It relies heavily on outside AI models. In a technical sense, someone can certainly call it a wrapper.

But by July 2026, Emergent said it had more than 200,000 paying customers and had crossed a $120 million annual revenue run-rate. It raised $130 million at a valuation of around $1.5 billion.

At that point, the argument changes.

Those customers already have access to Claude, ChatGPT and Gemini.

They are still paying Emergent because they do not want an AI model. They want software that works.

The company handles coding, testing, debugging, hosting and deployment. The underlying model is one part of the product.

Could Anthropic or OpenAI eventually move into the same area? Of course.

That is Emergent's biggest danger.

But 200,000 paying customers are also an asset. Distribution becomes important.

This is exactly what NeuroPixel's founder was talking about when he said the market had moved towards distribution.

NeuroPixel owned technology but did not control enough customers. Emergent depends more on other people's technology but controls a large customer base.

Today, Emergent is clearly the stronger business.

Haptik and Yellow.ai have to prove that years of enterprise work still matter

Conversational AI used to be difficult to build.

Today, a capable developer can combine an AI model, a voice system, company data and a few software connections surprisingly quickly.

That puts pressure on companies such as Haptik and Yellow.ai.

But these companies have something a new startup does not have.

They have spent years selling to enterprises, connecting with internal systems, going through security reviews, handling large customer volumes and fixing thousands of boring operational problems.

Haptik also has the advantage of being part of the Jio ecosystem. It reported around ₹157 crore of revenue in FY24 and a profit of roughly ₹11 crore.

Yellow.ai says it serves more than 650 enterprises across more than 80 countries. Their future depends on whether those customer relationships are sticky.

If their only advantage is that their chatbot is a little better, we would worry.

If removing them means changing customer-support processes across dozens of systems, replacing integrations and retraining staff, their position is much better.

That is where their real value now sits.

The danger gets much higher in content, marketing and presentation tools

This is the part of the Indian market where we would be much more cautious.

Consider Scalenut.

The company works in AI content and SEO. It raised around $3.5 million. FY25 revenue was roughly ₹7 crore and grew only about 8%.

The problem is obvious.

ChatGPT writes articles. Claude writes articles. Gemini writes articles.

Ahrefs can add AI. Semrush can add AI. HubSpot can add AI.

The original reason for paying a separate AI content company has become weaker.

Scalenut is moving into areas such as optimisation for AI search, which is sensible. But it still has to answer a hard question: why should a customer keep paying for another tool when the software it already uses keeps adding similar features?

I am not saying Scalenut will shut down.

A ₹7 crore software business can survive for years if expenses are under control.

I am saying it is much harder to see how this becomes a very large venture-backed company unless the product moves much deeper into the customer's work.

Contlo has a similar problem.

It has raised around $14 million. FY25 revenue was about ₹20 crore and growth was only around 5%.

That would make me nervous.

The company sits in marketing automation and AI agents, where Salesforce, HubSpot, Klaviyo, Braze and dozens of other companies are all adding similar features.

A startup that has raised this much money needs more than a useful feature.

It needs a reason for customers to stay for years.

At the moment, we would put Contlo in the higher-risk group unless the company can show stronger customer retention, faster growth or a part of the product that the larger platforms cannot easily copy.

Coding startups may face the fastest shakeout

Coding is moving faster than almost any other AI market.

Two years ago, turning a Figma design into React code looked like a full product.

Today, coding agents can increasingly understand the design, write the frontend, create the backend, test the code and deploy the application.

That change is brutal for narrow coding tools.

CodeParrot is an early example. The YC-backed startup raised around $500,000 and worked on converting Figma designs and screenshots into code. It eventually shut down after failing to build meaningful recurring revenue.

Kombai and DhiWise deserve close attention for the same reason.

Kombai has done real technical work. It is not a fake company.

The problem is that the market around it is moving very quickly.

If Claude Code, OpenAI's coding tools, Gemini and companies such as Emergent can increasingly build an entire application, a company focused mainly on one part of the coding process has to keep expanding what it does.

Otherwise its product becomes one button inside somebody else's much bigger product.

That does not mean Kombai will fail.

It does mean the company cannot afford to stand still.

Presentations, social posts and simple video tools worry for the same reason

A similar problem exists with startups that generate presentations, social-media posts, advertisements and basic videos.

Take Predis.ai.

Its product helps create social content and advertising material. Technically, this is a thin part of the market. Canva can do more of this every year. Adobe is adding AI. Meta is building creative tools. OpenAI and Google can generate images and text.

