When historians write about India’s technology decade starting in 2020, or “techade” as the Indian government has labeled it, they may well note an irony. The government mission tasked with one of the harder industrial challenges—convincing global semiconductor firms to physically plant themselves on Indian soil—quietly outperformed one with a more glamorous brief, on artificial intelligence. The India Semiconductor Mission (ISM) did not promise moonshots. It promised chips, cleanrooms, and committed capital. And it delivered.

Today, Micron’s assembly and test facility in Gujarat is close to transitioning to commercial-scale production. Tata Electronics’ packaging plant in Assam is doing the same. LAM Research and Applied Materials have set up tools presence. The Tata-PSMC fab in Dholera, India’s first commercial mega-fab, recently signed an MoU with ASML and is seeking to actively “deploy” ASML’s lithography equipment, a milestone that places India on a very short list of countries the Dutch precision-optics giant trusts with its machines. The Tata-ASML MoU also signals something far more consequential: that the semiconductor ecosystem is beginning to attract participants who do not merely follow demand but anticipate it. India’s recent and forthcoming free trade agreements could further provide the tailwinds to spur semiconductor exports to new markets, while also securing more cost-effective inputs across the supply chain.

Compare this to the India AI Mission (IAM), approved by India’s Union Cabinet in March 2024 with a budget of ₹10,372 crore. Much of this, around ₹4,568 crore, was earmarked to build compute capacity. On paper, the targets have been met and even exceeded. From an initial goal of 10,000 graphics processing units (GPUs), India assembled around 38,000 GPUs by late 2025. That is genuine progress, and the IAM deserves credit. But the honest question that must be asked, and in good faith, is whether acquiring GPUs is the same as building an AI future. The ISM’s approach also contrasts with that of the IAM. The ISM put together a small team of industry experts and went straight to work. One would be hard-pressed to find a working paper or concept note from the ISM that laid out what it was going to do. It simply delivered, across almost all parts of the value chain. The IAM, too, has an equally challenging task with respect to multiple layers of the AI value chain. Its official task is to “democratise access to computational resources, improve data quality, foster indigenous AI capabilities, attract top talent, support startups through risk capital, encourage industry collaboration, promote socially impactful AI projects, and ensure ethical AI development and use.” How has it scored on these parameters, and, in particular, on the objectives of “computational resources,” “improve data quality,” and “foster indigenous AI capabilities”?

From 2025 onwards, the Indian government released a range of AI-focused white papers. First there was a report on AI governance guidelines development by the Principal Scientific Advisor (PSA) in January 2025. This was followed by a paper by the Ministry of Information and Technology on AI governance in November 2025. In December 2025, the PSA released a white paper on democratizing access to AI infrastructure, followed by a white paper on a techno-legal approach to AI governance (which was defined as the integration of legal instruments, rule-based conditioning, regulatory oversight and technical enforcement mechanisms embedded with the technical architecture by design). Most recently, the PSA released a white paper in March 2026 on advancing indigenous foundational models. These are all significant outcomes derived through multi-stakeholder consultation processes. However, in most of these papers, it is interesting to note that compute is only mentioned as something to be procured, not built or manufactured.

The Assumption Nobody Examined

To begin with, a particular line of argument has already been surmised in several quarters, that the infrastructure for the entire AI value chain may be challenging to build. The sheer infrastructure gap and the resource constraints that face not only India, but any aspiring AI power, will serve as massive roadblocks. Not to mention the capital scarcity issues that have been acknowledged by some of India’s prominent tech czars. All this has generated widespread debate on which vertical India should focus on and the tradeoffs involved, mostly in favor of the application layer (the actual applications people use, as opposed to the underlying AI technology). A formal press release from the Indian government earlier this year stated that India was “pursuing an ‘AI diffusion’ strategy” of deploying AI applications across sectors.

However, embedded in India’s AI strategy is an assumption so comfortable it has rarely been interrogated: that access to the global AI supply chain is, and will remain, freely available. That India can procure compute from abroad, run models largely trained elsewhere, and layer Indian applications on top, and that this arrangement will hold indefinitely.

It may not. Two events corroborate this.

