Vertical AI for India: where the moats actually are

Two years ago, the venture consensus held that horizontal AI infrastructure was the right place to be. Foundation models, picks-and-shovels infrastructure, GPU compute brokerages — that was where the largest absolute dollars were going, and the heuristic argument was that the application layer would commoditize as soon as the models stabilized. We did not believe that argument then, and the data over 2024 and 2025 has not moved in its favor.

The actual pattern is closer to the opposite. The horizontal layer has commoditized faster than its underwriters expected. Foundation-model performance has compressed; pricing per token has fallen by an order of magnitude per year for the last two years; the meaningful differentiation between OpenAI, Anthropic, Google, and the open-weight tier is now narrow enough that most application builders treat them as substitutable. The application layer has not commoditized in parallel. The companies that compounded fastest in 2024 and 2025 were the ones with proprietary data, regulatory standing, and a cost structure that made workflows economically viable that had been impossible at the previous price points.

The sharper version of the AI bet, in other words, was vertical AI all along. And inside vertical AI, the sharpest version was vertical AI built in India, for the verticals where Indian data, Indian regulatory positioning, or Indian cost structure was the unfair advantage. This essay is about why that thesis is durable — and where, specifically, the moats are turning out to be.

1. AI is no longer a category. It is a substrate.

It is worth being precise about what has changed.

When venture firms first started writing AI checks at scale in 2022 and 2023, "AI company" was still a meaningful descriptor. A startup that built on top of GPT-3 was visibly differentiated from a startup that did not. The technical moat between AI-native applications and incumbents was wide enough that the early movers won real share. That window has now closed for almost every category of horizontal application.

What has replaced it is a world in which AI is the substrate that every software company builds on. The act of writing code, designing UI, generating content, summarizing documents, and answering customer queries has dropped in cost by an order of magnitude. The TAM for software has expanded — workflows that were previously priced out of automation can now be served — but the durable defensibility of being "AI-native" has eroded, because every credible competitor is also AI-native.

What this implies for venture is that the right question is no longer "is this an AI company." Every company will be. The right question is what is the moat on top of the AI — and the candidate moats are a familiar list: proprietary data the model was trained or fine-tuned on, regulatory standing that the customer cannot acquire elsewhere, distribution and trust the company has built into a particular vertical, and a cost structure that the competition cannot economically match.

This is the lens we apply when we underwrite the first pillar of our thesis. The question is never "are they using AI." The question is what is the durable moat on top of it.

2. India produces three of those moats in unusual concentration

Almost every other geography produces one or two of the moats listed above. India produces three of them at structural scale, and that is the reason vertical AI built in India has compounding advantages that vertical AI built elsewhere does not.

Proprietary domain data, geo-locked. Healthcare, fintech, legal, supply chain, and consumer commerce in India generate datasets that no Western incumbent can access. India's healthcare system processes roughly 800 million outpatient visits annually across a hospital network that is largely private and almost entirely Indian-run. India Stack — Aadhaar, UPI, Account Aggregator, ONDC, DigiLocker — has become the largest population-scale digital public infrastructure deployment in the world, with UPI processing more than 17 billion transactions in a single month by late 2025.^[1] Indian commerce data, Indian credit data, Indian language data, Indian mobility data — none of it is meaningfully accessible to OpenAI's training corpus or Anthropic's. A clinical-AI company training on de-identified Indian hospital data, or a legal-AI company built on Indian case law, or a credit-decisioning company built on Account Aggregator-flow data, is not competing with the foundation-model layer. It is building a moat the foundation-model layer cannot trespass on.

Cost arbitrage that changes which workflows are economically viable. India produces approximately 1.5 million engineering graduates annually — more than the United States and China combined.^[2] The fully-loaded cost of a software engineer in India is roughly $10,000-$20,000 per year, against $120,000-$180,000 in the United States and $80,000-$120,000 in Western Europe.^[3] That gap was the historical basis for the IT-services industry. The new application of it is more interesting. A vertical-AI workflow that requires a small human-in-the-loop layer — a contract reviewer, a clinical coder, a claims adjuster, a tier-1 customer service agent — only works at SaaS unit economics if that layer costs 10% of the Western alternative. In every category where the workflow is repetitive enough to be partly automated but not yet good enough to be fully automated, this cost difference is what determines whether the company can clear an 80% gross margin at $50/seat or only at $500/seat. Indian cost structure unlocks workflows that are economically impossible at US cost structure. The TAM expansion is real, and it is not arbitrageable by US-only competitors.

