Not every AI startups in India are really an AI startups. Many are wrappers, a thin layer of software over a foreign foundation model like GPT, Gemini or Claude, dressed up with an Indian logo. The companies actually building are the ones creating real technology underneath: their own foundation models, GPU infrastructure, proprietary data or hard domain depth.
Author: Aadarsh Patel | EQMint
In 2026 the genuine builders include Sarvam AI and Krutrim on sovereign language models, Neysa on AI compute infrastructure, Fractal on enterprise decision intelligence and Qure.ai on medical imaging. The distinction matters more than ever, because a Google Cloud executive warned in 2026 that thin wrapper startups face extinction as foundation models absorb their features, and investors now treat we use GPT better than anyone as a red flag, not a pitch. This is the honest cut: who is building, who is wrapping, and how to tell the difference.
The wrapper question stopped being academic in India this year. At the February 2026 India AI Summit it became a public scandal, and it exposed how much of the sovereign-AI story is real versus marketing.
Here’s the honest breakdown: what a wrapper actually is, the companies genuinely building, the ones that got caught, and why the difference decides who survives.
Wrapper vs builder, the real distinction
Define it precisely, because the word wrapper is thrown around loosely. A wrapper takes your prompt, quietly sends it to a foreign model like OpenAI’s or Anthropic’s over an API, and returns the answer with a custom interface on top. A builder creates genuine technology of its own: a model it trained, infrastructure it operates, data it owns or a workflow it has embedded so deeply that the model underneath is replaceable.
The honest nuance most hot takes miss. Being a wrapper is not automatically fatal. Some of the world’s most valuable AI products are technically wrappers, built on models someone else trained, yet they thrive because of what they built around the model. The real split is between thin and thick.
| Thin wrapper | Thick builder |
| Prompt plus interface over one API | Proprietary model, data or infrastructure |
| No proprietary data or workflow lock-in | Embedded in a real workflow or dataset |
| Replicable in an afternoon | Takes years and capital to build |
| Dies when the model ships the feature | Survives model updates, often benefits |
The one-line test, borrowed from how investors now think: if a foundation model shipped your exact pitch as a default feature in its next release, would your customers cancel? If yes, you’re a feature, not a company. That question separates the thin wrappers, mostly doomed, from the thick builders and genuine deeptech.
The companies actually building
Take a clear position. A real set of Indian AI startups is doing the hard, capital-heavy work of building genuine technology, not just wrapping. Named to illustrate the category, not as recommendations.
Sarvam AI, foundation models. The clearest builder. Sarvam trains its own large language models optimised for Indian languages, was selected by the government’s IndiaAI Mission and allocated 4,096 NVIDIA H100 GPUs to build a sovereign model, and its work on Indic tokenisation genuinely lowers the compute cost of running Indian-language AI. Reports say its large model beat a leading global model on an OCR benchmark. That’s building, not wrapping.
Krutrim, the full-stack bet. Bhavish Aggarwal’s Krutrim aims at a vertically integrated stack, its own multilingual Krutrim-2 model, cloud infrastructure and even chip ambitions. It’s worth being honest here: Krutrim faced public allegations of being a wrapper after early stumbles, which makes its shift toward genuinely trained models and infrastructure the thing to watch. If it delivers the full stack, that’s a moat no foreign lab can replicate locally.
Neysa, the compute layer. Neysa builds AI cloud and GPU infrastructure, the picks and shovels of the AI gold rush. Infrastructure captured the largest share of Indian AI capital in 2026 precisely because it’s the hardest to build and the most defensible once running. This is deep building by definition.
Fractal and Qure.ai, proprietary depth. Fractal, a decision-intelligence veteran with its own reasoning model and an IPO expected, and Qure.ai, whose medical-imaging AI has screened over 39 million patients across 100-plus countries, both own something a wrapper can’t copy: deep domain expertise and proprietary or hard-won data. Qure.ai’s edge isn’t a clever prompt, it’s years of clinical validation.
The wrapper epidemic, and who got caught
Be blunt, because 2026 made this impossible to ignore. The India AI Summit in February 2026 became a showcase for the gap between claimed and real indigenous AI.
The most cited example: Krutrim’s public chatbot, when asked who created it, reportedly replied that it was developed by OpenAI, fuelling allegations that the consumer product was at that point a thin layer over ChatGPT’s API. The company has since pushed hard toward its own models, but the episode became the symbol of the problem. At the same summit, a robot dog presented as indigenous technology was quickly identified by observers as an off-the-shelf Chinese unit with a new label. These weren’t isolated embarrassments, they crystallised a real worry: that a chunk of India’s sovereign-AI story is marketing over substance.
The honest point is not to mock. It’s that the wrapper epidemic makes the genuine builders more valuable, and makes scrutiny essential. When anyone can slap an Indian logo on a foreign model and call it sovereign AI, the ability to tell real from fake becomes the whole game.
Why thin wrappers are in real danger
This isn’t EQMint’s opinion alone, it’s the emerging consensus. A Google Cloud executive warned in 2026 that thin wrapper and multi-model-aggregator startups face extinction, and the logic is hard to argue with.
