India’s AI Opportunity Is Applications, Not Chips

Source: Marcellus Investment Managers · Speakers: Mohandas Pai and Saurabh Mukherjea · Video ID: maFTZNeBkN8 · Published July 22, 2026

India is unlikely to win the next four or five years of AI by manufacturing frontier chips or training trillion-parameter foundation models. Its advantage is more practical: talent density, enterprise implementation skill, domain-specific systems, and the ability to build the application layer where AI finally touches business workflows.

Who Are Mohandas Pai and Saurabh Mukherjea?

Mohandas Pai is a former Infosys CFO, investor, public-policy voice, and one of India’s most visible commentators on technology, education, capital formation, and Bangalore’s role in the global software economy. His strength is the combination of macro numbers and operator memory: power capacity, chip economics, GCC hiring, software services, and the gritty structure of enterprise IT.

Saurabh Mukherjea, founder and CIO of Marcellus Investment Managers, frames the discussion from an investor’s seat: what AI does to Indian technology jobs, enterprise spending, global competition, and portfolio construction. Together, they produce a view that is neither simplistic techno-optimism nor lazy job-loss panic.

AI Has Five Layers, and India Should Not Pretend They Are the Same

Pai starts by separating AI into layers. The first layer is power. AI data centers consume enormous electricity because GPUs do parallel processing at massive scale and require heavy cooling. The United States is spending aggressively on power and grid rejuvenation, but its grid is roughly 50 years old and faces local resistance to new data centers because people worry power costs will rise.

India has a different constraint. Pai estimates installed capacity around 535 gigawatts, with recent peak demand around 270 gigawatts, and says the country is adding roughly 55–60 gigawatts a year, largely solar. The problem is not only generation; it is transmission. Around 25–30 gigawatts of solar power is stranded because the grid has not been built out fast enough. If India wants to host more AI infrastructure, transmission is as strategic as generation.

The second layer is chips, and Pai is blunt: for the next four to five years, India does not have a chance at the frontier. Two-nanometer chips require machines from ASML that cost roughly $400–500 million each, and a leading-edge fab can require $35–40 billion. The real contest is between the United States, Taiwan, and China, with Japan behind and China racing despite lacking the most advanced ASML machines.

That does not mean India is irrelevant in semiconductors. Bangalore, Pai argues, is the largest fabless chip-design location in the world, with around 350,000 chip designers, testers, and embedded software people. Qualcomm, Intel, Nvidia, AMD, and other multinationals already do deep design work there; he says AMD’s two-nanometer chip is being designed in Bangalore. India’s gap is not design talent. It is capital, fabs, owned IP, and the ability to help domestic chip-design startups scale.

The Hyperscaler War Is a US-China Game

The third layer is hyperscale cloud. Microsoft, Meta, AWS, Google, Oracle, and others are committing staggering capital. Pai describes the big five as investing around $700 billion this year, with broader global investment potentially in the trillions over the next few years. To justify that infrastructure, the industry needs hundreds of billions in subscription and token revenue.

That revenue model is still unresolved. ChatGPT and Anthropic have vast usage, but much of it is free. Pai’s quip captures the geography: eyeballs come from India; money comes from the United States. India may grow from roughly two gigawatts of data-center capacity toward seven or eight by 2030 if announced commitments materialize, with Google in Visakhapatnam and Reliance and Adani entering the field. But the global hyperscaler economics will still be set primarily by the US and China.

DeepSeek changed the psychology of that contest. Its release shocked American markets, briefly erasing hundreds of billions in market value from Nvidia and related names. Pai’s geopolitical interpretation is stark: if China invests tens of billions into AI and gives powerful models away as open source, it attacks the American subscription model at the revenue layer. The strategic war is not only about who has the best model; it is about whether the economics of the model layer collapse into a commodity.

Foundation Models Will Become Inputs; Applications Are the Indian Opening

The fourth layer is large language models. Pai does not see India matching the capital intensity of frontier LLM training soon. Leading models have scaled to enormous parameter counts and training costs, while Indian efforts such as Sarvam remain far smaller. But open-source models change the entry point. If Llama-like models and Chinese open-source systems commoditize the base layer, Indian builders can reconfigure them, add domain data, build inference layers, and create useful systems without training from scratch.

The fifth layer—the application layer—is where Pai is most bullish. He says hundreds of young Indian founders are building AI applications for enterprises. At his office, 314 Capital, around a thousand AI-app companies applied for capital; they invested in 10. The technology is promising, but revenue scale remains the test.

This is the most important distinction for founders and investors. India should not confuse national pride with strategic focus. Frontier fabs and trillion-parameter training runs require capital structures India does not yet have. Enterprise AI applications require domain expertise, engineering labor, customer proximity, cost discipline, and integration competence—areas where India has real advantages.

Enterprise AI Will Be Implemented Through Old Systems, Not Around Them

Pai estimates the total installed enterprise software base across the United States, Europe, and Japan at $20–25 trillion. Much of it is old, mission-critical, and deeply intertwined with hardware, security, databases, and operating processes. IBM’s point about American payment systems still running large amounts of COBOL matters because these systems work, scale, and cannot simply be replaced live.

That means enterprises are unlikely to drop a generic AI app into ChatGPT and rewire the company overnight. A large US bank might have 14 databases that do not talk to each other while spending $17 billion a year on software. The frontend may have modern digital layers, but the interior remains complex, fragile, and politically hard to touch.