Regulatory and trust positioning, especially with Indian government and Indian enterprise. Selling software into Indian banks, hospitals, utilities, insurers, government bodies, and regulated industries requires India-domiciled vendors, India-resident data, and Indian compliance posture — particularly under DPDP, RBI, IRDAI, and emerging AI regulation. Most foreign vendors cannot meet the requirements without a multi-year buildout. Indian-built vertical-AI vendors, by default, can. The regulatory positioning is asymmetric.

A founder who is building a vertical-AI company that taps even one of these three structural advantages has a credible moat. A founder who is tapping two or three of them — Indian data plus Indian cost structure, or Indian regulation plus Indian language nuance — has a moat that compounds with every additional customer.

3. The three wedges where the moats actually compound

Inside the broad category of "vertical AI for India," not every wedge is created equal. Three sub-categories of opportunity deserve to be highlighted because they are where the moat structure compounds most cleanly.

Indian-data-native verticals. This is the category where the moat is the data, and the AI is the engine that turns the data into product. The canonical examples are healthcare-AI built on Indian clinical and diagnostic data, legal-AI built on Indian case law and statutory drafting, fintech-AI built on Account Aggregator and credit-bureau flows, and agritech-AI built on Indian satellite, soil, and crop-cycle data. Each of these categories shares a property that matters: the data is geo-specific, hard to replicate, and the company that ingests it first develops compounding accuracy advantages.

The angel position in our portfolio that demonstrates this most clearly is Carestack — a dental-PMS-and-clinical-data platform built on Indian dental clinic operations, backed in 2013, that has subsequently raised more than $145 million and is one of the largest US dental-software vendors today. The Indian-data-native moat does not require staying in India. It requires the data foundation to be Indian. Carestack's clinical-data layer was built before the company had to compete in the US market — and that head start is exactly the moat that vertical AI compounds on. We are looking for the next Carestack-shaped position in healthcare, legal, and fintech, where the Indian data foundation is being built today and the AI layer on top of it has not yet been priced.

Cost-arbitrage workflows that were previously priced out. This is the category where the moat is the cost structure, and the AI is what makes the workflow good enough to clear the price point. The canonical examples are mid-market accounting automation, claims processing, contract review for SMBs, customer service in long-tail languages, content moderation, and data-labeling-as-a-service. Each of these categories has a Western incumbent that charges at enterprise unit economics and serves only the largest customers. The Indian-built version, with AI plus a small human-in-the-loop layer, can serve the long tail at a tenth of the price and clear healthy margins.

The portfolio expression we have leaned hardest into here is Runo — a SIM-based mobile call CRM that has reached over $750k ARR with 450-plus paying customers, including Honda and Bajaj Finserv. The wedge is that Runo is the only credible CRM for sales teams that work primarily on mobile phones, which is the dominant sales motion in India and across emerging markets, and the cost structure that makes it work is structurally Indian. Bynry — our utility-software bet — sits in the same family. AI compresses Bynry's deployment timeline from months to weeks and reduces the customer's call-center volume by 60-80%, which is exactly the operating-cost reduction required to make small-utility software economically viable. Whatmore, our shoppable-video position, is the consumer-retail expression of the same pattern: AI generates the content at unit economics that the previous generation of post-production studios could not match, which expands the customer base by an order of magnitude.

Frugal-engineered AI infrastructure for global deployments. This is the category where the moat is in the picks and shovels — observability, model routing, fine-tuning ops, evaluation frameworks, security tooling — built by engineers who grew up optimizing for constraint, and sold into a global customer base that is trying to reduce its compute spend. This is the wedge where the cost-conscious mindset that India has historically had on hardware and on services translates into a pricing discipline that the US-only infrastructure layer is not yet matching.

Our position in Skydda — AI-native cybersecurity for the post-LLM enterprise stack — sits in this category. The buyers are global; the engineering is Indian; the price-performance is fundamentally different from the US-incumbent layer it competes with. Neverinstall, our cloud-liberated computing position, is in the same family. Both companies illustrate that the infrastructure-tooling layer for AI deployments is not yet a winner-takes-all market, and India-built challengers can carve durable market share because their unit economics are structurally different.

4. What the data is now showing

When we wrote the first version of this thesis in early 2024, much of the supporting evidence was forward-looking. Two years on, the data is starting to confirm the shape of the bet.