First, under the Joe Biden administration’s January 2025 Framework for AI Diffusion, entities faced cumulative restrictions on how many AI chips they could import, except those from eighteen countries, which were exempt because they also accounted for most of the global AI data center capacity and they were either a Five Eyes country, a close NATO ally, or already a major semiconductor hub. Reaching higher thresholds under this now-rescinded framework required government-to-government assurances, a process that would have demanded diplomatic, capital, and regulatory capacity that India is still building. What if a future U.S. administration were to reinstitute this framework? Indeed, recently, the United States Government Accountability Office (GAO) released a decision where it stated that the AI Diffusion Rule cannot be said to have been formally rescinded because the very announcement of the rescission is also a rule for the purposes of the Congressional Review Act. Essentially, the rescission would now have to be submitted to the U.S. Congress and the U.S. Comptroller General before it can formally take effect.

Second, when Anthropic announced its Claude Mythos Preview model in April 2026—a frontier AI model so potent in its cybersecurity capabilities that the company restricted access to roughly forty entities globally—India was not among them. Thirty-nine of those entities were American, with the one exception being the United Kingdom’s AI Safety Institute. Subsequently, the U.S. government ordered Anthropic to suspend access to its Fable 5 model to all foreign nationals, including to those that were its employees. Even though retraction of this model, as per the directions of the U.S. government, was not in line with usual U.S. Department of Commerce guidance, it did raise questions about the predictability of the U.S. export control regime, which had thus far targeted advanced AI chips.

This has been a wake-up call to those who perhaps thought that India’s large market size would give it preferable access to such frontier AI models. India has not yet built the institutional and technical ecosystem: the safety institutes, the evaluation and testing frameworks, and the enterprise-grade security infrastructure, which otherwise might have earned early access to these types of frontier AI models.

Even at the application layer, Indian developers can’t assume that access to models will always be available, or that it will be fair. Big AI companies could bundle their models with other downstream products or apps, by either packaging them into their own apps, or partnering with telecom companies to bundle AI access alongside things like cloud storage, security tools, and GPU access.

Lastly, even the “DPI [digital public infrastructure] for AI” approach, meant to democratize access to AI infrastructure, is at risk of being impractical if the underlying compute was procured and not built. The core promise of DPI for AI is reducing prohibitive costs for smaller players. A Government of India document sums up the promise of DPI for AI as follows: “Its value lies in creating predictable, transparent and interoperable access pathways, particularly for smaller firms, research institutions, and startups that face prohibitive entry barriers. By reducing costs, standardizing interfaces and establishing common governance norms, DPI can meaningfully expand the base of participants who can benefit from AI infrastructure.”

However, if India is importing GPUs at global market prices, which could potentially be tariff-laden, limited by export controls, or subject to demand surges driven by U.S. hyperscalers, the underlying compute cost remains stubbornly high. A well-designed access layer on top of expensive imported hardware is like subsidizing bus service on an expensive toll road: The governance is democratic, but the underlying economics are market-driven and not meant to necessarily benefit smaller firms, or be inclusive.

The ASML Lesson and What India Can Learn From It

There is a story from the history of semiconductor lithography that is worth retelling here, not as inspiration but as instruction.

In the 1990s, ASML was a distant third in a market dominated by Nikon and Canon, both of which had deep relationships with the biggest chipmakers of the era, namely, Intel, IBM, and others. ASML was smaller, less capitalized, and dependent on Zeiss for lenses it could not manufacture itself.[1] A rational observer might have concluded that the technical challenge of using extreme ultraviolet (EUV) light to etch microchips was too expensive, speculative, and unlikely to displace entrenched incumbents.

ASML did not reach that conclusion. Instead, it partnered with Zeiss to co-design the lenses that would eventually power deep ultraviolet (DUV) machines, a precursor to the EUV machines that came later. On EUVs, it formed the EUV Euclides consortium in 1998, attracting other European partners and governments to carry out research on how EUV light can be used to fabricate microchips. Once it reached a breakthrough in that research, it sought to work with Intel, which was motivated to support multiple lithography suppliers precisely because it understood the strategic risk of single-vendor dependence.[2] Today, ASML is the sole producer of EUV machines in the world. It works with over 5,000 suppliers and has become one of the most critical chokepoints in the entire global technology order, let alone the semiconductor supply chain.

The lesson is not that India should try to become ASML. The lesson is about the nature of the bet ASML made: It identified a technically daunting problem that the market needed solved, built coalitions rather than doing it alone, and made itself indispensable to the people who might have dismissed it.

What Can India Build That the World Will Need?

India has a tradition of building technology for development, not leverage: space imagery shared for disaster management; DPI exported at scale; vaccine diplomacy and generic pharmaceuticals that reached corners of the world where patented drugs could not. This is genuinely admirable, but the approach may be insufficient as a framework for building AI infrastructure.

