Insights on current policy issues in India
—RSJ
For ordinary citizens of India, democracy means elections. Every few years, they participate in what the media likes to call the dance of democracy. They vote, governments change or survive, political parties celebrate or sulk (or break up spontaneously, like they do these days), television studios discover data science and garish charts, and India moves on to the next pointless ODI cricket match.
We spend very little time thinking about the machinery that makes those elections possible because, for the most part, that machinery has remained stable since Independence. The Census happened every ten years. Electoral rolls were periodically revised with a visit from a tired-looking school teacher assigned the role of a booth-level office (BLO). Constituencies remained frozen after a point. State elections and national elections came at different times. There were arguments over results, but most people rarely gave a second thought to the rules themselves.
This is about to change.
Over the next few years, India will find itself discussing four issues that are usually treated as separate conversations. There is the Special Intensive Revision (SIR) of electoral rolls, which has already become controversial in several states. Then we have the Census, finally being conducted after a gap of sixteen years. There is also the debate over delimitation, which will inevitably follow once the Census is complete. And there is also the misguided clamour for One Nation, One Election, which seeks to synchronise elections across the country.
Of course, each has its own administrative logic. Electoral rolls need updating every few years. We have already delayed the Census, which gathers the core demographic data of the nation, for over 6 years. That can’t be delayed further. Parliamentary constituencies cannot remain frozen in time based on Census data that’s fifty years old. And, some may argue that having big state or national elections scheduled every 9-12 months isn’t helping. Simultaneous elections promise lower costs and fewer interruptions to governance. Business as usual and nothing more to see.
But this is where I will channel my inner Kishore Kumar and sing, “Babu, samjho ishaare”. Look at them with a different lens; each one of these exercises taken together asks a basic question about how representative democracy works. Who gets to vote? How is that right established? Who gets counted? How should representation be divided across the country? How often should governments seek a fresh mandate? These are questions that most democracies settle early in their lives and then revisit only occasionally. India seems to be reopening all of them at roughly the same time. Is it all really needed? And if it is, is this the best way to go about it?
Take the Special Intensive Revision. The Election Commission’s objective, prima facie, makes sense. Electoral rolls should be accurate. Dead voters should be taken out. Duplicate entries should be eliminated. No democracy benefits from inaccurate voter lists. Things get interesting with the process adopted rather than the objective.
Across several states, voters who have participated in election after election are being asked to produce documents to establish their eligibility once again, sometimes requiring records related to their parents or earlier generations. In a country where births were not universally registered until fairly recently, where migration has been constant and where official records remain uneven, this is a much larger exercise than it appears on paper. The individual horror stories that are emerging from people who have gone through this exercise suggest either a lack of any forethought on how this will play out for ordinary citizens or a deliberate attempt to make it difficult for people.
Former Chief Election Commissioner S.Y. Quraishi captured the concern when he argued that the exercise appeared more focused on exclusion than on inclusion. Whether one agrees with his assessment or not, it points to a legitimate question that deserves discussion. In a democracy, where should the burden of proof lie? If someone has voted for decades, should the State have to demonstrate why that person should no longer remain on the rolls, or should the citizen once again establish a right that has already been recognised repeatedly?
To me, it should be with the State. Democracies have to protect both the integrity of electoral rolls and the right of eligible citizens to participate. Those objectives can sometimes pull in different directions, but to make voting citizens with multiple proofs of identity run pillar to post to establish their voting right should be unacceptable.
The debate also exposes another ambiguity that the state seems to be revelling in putting the citizens through. The Election Commission is responsible for preparing and maintaining electoral rolls. It is not the authority that determines citizenship. Yet when voter verification increasingly depends upon proving family lineage or historical residence, the distinction between verifying electoral eligibility and scrutinising citizenship becomes harder for ordinary citizens to understand.
Many people possess voter identity cards, Aadhaar cards, passports, ration cards, tax records and decades of interaction with the State, yet struggle to produce a particular document from fifty years ago that has been made compulsory to feature in the electoral list. The SIR process right now can seem like a Kafkaesque nightmare if you fail to have one key document. The issue is therefore larger than documentation. It is about the relationship between the citizen and the State. At what point does the State presume that a citizen belongs, and at what point does it ask that citizen to prove belonging once again?
That question leads to the Census, because before a democracy decides who votes, it has to decide who counts. The Census is often treated as a statistical exercise and therefore attracts little public attention. In reality, it underpins almost every important decision governments make. Population figures influence fiscal transfers, welfare programmes, infrastructure planning, urban development, reservation policies and eventually political representation itself. After sixteen years without fresh data, India is finally preparing to count itself again. The final census data will be out by 2030. The numbers that emerge will shape public policy for years and will become the foundation for the next major constitutional debate.
That debate is delimitation. For decades, India chose to freeze the distribution of Lok Sabha seats among states because successive governments did not want states that had successfully reduced fertility rates to lose political influence as a consequence of that success. The freeze was always temporary. Once fresh Census data becomes available, the issue will return.
On the face of it, the principle is simple. Democracies are built on political equality. If populations have changed dramatically, representation should reflect those changes. Why should an MP in one state represent substantially fewer citizens than an MP elsewhere?