India added 11 new unicorns in 2025 — bringing the national total to 73 — with AI and deep tech driving the fastest-growing categories. Ai.tech became the fastest unicorn in Indian history, reaching a $1.5 billion valuation in under three years as a rare bootstrapped success. The ASK Private Wealth Hurun India research has documented 105 future unicorns across the gazelle and cheetah categories, with AI and deep tech showing meaningfully higher growth rates than the broader cohort. Netradyne — fleet-AI built on driver-behavior data — became one of the cleanest examples of the pattern. The 2025 data also showed 73 unicorns collectively employing more than 206,000 people, with India ranking third globally in unicorn count behind the United States and China.

Capital is following. PE-VC activity in India rebounded to roughly $43 billion in 2024, with India remaining a top APAC destination. The India Deep Tech Investment Alliance, formed in 2025, has pledged to mobilize over $1 billion specifically into Indian deep tech, with semiconductors, AI, and robotics as the named priorities. Bilateral US-India trade has more than doubled in three years to over $190 billion. The corridor is hardening at exactly the moment the underlying companies are scaling.

The within-portfolio evidence has tracked the same shape. Our angel position in Seekho — backed in 2021 — has been recognized as a Cheetah in the latest Hurun research and is on a path that we underwrote at $5M post and now sits at a markup that puts the position at 80x for the angels who participated. Carestack's expansion to over $40M ARR has played out exactly the way the original 2013 thesis on India-built dental software anticipated. Bynry has signed named municipal customers including the City of San Bernardino and Siloam Springs, with a serviceable obtainable market of $10 billion across roughly ten thousand North American utilities. Runo's paying customer base has crossed 450 with named enterprise logos. Each of these is, in its own segment, validation that the moats we underwrote on day one are compounding.

5. What the bear case is, and why we still hold the position

It would be intellectually lazy to write a cornerstone essay on this category without engaging the bear case directly. There are three.

The first is that horizontal AI eats vertical AI. The argument is that as foundation-model capabilities improve and as agentic frameworks become competent enough to handle entire workflows, vertical-AI applications will be reduced to thin shells over horizontal models, and the moats will collapse. We do not dismiss the argument. The foundation-model layer has improved faster than even the optimists expected. The empirical pattern, however, has not yet matched the prediction. Even as model capabilities have improved, the share of enterprise AI spend going to vertical-application companies has grown, not shrunk. The reason, as best we can tell, is that the moats we are underwriting — data, regulation, distribution, cost structure — are not horizontal capabilities. They are structural assets that the foundation models cannot acquire even when they get cheaper and better. We expect this to remain true for the foreseeable future. We will be watching the data carefully for any sign that it is changing.

The second is that India-specific regulatory and political risk could disrupt the corridor. The argument is that visa policy, data-localization requirements, and political volatility in either the US or India could materially slow the cross-border arc that vertical-AI-built-in-India relies on. This risk is real and it is one we underwrite explicitly. The mitigation is that the structural alignment between the two countries — institutionalized through frameworks like IDTA, TRUST, and RDI — has survived multiple administration changes on both sides. The corridor is durable, even if the path is bumpy.

The third is that capital crowds in. As the thesis becomes more widely shared, more capital will compete for the same positions, deal pricing will rise, and the structural advantage of being a focused fund will erode. This is the strongest of the three bear arguments because it is empirically certain to happen — the only question is the timing. Our response is operational. The reason we anchor the firm in Hurun India research, in the diaspora corridor, in operator depth, and in pre-thesis writing rather than in deal-flow auctions is that we are organizing the firm to source the deals that the auction-driven competitors do not see. As long as we maintain that operational discipline, we expect to keep our edge as the category becomes more competitive.

6. The Fund II implication

We expect five to seven of Fund II's positions to land in this pillar — the largest single conviction in the thesis. The composition will be roughly even between Indian-data-native verticals and global-cost-arbitrage workflows, with one or two infrastructure-layer positions in the picks-and-shovels category.

The screen we apply to every candidate company in this pillar is the same. Is the moat data, regulation, distribution, or cost structure? Does the moat compound with each additional customer, or is it a one-time advantage? Is the founder uncommonly close to the customer, or are they hypothesis-led from a deck? Is the unit economics shape healthy at the small-customer end, or only at the enterprise end? And is the corridor in play — does the company have a credible plan to be dual-geography, with engineering anchored in India and customer base in either India or globally?

If the answer to those questions is yes, we lean in early, write the first check, and spend the next two years working with the founders on customer introductions, follow-on capital, and partnerships. The vertical-AI-for-India thesis is not just where we are putting the money. It is where the firm's operational bench is structurally pointed. The portfolio you can already see in Fund I is the proof of concept. Fund II is where the thesis compounds.

— Vinod Jose, Founding GP

This essay extends the Vertical AI pillar of the Callapina thesis.