The massive excitement- some might call it hype- surrounding Generative AI, which began with the arrival of OpenAI’s ChatGPT in late 2022, continues to unfold. There have been a few recent dips, first with the announcement of competition from China’s Deepseek and then with the general decline of the NASDAQ and the Magnificent Seven tech companies, even prior to the tariff tantrums.
The market valuations remain enormous. OpenAI was valued at $300 billion in a funding round, while xAI (Elon Musk’s company that has productised Grok using real-time data from X, (formerly Twitter) is valued at $80 billion.
On the other hand, there are questions about intellectual property: ChatGPT reproduced the trademark look and feel of the Japanese animation studio Ghibli, with no clarity regarding whether a license for the IPR was obtained. There is also a sinister outcome: the photographs you Ghibli-ize become the property of OpenAI.
There are three broad and interesting questions: first, whether we are witnessing a genuine, life-changing innovation as dramatic as the arrival of electricity; second, where the significant returns on investment will come from; and third, what India’s current and future roles may entail, especially in light of the recent national AI mission announcements.
- Is Generative AI a Truly Disruptive Innovation?
First, let’s discuss the nomenclature. Traditional AI and Machine Learning are now referred to as predictive AI. This approach utilises vast amounts of numerical data to identify patterns. Significant advancements in recent years, such as AlphaFold, have originated from this field. It examines historical data to predict future outcomes or trends. Statistical models and machine learning algorithms predict events like customer behaviour, market trends, or equipment failures.
Generative AI, on the other hand, focuses on creating new content or data, such as images, text, music, or even software code. It is designed to produce novel outputs based on patterns learned from existing data, primarily unstructured data like text.
Predictive AI has yielded valuable results, enhancing everything from retail inventory planning to more precise X-ray interpretations. The challenge with generative AI is that it has yet to produce a compelling enterprise use case.
Currently, there is no clear use case for B2C Generative AI either. While generative AI is likely to soon become as common as email and video conferencing, few people would be willing to pay for these products. For the most part, major vendors are using generative AI to “enhance the user experience.” Having accomplished this, they are quite willing to release their code to the public at large. There is a potential use case for software companies to assist enterprises in improving their internal processes with AI; these companies would function like consultants but with a tangible impact on operations.
There is a belief that AI-based “agents” might revolutionise workflows and enterprise computing. Similarly, the increasingly popular “vibe-coding” may enable non-technical users to generate software using simple English prompts. All of this remains to be seen, and despite enthusiastic announcements, a decisive use case is elusive. Therefore, at this moment, generative AI is not an earth-shaking innovation like electricity or the Internet.
- What is Behind the Meteoric Rise of Generative AI?
Several factors have established the basis for its popularity. In addition to those mentioned below, there is the ongoing pursuit of an “economic moat,” the development of a hype cycle as a regular aspect of the technology industry, and the use of “standards” as a competitive weapon.
Technical Breakthroughs that Enabled a Better User Interface
Since ChatGPT first appeared in late 2023, the rapid uptake of Generative AI was primarily attributed to its excellent user experience. Additionally, the AI generated responses to questions quickly and with impressive confidence, even though users knew these responses were statistical rather than deterministic and could be prone to errors (hallucinations).
The example of Eliza, an early AI chatbot from the 1960s, is instructive. Eliza functioned like a psychotherapist—rephrasing what the user said into questions or prompts to encourage further elaboration. This method created a surprisingly conversational experience despite its simplicity. People weren’t merely interacting with a program—they were filling in the gaps with their humanity, making it feel personal and responsive.
Eliza also focused on the conversation, enhancing the illusion of a one-on-one exchange with a thoughtful listener. When its internal code did not allow it to generate a sensible answer, it simply responded with “Tell me more about …” (the last topic). It didn’t inundate users with options or technical jargon—it just “listened” and responded, which felt intuitive and natural.
Much the same is true of today’s chatbots, which give the (mistaken) impression that a profoundly empathetic person is at the other end of the conversation. This anthropomorphisation, unfortunately, has sometimes led to addictive behaviour, resulting in depression, mental illness, and even suicide.
Cinema thrives on viewers’ “willing suspension of disbelief.” Similarly, generative AI burst onto the scene with believable answers to common questions, and unsurprisingly, it became the technology with the fastest adoption rate ever.
