Search for “AI-powered” on any startup pitch deck in 2025, and you’ll find it right there, in bold letters, anchoring a product that looks suspiciously like a glorified chatbot or a redesigned dashboard. Welcome to the “AI for Everything” era. Slick demos, inflated valuations, and GPT prompts disguised as proprietary tech.
But here’s the real question no one wants to ask in boardrooms: Are we funding innovation or just integrations?
The Wrappers Are Winning. But Is That Sustainable?
In the last 18 months, India has seen a surge in “AI startups.” From HR tech to legal SaaS, agri-advisory to marketing copy tools, everyone’s building “AI-powered platforms.”
But peel back the layers, and many of these companies are simply wrapping OpenAI or Google Vertex APIs with custom UX and spinning them as IP- a few fine-tuned models here, a plug-and-play Whisper API there, and voila: VC gold.
Why is this happening?
Because it’s fast, flashy, and fundable.
Investors love narratives that sell, and nothing sells like AI. Especially post-ChatGPT, when the world collectively realized that an LLM could simulate intelligence in ways we hadn’t seen before. But what we’re often seeing in India is not foundational AI, it’s feature mimicry, with a startup layer built over someone else’s model.
And the problem with wrappers?
They break. They burn. And worst of all, they don’t last.
Tokenized Trends, Thin Moats
Let’s be blunt. The real value in AI isn’t in calling someone else’s API. It’s in:
- Owning the model (or at least customizing it with proprietary data).
- Having differentiated data pipelines that improve over time.
- Building infra that optimizes or orchestrates model performance.
- Creating vertical-specific tools that are indispensable and not just impressive.
Most Indian startups aren’t doing this. Instead, what we have is a rush of “ChatGPT for HR,” “AI for resume screening,” “Bard for school teachers,” and “LLM-powered exam prep.” All useful, sure. But not defensible.
Take a look at what’s raising money:
- ZuAI, an edtech tool, raised $484k for an AI tutor powered by GPT-4.
- A legal-tech tool offers AI-generated case summaries, with no model training of its own.
- Multiple copywriting tools built on open-source LLMs claim “proprietary tone tech.”
Do they solve a user problem? Maybe.
Do they have product-market fit? Possibly.
Do they have a moat? Absolutely not.
Because the moment OpenAI, Google, or Anthropic offer a native plugin or make the same feature freely available, these startups become redundant overnight.
Let’s Talk Burn
Here’s where things get spicy.
The average “AI startup” in India today burns like a late-stage D2C brand circa 2021. Why?
- API costs are high (especially for image, video, and multimodal tools).
- GPU compute is still bottlenecked and prohibitively expensive.
- Talent is scarce; India has fewer than 500 serious AI researchers contributing to major papers.
- And hosting inference models on-prem or on edge? Don’t even ask.
What this means is that most startups are bleeding margins to keep the “AI” alive. As token costs rose and free usage caps tightened, their unit economics tanked. The same story is repeating across sectors: buzz first, margins later.
Where Real AI-Native Moats Exist
To be clear, not all hope is lost. Some startups are building for the long game.
1. AI-Native Infra
Startups like Sarvam.AI are working on Indic-language foundation models, with deep tech, original research, and model optimization. That’s a moat.
2. Verticalized + Proprietary Data
Think Karya focuses on democratising earning and learning opportunities to low-income communities. Karya is powering the next generation of inclusive AI, by creating multimodal datasets in Indic languages, conducting human-in-the-loop tasks, and ensuring culturally aware evaluations to reduce bias in large language models.
3. AI Ops and Tooling
Portkey.ai, a California-based startup founded by Rohit Agarwal and Ayush Garg, doesn’t build LLMs, but offers a plug-and-play backend for AI apps (prompt routing, observability, failover). This infra-layer play is less glamorous but deeply needed.
These are the kinds of companies that won’t be killed off by a single API update. Because they’re not betting on hype, but they’re building the plumbing.
The Investor Blind Spot
VCs chasing “AI-for-X” plays often overestimate first-mover advantage and underestimate infra fragility. But building on borrowed intelligence is like building a castle on sand.
And yet, in pitch decks across India, the AI slide is often the least thought-through.
One token count. One chat UI. One LLM integration. Done.
This leads to:
- Overfunding of shallow plays.
- Underfunding of infrastructure-layer startups.
- Fewer exit-worthy companies in 3–5 years.
So, What Should Be Funded Instead?
- Data network effects: Startups that collect, clean, and annotate niche datasets, especially in Bharat or non-English contexts.
- Custom model training: AI that improves as it learns from user interaction (not just outputs a response).
- AI governance + compliance tooling: With India’s DPDP Act now active, there’s a major gap in monitoring model behaviour and consent mechanisms.
- LLM ops tooling: Model versioning, prompt feedback loops, hallucination flags, audit trails, and none of this is standardised yet.
Closing Thought: Innovation vs. Integration
Yes, it’s easier (and faster) to build an app that wraps GPT-4 and calls it AI.
But India doesn’t need another 500 AI chatbots.
We need a foundational ecosystem:
- Trained researchers.
- Accessible GPUs.
- Public-private data partnerships.
- And serious funding for infra, not just interfaces.
The gold rush is on, but if everyone’s digging for API gold, we might miss the real treasure: building AI that actually works for India, in India, and by India.