Yet Predis did something sensible when the market changed.

Instead of continuing to raise large amounts of money, the company focused on becoming profitable. Investor Gagan Goyal said the business moved from around $100,000 of annual recurring revenue to roughly $400,000 within four months during that period.

That changes the calculation.

If Predis had raised $50 million and hired 300 people, we would worry much more.

A small profitable software company can live comfortably even if it never becomes a unicorn.

That distinction is often missing from startup discussions.

A wrapper can be a good business. It can still be a bad venture investment.

Funding can make a weak AI company more fragile

This is one of the stranger parts of the current boom.

More funding is usually treated as proof that a startup is doing well.

Sometimes it makes the problem worse.

Imagine two AI companies.

The first company raises $2 million, reaches ₹15 crore of revenue, keeps a small team and becomes profitable.

The second company raises $30 million, reaches the same ₹15 crore of revenue and employs 200 people.

The first company can continue happily.

The second has to grow extremely fast because its investors need a much larger outcome.

It must spend more on sales. It must enter new markets. It must probably raise again.

If the product becomes easier to copy during that period, the company gets trapped.

This is likely to happen to a number of Indian AI startups.

The technology may work. Customers may even like it.

The problem is that the company raised money for a $500 million future while building a business that may naturally be worth $30 million.

Most new GenAI startups are wrappers

There is no proper public database showing which model every Indian startup uses, so giving an exact number would be dishonest.

But looking at where funding has gone and what companies are building, we estimate that perhaps 70% to 80% of newer Indian GenAI application startups depend heavily on models built by somebody else.

That does not bother much.

The more worrying number is the share that has very little left beyond those models. We think that perhaps 30% to 40% of the current application group falls into that category.

Those are the startups where the product can be swallowed by a larger platform very quickly.

A basic meeting summariser can disappear into Google Meet or Microsoft Teams. A simple sales-email generator can disappear into Salesforce or HubSpot. A presentation generator can disappear into PowerPoint or Google Slides. A PDF chatbot can disappear into ChatGPT. A basic customer-support bot can disappear into Salesforce.

A Figma-to-code product can disappear into a full coding agent. A social-post generator can disappear into Canva, Adobe or the social platform itself.

Some of these companies will survive by moving deeper into the customer's work. Some will become small profitable businesses. Some will get acquired.

A meaningful number will probably disappear. We have already started seeing that happen.

The Indian AI companies we would feel most comfortable about

If we had to divide the market today, Qure.ai and Gnani.ai would sit near the top for me because both have something difficult to recreate. Qure has clinical data, approvals and hospital use. Gnani has years of difficult speech data and enterprise deployment.

Nanonets also looks much healthier than the average GenAI company because it sits inside a real business process rather than selling a clever output.

Neysa is another serious business, although for completely different reasons. It sells computing infrastructure, so its risks are hardware costs, utilisation and competition from large cloud providers rather than a simple model replacing the product.

Sarvam belongs in the serious technology group as well, but its commercial numbers still need to catch up with its valuation.

Emergent, Haptik and Yellow.ai are more dependent on outside models, but they have meaningful distribution and customer relationships. That gives them a better chance than many companies that may own more technology but have fewer customers.

The companies we would watch much more carefully include Scalenut, Contlo, Kombai, DhiWise, simple presentation generators, basic AI marketing tools and narrow coding products.

Again, that does not mean those companies are about to close.

It means the ground under their products is moving quickly.

The next winners may stop calling themselves AI companies

The strongest businesses in this market will probably become less interested in the label over time.

Qure.ai will be judged on whether it helps diagnose patients.

Nanonets will be judged on whether companies process invoices faster and with fewer people.

Gnani will be judged on whether a bank can handle millions of customer calls at lower cost without making customers angry.

Emergent will be judged on whether a business owner can build useful software without hiring a conventional software team.

Nobody will care very much which model is running underneath.

That is probably the clearest sign that a company has built something useful.

The weaker companies will remain dependent on the excitement around AI itself because there is not enough business underneath once that excitement disappears.

That gives us a simple way to look at the current rush.

Take the website of any Indian AI startup.

Remove the words “AI-powered”. Remove GPT, Claude and Gemini. Then ask what is left.

For Qure.ai, a lot is left.

For Gnani, a lot is left.

For Nanonets, a lot is left.

For Emergent, 200,000 paying customers and a working product remain.

For a large number of newer AI startups, the answer is much less convincing. That is the part of India's AI boom most likely to be tested over the next two or three years.