The Gold Rush Paradox
During the 1849 Gold Rush, miners flocked to California, dreaming of striking it rich. Still, the reality was harsh—most barely broke even, while real wealth accumulated in the hands of intermediaries. These included merchants selling picks, shovels, pans, jeans, and provisions and those who built shanty towns and ran nightclubs. They thrived because they provided the essential tools and infrastructure that every miner needed, regardless of whether those miners found gold. The demand was predictable and widespread, and intermediaries didn’t bear the same risks as the prospectors digging for an uncertain payoff.
A similar dynamic is unfolding with generative AI today. Companies developing and selling enterprise AI solutions—such as those providing custom chatbots, content generators, or industry-specific AI platforms—are akin to the miners. They’re pursuing the “gold” of widespread adoption and transformative use cases, but their success is far from assured. Developing AI models is expensive, competitive, and risky; it requires significant investment in talent, data, and computing power, and the payoff relies on market acceptance and differentiation in a crowded landscape.
In the meantime, chip makers like Nvidia and cloud computing giants such as Amazon (AWS), Microsoft (Azure), and Google (GCP) serve as modern intermediaries. Every AI company, from startups to tech titans, depends on these tools. For example, Nvidia’s chips represent the gold standard for training large language models (LLMs). Similarly, cloud providers supply the storage, networking, and computing resources necessary to make AI development and deployment feasible at scale.
The intermediaries thrive because their products are essential, and their revenue streams are more stable. In contrast, enterprise AI companies confront intense competition and unpredictable margins. They are counting on providing value to end users, but they’re frequently just one innovation or pricing battle away from being undercut. The intermediaries, with established positions and extensive customer bases, capture most of the value without facing the same risks. Just as Levi Strauss created a denim empire while many miners failed, Nvidia and the cloud giants earn billions while AI startups struggle for survival.
Microsoft’s Strategic Vision, Reinvention, and Competitive Skills
Microsoft’s strategic vision has been pivotal in expanding generative AI. By leveraging its partnership with OpenAI, it created an entirely new market. Consequently, Microsoft is the only company in the top 10 in market capitalisation in 2001 and 2025.
In the earlier era, it utilised its co-ownership of the dominant Wintel franchise to establish this position. Then, transitioning from desktop to cloud computing, it positioned itself among the top three cloud computing platforms, alongside Amazon’s AWS (the pioneer) and Google Cloud (which lags far behind).
As discussed, cloud computing and chips represent the most lucrative aspects of the generative AI ecosystem. In a sense, this mirrors the re-creation of the Wintel duopoly. Microsoft has positioned itself advantageously by partnering with OpenAI and capitalising on enterprise customers who are already committed to Windows, Office, Teams, and other products.
There is also a fascinating saga of corporate competition, where Microsoft has turned the tables on Google in their decades-long rivalry. It is ironic that Google, through its DeepMind subsidiary (e.g., AlphaFold and AlphaGo) and its invention of Transformer technology, was a pioneer in both predictive and generative AI. Additionally, it succeeded in creating a search engine franchise.
However, Microsoft has effectively positioned itself as the leader in AI. Google’s core search and corresponding $200 billion advertising business now face jeopardy as users abandon its offerings in favour of new AI search engines like Perplexity or Grok.
The Chinese Challenge
The arrival of Chinese Generative AI products such as DeepSeek significantly reshaped the AI marketplace by 2025, introducing a blend of innovation, competition, and disruption that has resonated globally.
First, there was DeepSeek, which claimed cost-effectiveness, although some experts are sceptical about their assertions of an order-of-magnitude improvement. Reports suggest that DeepSeek-V3 was trained for under $6 million using fewer, less advanced Nvidia chips (e.g., H800s) compared to the billions spent by U.S. firms. This efficiency arises from techniques such as sparsity in model training (focusing only on relevant parameters) and data compression, enabling high performance with lower resource demands. This lowered the barriers to entry.
Second, there is open-source momentum. Unlike OpenAI’s proprietary models, DeepSeek has embraced an open-source approach, similar to Google’s GEMINI and Facebook’s LLaMA. Chinese tech giants like Alibaba and Tencent have also open-sourced their models (e.g., Qwen 2.5, Hunyuan), creating an “Android moment” for AI. These Chinese alternatives have eroded the pricing power of Western firms.
Third, there were geopolitical ripples: the US advocated for new investment, such as the $500 billion Stargate initiative. Concerns arose regarding Chinese products collecting data from various sources, especially as the Chinese government began treating its AI companies as “national champions” deserving of support.
