Why India's IT Giants Don't Build AI — And What China's Free Models Are Doing to America's $725 Billion Gamble

The Biggest Race You're Not Watching

Somewhere in a server farm outside Beijing, a model trained on a fraction of the budget of its American rivals is quietly outperforming them on benchmark after benchmark. Somewhere in San Francisco, a startup burning through $27 billion a year is nervously watching that. And somewhere in Chennai and Bengaluru, the largest IT companies in India — collectively worth a quarter-trillion dollars — are mostly doing what they have always done: writing code for someone else's vision.

The AI arms race is the defining industrial competition of the 2020s. The US government calls it a national security issue. China has made it a state priority. The European Union has drafted laws about it. And India — the country that supplies more software engineers to the world than any other — is largely watching from the sidelines, billing by the hour.

This is not a story about India failing. It is a story about a structural trap, a business model that made perfect sense for thirty years and now looks increasingly like a slow-motion problem. It is also a story about what happens when China floods the world with powerful free AI models — and what that does to the $690 billion bet that America's biggest companies are making on AI right now.

Buckle up. This one moves fast.

India's $250 Billion Blind Spot

TCS, Infosys, Wipro, and HCL together employ over 1.5 million people. Their combined market capitalisation runs into the hundreds of billions of dollars. They are, by any measure, enormous companies. And their research and development spending is, by any comparable measure, almost invisible.

TCS spends roughly 1.1% of its revenue on R&D. Infosys spends approximately 0.52%. For context, US Big Tech companies average around 15% of revenue on research. Even Chinese AI labs — operating under export restrictions, with limited access to the world's best chips — spend an estimated 8%.

Indian IT's 1% is not a rounding error. It is a strategic choice.

Why Being a "Services Company" Is a Trap

To understand why Indian IT barely spends on research, you have to understand their business model. These are not product companies. They are services companies — which means they get paid for people, not for inventions. A consultant billed at $40 an hour earns money every hour they work. A software product earns money whether its creator is asleep or awake. These are fundamentally different economics.

Services companies have very little incentive to reduce the number of people they bill. Research that automates work is, from a short-term P&L perspective, research that eats your own revenue. This is not stupidity. This is rational. For three decades, it worked spectacularly.

But here is the trap: between FY2020 and FY2025, Indian IT majors paid out 87% of their combined net profits as dividends and share buybacks. That is money returned to shareholders rather than reinvested in building AI capabilities. It signals a mature, cash-generating business — not a company betting its future on a technology shift.

The irony is sharp. The companies best positioned by workforce size and client relationships to lead India into the AI era are instead writing cheques to shareholders while OpenAI and Baidu write models. Shareholders are happy today. The question is whether CEOs will be happy in 2030.

India's Brain Drain Problem

India produces more AI talent than almost any country on Earth. In 2025, India had roughly 50,000 top AI researchers and inventors — second only to the United States, which had 220,000. On LinkedIn, India leads the world in AI skill penetration: AI skills appear in Indian member profiles at three times the global average rate.

And most of those people leave.

The Stanford AI Index, which tracks the net flow of AI talent between countries, gave India a score of -16.9 in its 2025 report. That is the worst brain drain score for any country in the dataset. Canada, which people also complain about losing talent, scored -7.1. India's number is more than double that.

What does -16.9 mean in practice? It means that for every AI researcher who moves to India, nearly seventeen are moving away — to the US, the UK, Canada, the UAE. They are not leaving because they lack ambition. They are leaving because that is where the research jobs, the frontier compute, and the venture capital are.

India is, in effect, an AI talent farm for the rest of the world. The seeds are planted here. The harvest happens elsewhere.

India Is Not Sleeping Entirely

To be fair, India is not completely absent from this race. It just started very late, and it is running with one shoe untied.

The Indian government launched the IndiaAI Mission with an approved budget of ₹10,371 crore — roughly $1.2 billion — spread over five years. The ambition is real: build compute infrastructure, fund research fellowships, create a sovereign AI ecosystem. In Year 1, however, the mission spent only about 31% of its allocated funds. Parliamentary panels have noted that fellowship targets were missed by enormous margins — 150 undergraduate scholars selected against a target of 5,000.

On the private side, there are genuine bright spots. Sarvam AI, which is building language models tuned for Indian languages, raised $234 million in a Series B round in June 2026 — led by HCLTech, with Bessemer Venture Partners, Khosla Ventures, and Peak XV also backing it. That put Sarvam's valuation at $1.5 billion, making it India's first AI unicorn in the true research sense. Krutrim, Ola founder Bhavish Aggarwal's AI venture, has had a rockier path and has struggled to translate early hype into consistent product delivery.

Even within Indian IT's cautious culture, some executives are beginning to say the quiet part out loud. HCLTech CEO C Vijayakumar said in February 2025: "To have a long-term competitive advantage, it makes a lot of sense to build... We need to find ways to very economically create a training infrastructure." That is an unusually direct admission from the corner office of a major IT firm. Whether capital follows that conviction is another question.

