The Small Model Arbitrage: Why India Should Be Building Vertical LLMs, Not Chasing Frontier
India is trying to build its own GPT-4. This is a mistake.
The capital requirement to train a frontier model is $500M-$1B+. The talent war for ML researchers is won before you enter it—OpenAI, Anthropic, and Google have already hired everyone worth hiring at compensation packages Indian companies can’t match. The compute infrastructure is controlled by three hyperscalers who are also your competitors.
This is not a winnable race. But there’s a different race that is.
The Small Model Arbitrage
I’m calling this the Small Model Arbitrage—the opportunity to capture value by building specialized, vertical-specific LLMs that use local data, languages, and domain expertise where general-purpose models systematically underperform.
The arbitrage exists because frontier model companies optimize for breadth, not depth. GPT-4 is remarkable at general reasoning but mediocre at Tamil legal document analysis, Ayurvedic diagnosis support, or GST compliance automation. The long tail of vertical use cases is economically unattractive to companies spending $1B on training runs.
That’s where the opening is.
A well-executed vertical LLM in a defensible domain will reach profitability faster and generate higher ROI than an Indian frontier model attempt over the next 5 years.
The math supports this. Training a competitive frontier model requires $500M-$1B in compute, 100+ PhD-level researchers at $300K-$500K/year, and 3-5 years to market. Ongoing capital burn to stay competitive as OpenAI and Anthropic release new versions.
A vertical LLM requires $2M-$10M in initial training, 10-20 engineers and domain experts, and 6-12 months to first deployment. The moat is proprietary domain data, not compute scale.
The capital efficiency difference is 50-100x. The time-to-revenue difference is 5-10x.
Why Capital Efficiency Isn’t the Whole Story
Capital efficiency alone doesn’t win. The real arbitrage is in defensibility.
Frontier models are commodity infrastructure. When GPT-5 launches, GPT-4 pricing collapses. When Claude 4 launches, Claude 3.5 becomes table stakes. The moat is constantly eroding because the moat IS the model, and the model is constantly being replaced.
Vertical models have different moats. The moat is the proprietary training data, the domain-specific evaluation benchmarks, the integration into existing workflows, the trust built with regulated industries. These don’t erode when OpenAI ships a new model. They compound.
Consider Indian legal text. A frontier model can summarize a contract. A vertical legal LLM trained on 20 years of Indian case law, Supreme Court judgments, and regulatory filings can identify precedent, flag jurisdictional issues, and generate compliant documentation.
The difference in value is 10x. The difference in defensibility is 100x.
Or healthcare. GPT-4 can answer general medical questions. A model trained on Indian clinical protocols, drug formularies, insurance claim patterns, and regional disease prevalence can assist with diagnosis, treatment planning, and prior authorization. It’s not a better general model—it’s a purpose-built tool that works within the constraints of the Indian healthcare system.
The pattern here is data specificity as competitive advantage. Frontier models are trained on the open web. Vertical models are trained on proprietary, domain-specific corpora that are expensive or impossible for competitors to replicate.
The Import Substitution Mistake
India tried this playbook before. Post-independence industrial policy was built on import substitution—build everything domestically, compete head-to-head with established global players. It failed spectacularly.
India’s inward-looking trade regime discouraged labor-intensive export industries and rewarded installation of new capacity over actual output. The economy stagnated for decades.
The companies that succeeded—Infosys, Wipro, TCS—didn’t try to be IBM. They specialized in specific services where India had comparative advantage: cost-efficient software development, business process outsourcing, IT support. They built world-class competitors by focusing, not by trying to replicate the entire stack.
The Small Model Arbitrage is the same bet. Don’t build Indian GPT-4. Build the best Tamil-English legal LLM. Build the best model for Indian tax code. Build the best clinical decision support system for Indian healthcare protocols.