The Chinese players changed the rules of the game: it is no longer solely about massive investments in the billions of dollars in proprietary systems, as seen in the US model, but rather enticingly about the potential to create LLMs using open-source Chinese products.
- India’s path to having a role in this domain
Objectives of Indian AI
The Indian government and society need to clearly understand how they want to position themselves within the expansive realm of Generative AI. In the authors’ opinion, the focus should be on leveraging AI’s capabilities for India’s benefit; therefore, striving to build products that compete globally with current market leaders would be unrealistic.
Two essential steps are required to develop AI products that address the needs of Indian society: debiasing and localisation. Each of these steps is briefly described.
Current open-source models, even on matters that concern India, are primarily trained on sources from outside the country. Two examples will serve to illustrate the downside of this: (1) If Deep Seek is asked a question about Arunachal Pradesh, it responds that no such place exists. This is because Deep Seek, a Chinese product, does not acknowledge that Arunachal Pradesh is part of India. (2) If any Western model is queried about the RSS, the answer will likely be that the RSS is a Hindu terrorist organisation. This is due to these models being trained on sources like Wikipedia, which are irredeemably hostile to India.
Eliminating such distortions is referred to as “debiasing.” It may be overly ambitious to believe that we can prevent individuals outside India from receiving a distorted answer to a query. However, at the very least, we can ensure an alternative query engine provides more accurate information.
For several decades, there has been a notable trend of Westerners appropriating India’s intellectual property. Examples abound: the healing powers of turmeric, Basmati rice, yoga, pranayama, and more. A strong nation with self-respect should aim to prevent future thefts and rectify past thefts.
Additionally, LLMs may begin to run out of training data, which could lead them to rely on “synthetic data” generated by AI or other artificial processes. This presents several issues: the amplification of existing biases in the models, a lack of real-world grounding, and the possibility of “model collapse,” where genAI starts producing gibberish. Consequently, genAI companies would need to seek new real training data, and IKS could be “digested. ”
In particular, traditional Indian Knowledge Systems (IKS) contain a wealth of material that can now be mined and appropriated by language models. To forestall this, it is essential to codify IKS in a format that unambiguously establishes the origin of the knowledge. This, in turn, requires incorporating IKS into an Indian Generative AI model, a process known as “localisation.”
Specifically, building LLMs that specialise in IKS would be desirable. For example, there could be one trained almost exclusively on Panini’s Ashtadyayi, which researchers could use to mine the depths of that masterwork and gather deep insights. In another example, recent cryptographic deciphering of the Indus-Sarasvati script might have been accelerated if there were an LLM that focused narrowly on the topic. Steps for achieving both objectives are described further below.
Approaches to Building an Indian AI Solution
Broadly speaking, two possible approaches to building a language model are foundational and fine-tuned (this phrase is not universally used). We discuss the advantages and disadvantages of each approach, placing particular emphasis on the Indian scenario.
A foundational model is essentially an ab initio model in which the model builders create their own pool of tokens from various data sources (public, proprietary, or both), select the model architecture, and then train the model by selecting the “weights” of the model. Generating a sufficiently rich corpus to produce realistic language models would require between 10 trillion and 100 trillion tokens and 500 billion to one trillion parameters.
Current models such as GEMINI, LLaMA, ChatGPT4, and Deep Seek all fall within this range. However, the cost would be substantially higher than for developing fine-tuned models. The IndiaAI mission envisages an outlay of Rs. 2000 crore, distributed over 6 to 10 projects (or Rs. 200 to 300 crore per project), with a six—to twelve-month development timeline for foundational models. In the authors’ view, it is unrealistic to expect any impactful foundational model to be developed with this level of funding.
To build a fine-tuned model, builders start with an open-source model that best meets their requirements and then adjust the weights so that the model performs well on their own additional data set, which may be proprietary. The key is to ensure that while adjusting the weights, the performance of the corpus used to train the original model does not deteriorate.
This is tricky because, while the weights of an open-source model are freely available, the corpus used to derive these weights is not. Fortunately, a decades-old idea from statistics comes to the rescue. If the size of the additional data used for fine-tuning is several orders of magnitude smaller than the original corpus, an approach known as “Low-Rank Adaptation (LoRA)” can be employed. Currently available open-source models are estimated to be based on 100 trillion (10^14) tokens. Any additional Indian data would not exceed a trillion (10^12) tokens, or 1% of the original (unknown) corpus. This suggests that fine-tuning would work well in an Indian context.