The China Bombshell

On January 20, 2025, a Chinese AI lab called DeepSeek released a model called DeepSeek V3. It was not the release that shocked the industry. It was the footnote in the technical report.

DeepSeek claimed the entire model was trained for approximately $5.6 million. American AI labs had been spending hundreds of millions of dollars — sometimes more — to train models of comparable capability. Suddenly, the assumption that cutting-edge AI required essentially unlimited capital looked fragile.

Seven days later, on January 27, 2025, Nvidia's stock fell 17% in a single trading session. Roughly $600 billion in market capitalisation evaporated in one day. The logic was brutal: if you can train a world-class AI model for $5.6 million instead of $500 million, you do not need to buy nearly as many of Nvidia's chips. The whole thesis of "AI requires infinite compute" suddenly had a very large asterisk next to it.

DeepSeek's model also matched or beat major US models on several standard benchmarks — the tests the industry uses to compare reasoning ability, coding skill, and language understanding. The DeepSeek API launched at roughly 90% cheaper than comparable US model APIs. It was not just a technical achievement. It was a pricing shock.

And Then the Floodgates Opened

DeepSeek was the warning shot. What followed was the artillery.

Alibaba released Qwen3, its open-source model family — meaning the model weights (the actual learned parameters, the "brain" of the AI) were made freely downloadable for anyone to use, modify, and build on. Qwen3 has been downloaded over 300 million times. The open-source AI community has built over 100,000 derivative models on top of it on HuggingFace alone. An entire ecosystem of products now runs on a free Chinese model.

Then, in July 2026, Moonshot AI released Kimi K3. It has 2.8 trillion parameters — a parameter being one of the individual numerical weights that determine how a model thinks. To put that in perspective: GPT-3, which launched in 2020 and seemed impossibly large at the time, had 175 billion parameters. Kimi K3 is sixteen times bigger, and it is free to download. It is the largest open-weight model ever released to the public. Anyone with enough server hardware can now run a model that competes with the best closed American systems — for free.

This is not a trend. This is a structural shift in who controls access to frontier AI.

The Trillion-Dollar Question

Here is where this gets genuinely uncomfortable for the US industry.

In 2026, America's largest tech companies are spending at a scale that is almost impossible to comprehend. Amazon has committed $200 billion in AI-related capital expenditure. Microsoft is spending $190 billion. Google is spending $175 to $205 billion. Meta is spending $125 to $145 billion. Add it up and you get somewhere between $690 billion and $725 billion — in a single year.

OpenAI, the company that started this whole public frenzy with ChatGPT, is valued at $852 billion. It is generating around $40 billion in annual recurring revenue. It is also burning through roughly $27 billion a year in compute, salaries, and operations — and is still unprofitable. Anthropic, the rival AI safety company, reached a valuation of $965 billion in its 2026 Series H funding round.

These are staggering bets. And they rest on a simple premise: that companies and consumers will keep paying for AI access, because access to the best models is worth paying for.

But if China keeps releasing free models that are nearly as good — or, in some benchmarks, better — that premise gets harder to defend. The price of using AI via API (a way of accessing the model over the internet, like a subscription service) has already collapsed. LLM prices — the cost of getting an AI model to process and respond to your text — fell roughly 300 times between 2023 and 2026. OpenAI itself cut prices on several of its models by 80%. The commodity pressure is already visible.

If the best AI model in the world costs $0, who pays for the $725 billion buildout?

How America Fights Back

The US industry is not sitting still. And the answer it is converging on is clever: stop selling models, start selling systems.

The move is from raw AI capability to AI-powered workflows — called agents — that do real work inside a company's existing processes. An agent does not just answer a question. It reads your emails, updates your CRM, files your compliance report, and flags the anomaly in your invoice queue, all without being asked twice. That is worth paying for. A raw model that generates text is increasingly a commodity. A trusted system embedded in your business operations is not.

The other moat is data. Retrieval-Augmented Generation — RAG for short — is a technique where an AI model is connected to a company's private documents, databases, and knowledge bases, so it answers questions using real, up-to-date internal information rather than general training data. A free Chinese model cannot access your hospital's patient records or your law firm's case files. A purpose-built enterprise system with RAG pipelines, running on your own infrastructure, can. The RAG market was $1.94 billion in 2025 and is projected to reach $9.86 billion by 2030.

There is also the compliance angle. Regulated industries — healthcare, finance, defence, insurance — cannot simply hand their data to any model, wherever it runs. Certifications like HIPAA (US healthcare data rules), SOC 2 (security audit standards), and FedRAMP (US government cloud standards) take years and millions of dollars to obtain. They are bureaucratic, annoying, and extremely valuable as a moat against cheaper competitors.