This is where Indian services firms—TCS, Infosys, and others—retain strategic relevance. They understand how to touch, reconfigure, test, migrate, and support enterprise systems. AI can compress a nine-month project into three or four months, but someone still has to understand the process, write and test code, integrate it, and keep the business running.

The cost model is not obvious. Tokens may become expensive enough that “people plus tokens” costs more than people alone. But India’s offshore billing rates of $50,000–60,000 per person-year compare favorably with US costs of $200,000–250,000 plus tokens. The winning model may be an Indian people-token mix that is cheaper, more operationally grounded, and better suited to enterprise implementation.

Jobs Are in Flux, Not Free Fall

Pai rejects the immediate collapse narrative for Indian tech jobs. India has roughly 6 million people in the software export industry, plus around 1.5 million outside it. Hiring has slowed and utilization has risen from the high 70s to around 85–86%, which means companies hire less in anticipation and keep fewer people on the bench. But retirements, burnout, attrition, and new demand still create openings.

The job impact will be uneven. Freshers will still be hired, though likely in more selective numbers and with higher expectations for AI fluency. The middle layer is more exposed, especially people who became comfortable as document-pushers and stopped coding or learning. AI specialists, by contrast, will command premiums. Pai suggests a six-year AI specialist in Bangalore can earn ₹45–50 lakh, and potentially much more as scarcity increases.

GCCs amplify that scarcity. India has around 2,100 global capability centers, with Bangalore alone hosting about 1,100. Pai estimates GCC employment around 2.3 million and argues they are increasingly doing AI work for global enterprises because talent is scarce in the West. Bangalore’s density is the advantage: 2.6 million software workers, 65,000 IT companies, 22,000 startups, 56 unicorns, and roughly $100 billion in software exports.

His advice to employers is counterintuitive in a layoff-heavy environment: do not fire trained people just because AI improves productivity. Use AI to make them four times more productive, let natural attrition reduce excess where needed, and grow the business. People who can use AI well will be scarce.

Small Language Models and Voice Agents Will Matter Inside Enterprises

Pai aligns with Nandan Nilekani’s view that small language models will matter more for many enterprise deployments than ever-larger generic LLMs. LLMs will become inputs, like sophisticated search engines with code-writing ability. Domain-specific SLMs can combine financial-services knowledge, banking processes, and enterprise-specific data into systems that are easier to embed and control.

He also sees user interfaces moving from online forms to voice. Instead of navigating menus and typing details, a customer should be able to speak in their own language and tone, fill out a form, and finish. Pai names Smallest.ai, founded by Sudarshan Kamat, as an example of fast text-to-voice and voice-to-text work coming out of Bangalore. The point is broader: AI applications will matter when they disappear into workflows people already need to complete.

Investing in an AI World Requires Global Exposure and Indian Conviction

From an investment perspective, Pai’s view is two-sided. Globalization has changed because American S&P 500 companies now capture a large share of global revenue. With roughly 50% or more of S&P revenues coming from outside the United States, US equities have become a proxy for global business. He argues Indian investors should have some global exposure, preferably through smart vehicles rather than trying to pick everything themselves.

At the same time, he is structurally bullish on India. He expects India to become a $10 trillion economy in eight to 10 years, growing on the back of consumption and investment. India is one of only three countries investing more than a trillion dollars annually in capex, alongside China and the United States. Formal jobs are being created, GST collections are rising, banks are profitable, and the top 300–400 million consumers are driving cars, housing, travel, and discretionary spending.

The caveat is quality of income. Many new jobs still pay under ₹25,000 per month; India needs many more jobs above ₹50,000 to broaden consumption. AI can help create higher-productivity roles, but only if the country focuses on skills, enterprise adoption, and application-layer value rather than slogans.

Key Lessons

Why This Matters for Diffie

For Anand and Diffie, Pai’s diagnosis points to a precise market wedge. The opportunity is not to compete at the foundation-model layer. It is to build an application that solves an expensive enterprise workflow better because it understands the workflow, the existing system, and the integration constraints.

Frontend testing sits exactly in that application layer. Large enterprises and high-growth software teams cannot rip out their existing product surfaces, test suites, CI systems, release processes, and design workflows. They need AI that sits on top, understands the messy installed base, and helps teams move faster without breaking production. That is much closer to Pai’s enterprise reconfiguration thesis than to generic AI demos.

The people-plus-token cost frame is also useful for Diffie’s positioning. Buyers will ask whether AI testing reduces total cost or merely adds token spend on top of QA labor. Diffie should show where it compresses a nine-month quality-improvement effort into three or four months, where it reduces repeated manual review, and where deterministic browser evidence prevents costly hallucination. The pitch should be productivity with control, not magic.

The India talent angle matters too. If Bangalore and India become a dense hub of AI-augmented engineering and GCC transformation, Diffie can target teams under pressure to modernize frontend QA while preserving enterprise reliability. The ICP is not “anyone with tests.” It is teams whose product velocity is rising because of AI-generated code, whose UI surface area is growing, and whose existing QA process is too manual to keep up.

Pai’s broader lesson is strategic humility: win where the country, the team, and the product have structural advantage. For Diffie, that means owning the browser-testing application layer, integrating deeply into existing workflows, and becoming the practical AI tool that makes frontend teams measurably faster without pretending the rest of the enterprise stack does not exist.