Developing a high-quality, refined model would be significantly cheaper than creating a meaningful foundational model. This is because the initial step would involve an open-source model that has already gone through rigorous development and testing. However, according to the rules of the software community, any model built upon an open-source model must be put back into the open-source world. This would not be a drawback for India and might even be an advantage because it may lead to India being perceived as a significant player in this domain. In contrast, developing numerous relatively small models, distinguished only by their foundational nature, would not improve how the rest of the world perceives India.
What are the Skill Sets Required?
To build even a fine-tuned model, two distinct sets of skills are necessary: algorithms and software engineering. Most algorithms used in training LLMs are available in “pseudo-code” form in the open literature. Therefore, it is relatively straightforward (assuming one is familiar with the literature, which is not always a valid assumption) to convert this pseudo-code into working code, typically in Python.
No additional software engineering is required for relatively small models, such as those with 5 to 20 billion parameters. Environments like PyTorch handle issues such as parallelisation and memory allocation. It is highly desirable for the engineers involved in this project to begin with an open-source model of this size and to establish programmatic solutions for fine-tuning, including debiasing and localisation. This approach will help them gain insight into the algorithmic issues at play.
However, no meaningful model will be so small. The models on which we will be working, even if the starting point is open-source, will be a minimum of half a trillion parameters. Scaling up the solutions mentioned in the previous paragraph to this size would require an understanding of software engineering, including optimisation and algorithmic knowledge.
This is tacit know-how: “underground knowledge” that is usually not written down anywhere. Normally, only those who have “been there and done that” would know these aspects. Ideally, we should attempt to attract at least a few people who have worked on the large open-source models currently available. These people could, in turn, train others.
Budget and Time-Frame
A Phase-1 Proof of Concept based on an open-source model of approximately 20 billion parameters that achieves both debiasing and localisation can be prepared in six to nine months and would cost roughly Rs 600 crore. Phase 2, a fully functional version, would aim for a complete solution and would require around nine to twelve months, with a budget of about Rs 1200 crore.
However, this cannot remain a government-run initiative; the private sector must also contribute. Initially, this may occur through pilot projects using CSR funds. Still, in the future, comprehensive LLM development, training, data centres, and marketing must come from the private sector at a scale significantly greater than the initial public-sector investment. Targeted incentives are necessary to stimulate private-sector participation.
Potential Applications in Indian Society
While there are numerous ways in which Generative AI investment can benefit Indian society, we will focus on one area: education. Concerns persist regarding the poor educational attainment of Indian students, particularly in standardised global tests such as PISA, in which India has stopped participating due to dismal scores.
A significant factor in achieving success may be mother-tongue education at the primary and secondary levels, particularly in the hard sciences. Countries with high PISA scores, such as Finland, Japan, South Korea, and Germany, implement this approach. There may also be a cognitive advantage: you grasp concepts rather than grapple with unfamiliar English words.
By using LLMs trained with appropriate sources, we can provide high-quality translations into Indian languages, facilitating mother-tongue-based primary and secondary education. Numerous nations have demonstrated that this is not a disadvantage for future R&D. Additionally, in an environment where the Internet and social media encourage deracination, it may assist students in maintaining a certain cultural grounding.
Some argue that tertiary education should primarily be in English, but this can also be managed if LLMs provide real-time translation of lectures, allowing students to listen in whichever language they prefer. A significant side benefit of this approach would be the ability to conduct simultaneous translation between any Indian languages, making everyday communication much easier and effectively reducing some of the ongoing language conflicts.
Conclusion
Generative AI is here to stay, warts and all. It is up to the Indian state and private sector to take advantage of its presence and to participate in ways that do not directly confront the free-spending American and Chinese market leaders. There are niche/leapfrog or disruptive innovation plays that can create substantial value for Indian society by improving education, nurturing and protecting Indian Knowledge Systems, and easing multilingual communication.
Authors Brief Bio: Shri Rajeev Srinivasan is an adjunct faculty member at IIM Bangalore, Dr M. Vidyasagar, FRS, is a former National Chair Professor at IIT Hyderabad, and Dr Abhishek Puri is a Radiation Oncologist at Fortis Hospital, Mohali. The authors can be contacted at rajeev@alumni.stanford.edu.