Cloud inference revenue — the money cloud providers earn from running AI models for customers — is still growing even as model costs fall. AWS earns over $110 billion a year in cloud revenue. Azure is past $100 billion. The infrastructure layer is holding up. It is the model layer that is getting commoditised.

The Forecast: Who Wins, Who Struggles, Who Reinvents

India — the next three to five years: Indian IT majors will not collapse. Their existing client relationships are durable and their delivery machines are formidable. But they face a structural squeeze. If AI agents can do the work of ten junior developers, and clients start noticing that, headcount-based billing comes under pressure. The companies that pivot fastest — building proprietary AI layers on top of their delivery platforms, moving toward outcome-based pricing, investing in companies like Sarvam — will survive the transition. The ones that treat AI as a productivity tool for their existing model, rather than a reason to reinvent the model, will find margins compressing through the 2020s. India's government AI mission needs a serious reset in execution, not just ambition. The talent drain problem will not solve itself without frontier research opportunities at home.

China — the next three to five years: China is playing a different game with open-source AI, and it is a smart one. By releasing powerful models for free, Chinese labs build global developer ecosystems, reduce Western competitive advantages, and demonstrate capability under export-controlled chip constraints. DeepSeek's efficiency breakthroughs were partly forced by necessity — they could not buy enough Nvidia H100 chips, so they had to find smarter training approaches. That constraint produced an innovation. Expect China to keep releasing frontier open-weight models. The risk is geopolitical: as AI becomes more central to national security, pressure on governments to restrict Chinese AI usage in enterprise and government will grow, particularly in the US, EU, and allied markets. China wins on open-source reach but may face a ceiling on enterprise penetration in regulated Western markets.

The US — the next three to five years: The US hyperscalers will not lose the AI war. But they will likely lose the model-as-product business to commoditisation faster than they expected. The winners will be the ones who lock in enterprise accounts with agents, workflow integrations, and compliance certifications before free models become good enough for most business uses. OpenAI specifically faces an existential tension: it needs to be profitable, but its prices keep falling because the competitive pressure from open-source is relentless. The bet on AGI — Artificial General Intelligence, a hypothetical AI that matches or exceeds human reasoning across all domains — as a premium product is the long-range play. If that arrives and it is unmistakably better than anything open-source, the pricing power returns. If it does not, OpenAI's path to profitability looks very difficult at $27 billion of annual burn.

Don't Confuse These Three Things

Open-source weights are not the same as free inference. When a company releases model weights for free download, that means you can download the model's "brain" and run it yourself. But running a model with 2.8 trillion parameters requires an enormous amount of server hardware — GPUs, memory, cooling, electricity. "Free to download" does not mean "free to use at scale." For most businesses, they still need to pay a cloud provider to run these models. The cost has dropped dramatically, but it has not reached zero.

DeepSeek the company is not the same as DeepSeek the open model. DeepSeek is a Chinese company that offers a paid API service — you pay to use their hosted models, similar to paying OpenAI. But DeepSeek also released the weights of its models publicly, so anyone can download and run them independently, without ever touching DeepSeek's servers. When people say "DeepSeek is free," they usually mean the open weights, not the company's commercial service. This distinction matters for anyone making decisions about data privacy or geopolitical risk.

Indian IT "using AI" is not the same as "researching AI." When TCS or Infosys announces an AI initiative or a partnership with Microsoft Copilot, they are using someone else's AI to improve their own services. That is sensible. But it is categorically different from funding the research that creates the underlying models. Using a tool and building a tool involve entirely different capital structures, talent profiles, and strategic outcomes. Conflating the two leads to a dangerously comfortable misreading of India's actual position in this race.

The Bottom Line

The AI race is not being run between companies. It is being run between economic models. The US bet is on premium, integrated, enterprise-grade AI systems that are so embedded in workflows, data pipelines, and compliance frameworks that price competition from open-source cannot dislodge them. The China bet is on a global open-source ecosystem so widely adopted that Chinese AI infrastructure becomes the foundation of how the world builds software. India, right now, is not running either race — it is watching from the stands, billing by the hour, and hoping the game does not change too fast.

The facts, however, suggest the game is changing very fast. A $5.6 million training run that rivalled hundred-million-dollar American models. A 2.8 trillion parameter model, free to download. API prices that fell 300 times in three years. An Indian brain drain score that is the worst on the planet. And a government mission that spent less than a third of its Year 1 budget. These are not isolated data points. They form a pattern.

India's IT sector has survived and thrived through every previous technology wave by adapting its delivery model. The Y2K boom, the internet era, the cloud migration — each one threatened disruption and each one became an opportunity for Indian services firms. The AI wave is different in one important way: it threatens not just what Indian IT delivers, but how many people it takes to deliver it. That is not a client problem. That is an existential business model question.

The companies — and the country — that recognise this earliest and invest accordingly will be the ones writing history. The ones that wait for the wave to fully arrive before reacting will find themselves, as the old saying goes, very busy mopping the floor.