Who’s Building This
Sarvam AI is building this playbook. They’re not trying to be OpenAI India. They’re building models for Indian languages—starting with Hindi, Tamil, Telugu, Kannada. The training data includes regional dialects, code-switching patterns, and cultural context that frontier models miss. Their Indic LLM performs better on Hindi-English code-mixed text than GPT-4 because it was designed for that specific use case.
Niramai built an AI system for breast cancer screening using thermal imaging. It’s not a general-purpose vision model. It’s a vertical model trained on Indian patient data, optimized for cost-constrained clinical settings, and integrated with existing diagnostic workflows. The model’s accuracy isn’t better than frontier models on general image tasks—it’s better on the one task that matters for their customers.
Tricog built an ECG interpretation model for Indian hospitals. It doesn’t try to be the best general medical AI. It’s trained on Indian cardiac data, accounts for regional disease prevalence, and integrates with existing cardiology workflows. The specificity is the product.
These companies aren’t competing on compute scale. They’re competing on domain depth.
The Three Criteria for Vertical LLM Opportunity
Not every vertical is worth building. The opportunity exists where three conditions hold:
| Criterion | Why It Matters |
|---|---|
| Proprietary data access | The training corpus must be expensive or impossible for competitors to replicate. Public datasets don’t create moats. |
| Measurable performance delta | The vertical model must demonstrably outperform frontier models on domain-specific benchmarks. “Better for India” isn’t enough—quantify it. |
| Willingness to pay | The customer must value the vertical model enough to pay a premium over general-purpose alternatives. Cost savings or compliance requirements work. Marginal convenience doesn’t. |
Indian legal tech meets all three. Case law is proprietary, performance on precedent identification is measurable, and law firms pay for accuracy.
Indian healthcare meets all three. Clinical data is proprietary, diagnostic accuracy is measurable, and hospitals pay for compliance and outcomes.
Indian fintech meets two out of three. Transaction data is proprietary, fraud detection performance is measurable, but willingness to pay is unclear—banks may prefer general models with custom fine-tuning.
The test is simple: if a frontier model company could replicate your vertical model by spending $10M on data acquisition and fine-tuning, you don’t have a moat. If they can’t—because the data doesn’t exist, the domain expertise takes years to build, or the regulatory relationships are non-transferable—you do.
What I Don’t Know Yet
The open question is whether vertical LLMs can sustain pricing power as frontier models improve. If GPT-5 closes 80% of the performance gap on Indian legal text, does the 20% delta justify a 5x price premium?
I think yes, but the answer depends on how regulated and mission-critical the domain is. Healthcare and legal have high switching costs and regulatory lock-in. E-commerce and customer support don’t.
The other unknown is whether vertical models can defend against fine-tuned frontier models. If a customer can take GPT-4, fine-tune it on their own data, and get 90% of the value of your vertical model, your business model collapses.
The defense is proprietary training signal that the customer doesn’t have. If your model is trained on 10 years of aggregated industry data that no single customer possesses, fine-tuning doesn’t replicate it. If your model is just a fine-tuned version of a frontier model on the customer’s own data, you’re a services company, not a product company.
The Civilizational Bet
The broader question is whether India’s AI strategy should prioritize sovereignty or specialization.
Sovereignty argues for building frontier models domestically, even at higher cost, to ensure strategic autonomy. Specialization argues for building vertical models where India has comparative advantage, and relying on global infrastructure for general-purpose AI.
I think specialization wins. Sovereignty in AI is expensive and brittle. The cost to maintain a competitive frontier model is not a one-time investment—it’s an ongoing tax that grows every year as the frontier moves. India’s GDP per capita is $2,500. The U.S. is $76,000. The capital efficiency required to compete on frontier models is not realistic.
But specialization in vertical AI is realistic. India has 22 official languages, 1.4 billion people, and regulatory systems that differ significantly from Western markets. The data specificity is structural, not temporary.
The companies that win will be the ones that stop trying to replicate OpenAI and start building what OpenAI can’t.