As AI grows more sophisticated, as the supply chains that power it become more contested, and as access to frontier models becomes geopolitically charged, India must begin to ask a different set of questions. Not what applications it can build on someone else’s infrastructure but what the world needs, perhaps a few years from now, that we might be positioned to build and are yet to consider.

India must ask what is at the intersection of what the world needs, what it can build, and what the future will reward. It is not an argument for attempting everything. It is an argument against prematurely closing off possibilities by assuming India’s role in the AI value chain is fixed at the application layer.

Many AI hardware startups have struggled in the past because developers were reluctant to rewrite software for entirely new architecture. However, lately, by building AI chips that can work with CUDA (Compute Unified Device Architecture)-heavy environments, certain firms have arguably reduced the switching costs for enterprises and AI developers, which may also enhance developer adoption. Sure, it is early. The road is long. But it is the kind of bet that takes the market seriously, identifies a genuine gap, and does not wait for permission to try. A good example here is advanced packaging.

Advanced Packaging

The NITI Aayog, the Indian government’s think tank, recently released a roadmap for India’s semiconductor ambitions and outlined potential focus areas. A key area identified was advanced packaging, where the NITI Aayog report felt that “India should aim to secure self-reliance for domestic demand while aggressively targeting a top-three global position.”
This is a positive step. Packaging was earlier seen as a largely upstream part of the semiconductor value chain but has become essential to AI development as generative AI has revealed the limitations of single-chip architectures. The proliferation of large language models (LLMs) has led to a surge in demand for memory bandwidth (which helps provide data to the processor), giving rise to high memory bandwidth (HBM) that is challenging to create with traditional packaging. Accordingly, advanced packaging, with its focus on locating the memory as close as possible to the compute, stacks a large number of small chips on top of each other to improve performance and increases the density of the integrated circuit.

Compressed stacks of chips means that there is shorter distance between them, resulting in lower latency, and a significant jump in performance, all without upgrading to the next-generation process node. Indeed, most U.S. export controls are applicable to advanced downstream elements of the semiconductor value chain, such as semiconductor manufacturing equipment (SMEs), Electronic Design Automation (EDA) tools, and advanced AI chips. This makes it easier for countries like India to do more in this part of the value chain, unhindered by any export control issues.
Here, India could focus on interposers—electrical devices that connect chips to each other and the overall wafer substrate on which the chip is placed. These in turn require foundry-grade 65 nm process nodes, for which India can easily build infrastructure to manufacture. Consequently, the various layers of different materials stacked on top of each other can generate a significant amount of heat and require thermal management. Where traditional air cooling does not work, novel techniques, such as liquid cooling, can be explored.

The “Unknown Unknowns” We Have Not Asked

Former U.S. secretary of defense Donald Rumsfeld once said that there are “known knowns,” “known unknowns,” and “unknown unknowns.” India’s AI strategy has been largely organized around the first category. The “known knowns” here are things that India knows it needs, like GPUs and governance frameworks. It has begun to engage with the second, the “known unknowns,” which are things India knows it does not have yet, like sovereign models or the infrastructure to evaluate frontier AI safely.

But the third category, the things that India has not even thought to ask yet, is where strategic advantage is most often built. ASML didn’t set out to monopolize the lithography market. It set out to solve a problem that others had given up on. The question India must now sit with is uncomfortable but necessary: Is there something in the AI tech stack—a component, material, process, layer of infrastructure—that the current AI powers will need in quantities they do not yet have, at a scale they cannot yet imagine? Could India be positioned to build it?

This is not naïveté. It is the correct posture for a country that has demonstrated, repeatedly, that it can execute on hard technical problems when it decides to: in space, pharmaceuticals, DPI, and now, tentatively, semiconductors. The ISM’s progress on packaging, fabrication, and tooling is proof that the difficult work of industrial policy is possible in India, if done with patience and specificity. Though the IAM brought compute to India, compute purchased is not the same as capability built. The lesson of the semiconductor mission, of ASML, and of every firm that has made itself indispensable to a technology supply chain, is simpler and more demanding than any policy paper has yet captured.

Build. And they will come.

[1] Marc Hijink, Focus: The ASML Way – Inside the Power Struggle over the Most Complex Machine on Earth (Amsterdam: Uitgeverij Balans, 2023), 50.

[2] Marc Hijink, Focus: The ASML Way, 141.