As foundation models get smarter and cheaper, they absorb the features wrappers were built on. The nickname for the fear is getting GPT-5’d: the base model ships a native version of your one feature, and your revenue goes to zero overnight. It has already happened repeatedly, when a major model added native document handling, a wave of document-wrapper startups saw their value evaporate. Wrappers also carry brutal economics, thinner margins than normal software because every query pays an inference cost to the model provider, and estimates suggest the large majority of pure wrapper startups never make meaningful revenue. In India specifically, the crowded application layer plus dependence on foreign models makes thin wrappers doubly exposed.
India’s honest position in the global AI race
Take the balanced, honest view rather than the flag-waving one. India is not, in 2026, a foundation-model superpower on par with the US or China, and pretending otherwise is the wrapper problem in national form.
Where India genuinely leads is frugal, constraint-driven building: models optimised for 22 languages and low-end devices, cost-efficient infrastructure and a strong position as a quality infrastructure and evaluation layer for global AI. Some of the most substantive Indian AI work, like post-training and model evaluation for the world’s top labs, is invisible in the hype because its customers are foreign AI companies, not Indian consumers.
The IndiaAI Mission, with roughly 10,000 crore committed and subsidised GPU access, is a deliberate five-year bet to move India from that support role toward genuine foundation-model capability. Whether it succeeds is the open question of the decade.
So the honest framing is neither India will build the next OpenAI by next year nor India only makes wrappers. It’s that India is building real, differentiated AI in the places its constraints give it an edge, while a noisy layer of wrappers rides the same wave claiming more than it has built.
How to tell a builder from a wrapper
A practical filter, useful whether you’re an investor, a founder or just reading the headlines. Ask these of any Indian AI startup claiming to be the real thing.
| Ask this | A builder can answer |
| What did you actually train or build? | Names its model, infra or dataset |
| What do you own that a rival can’t copy? | Proprietary data, workflow or IP |
| What happens if GPT ships your feature? | Has a moat beyond the base model |
| Where does your compute come from? | Owns or operates real infrastructure |
| Who are your customers, really? | Paying enterprises, not just demos |
The honest bottom line. India’s AI story in 2026 is genuinely exciting and genuinely inflated at the same time. A real cohort, Sarvam, Neysa, Fractal, Qure.ai and others, is building hard technology with defensible moats, while a wide layer of thin wrappers claims the same sovereign-AI glory with far less underneath. The distinction is not pedantic, it decides which companies survive the next model release and which get erased by it. Read past the logo to the technology, ask what a company actually built, and the real Indian AI story, smaller, harder and more durable than the hype, comes clearly into focus.
FAQ
What is an AI wrapper startup?
A company whose product is mainly a thin layer of software and a user interface over someone else’s foundation model, like GPT, Gemini or Claude, with little proprietary technology of its own. It sends prompts to the external model and returns the answers.
Which Indian AI startups are actually building real technology?
Genuine builders in 2026 include Sarvam AI and Krutrim on sovereign language models, Neysa on GPU and cloud infrastructure, Fractal on enterprise decision intelligence, and Qure.ai on medical imaging, each with proprietary models, infrastructure or data.
Are all wrapper startups bad?
No. Some highly successful global AI products are technically wrappers but thrive because of proprietary data, workflow lock-in or distribution built around the model. The danger is the thin wrapper, a prompt and interface over one API with no moat, which is easily replicated.
What was the Krutrim wrapper controversy?
Krutrim’s public chatbot reportedly answered that it was developed by OpenAI when asked who created it, fuelling allegations it was a thin layer over ChatGPT at the time. The company has since moved toward its own trained models and infrastructure.
Why are thin AI wrappers considered risky?
As foundation models get smarter and cheaper, they absorb the single feature a wrapper offers, erasing its value overnight, informally called getting GPT-5’d. Wrappers also carry thin margins because every query pays the model provider an inference cost.
Is India a leader in foundation AI models?
Not yet on par with the US or China. India’s genuine strength is frugal, constraint-driven building for many languages and low-end devices, and acting as a quality infrastructure and evaluation layer. The IndiaAI Mission is a five-year bet to build deeper capability.
What is sovereign AI?
AI developed and owned within a country, including the models, infrastructure and data, so that critical capability and data do not depend on foreign providers. For India, with its linguistic diversity and data-localisation needs, it is a strategic priority.
How can I tell a real AI startup from a wrapper?
Ask what it actually trained or built, what it owns that rivals cannot copy, what happens if a foundation model ships its feature, where its compute comes from, and who its paying customers are. A genuine builder can answer these concretely.
EQMint is not a SEBI registered investment adviser. This article is for informational purposes only and is not investment advice, and does not recommend any specific company or investment. Company names are used only to illustrate the landscape, technical claims are based on public reporting that can change, and being described as a builder or wrapper here is an editorial characterisation, not a definitive judgement. Verify current details before relying on them.
For more such information visit EQMint
Join our Whatsapp channel for timely updates: Whatsapp