Yet the simplicity disappears once one considers India’s federal character. States in the South invested heavily in education, healthcare and family planning over several decades and now fear that demographic success could translate into reduced influence in Parliament. States in the North ask an equally reasonable question. Should their citizens continue to have less representation simply because they live in more populous states?
Neither argument is devoid of strong logic. One is rooted in equality among individual citizens. The other is rooted in fairness among the states that constitute the Union. Every federation grapples with this tension in one form or another. India postponed the debate for half a century. It cannot postpone it much longer. But what model should it adopt? How should it balance the debate between equal electoral weight versus fair representation for the states?
We should be using this opportunity to review the delimitation exercise and possibly, redress the flaws that have embedded themselves into our representative democracy. There should be wider consultations on proposals that are already in the mix and a more informed public debate on the alternatives available. But the nature of our polarised polity and the centralised instincts of this government is precluding this. There’s strong suspicion, possibly unjustified, among southern states that the delimitation exercise will leave them worse off in representation.
The proposal for One Nation, One Election completes the chain. Its supporters argue that India spends too much time conducting elections, enforcing the Model Code of Conduct and interrupting governance. Synchronising elections would reduce costs and allow governments to focus on administration rather than campaigning. Critics respond that staggered elections strengthen federalism because they allow voters to judge national and state governments separately and at different moments. The staggered election cycle, as it stands today, allows political parties to shape their political conversations relevant to the state independent of the national mood. It acts as a bulwark for federalism and against further centralising tendencies of the state. Also, I have my own doubts about the argument in favour of administrative efficiency for One Nation, One Election.
The temptation in Indian politics is to reduce every institutional debate to immediate political advantage. I fall for it myself. The government is accused of trying to entrench itself by changing the rules of the game in its favour. The opposition is seen as resisting every reform regardless of merit. There is enough evidence in public life to justify a degree of scepticism about political motives, but that scepticism should not become the entire conversation.
We have always argued on these pages that institutional design deserves a different standard of debate. The rules framed today will shape governments that have not yet been elected and political parties that do not yet exist. The Constituent Assembly understood this well. Its members spent years debating universal adult franchise, federalism, representation and the relationship between citizens and the State, not because they expected perfect answers but because they knew imperfect answers would endure for generations. Those debates were public, exhaustive and often inconclusive until broad agreement emerged. The legitimacy of the institutions that followed came not merely from the text of the Constitution but from the process through which those choices were made.
India is now approaching another moment when some of those foundational questions are returning. Questions of this scale cannot be treated as routine administrative exercises or settled through technical notifications or diktats alone. There is also a need to be sensitive to the travails of ordinary citizens and their feedback as they navigate these simultaneous exercises questioning citizenship, voting rights and representation. It is also one that deserves to involve many more voices than it does today, because governments come and go, but the rules of the game have a habit of staying around for a very long time.
Global issues relevant to India
—Pranay Kotasthane
This was a big week for AI geopolitics. Which week isn’t, you ask? No really. Some major things happened. On Wednesday, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model. Independent benchmarks place it narrowly behind Anthropic’s fabled Fable 5 on certain coding-heavy tests. But the kicker is that the full weights will be available for download by July 27. Anyone, anywhere, with enough inference compute can run it on their own servers.
The next day, Xi Jinping made his first-ever in-person appearance at the World Artificial Intelligence Conference in Shanghai. His message was that “AI development should not be a solo performance by a single country, but a symphony of international cooperation.” He announced WAICO (to be read as a Belt and Road Initiative for AI), a 29-country AI cooperation body headquartered in Shanghai, pledged 5,000 AI training spots for developing countries, and positioned China’s open-source models as a public good. Chinese state media framed it as a counter to America’s “AI Iron Curtain.”
Meanwhile, US government policy can be summed up in one word: confusion. I’ve been tracking US policymakers’ positions on Chinese open-weight models, and their arguments cycle through at least four positions.
One: their models have backdoors, ours don’t. This is technically plausible but strategically irrelevant. Open-weight models have the advantage that anyone can audit the weights. If anything, backdoor concerns should be more acute for proprietary models you can’t inspect. The US government demonstrated this rather vividly in June when it ordered Anthropic to pull Fable 5 from global access within days of launch, citing a jailbreak report. The model returned eighteen days later. If closed models are trustworthy because they’re closed, why did the government need to yank one off the market?
Two: China can subsidise this, but we face the “tyranny of the quarterly cycle.” This is the argument of a great power making excuses like a middle power. The depth of American capital across technology readiness levels is its strength, not a weakness. Neither is the claim true. DARPA funded OpenROAD, the open-source chip design toolkit that revolutionised semiconductor layout. RISC-V, now challenging ARM’s dominance, came out of UC Berkeley. Kubernetes, Android, TensorFlow, and Docker are all critical open-source digital infrastructural projects that originated in the US. In fact, the US has been, for decades, the world’s most effective practitioner of open technology as a strategic tool. The complaint that American firms cannot compete because they report quarterly earnings is not an argument about Chinese strength but is an admission of domestic policy failure.
Three: they’re distilling our models, so they’re not playing fair. OpenAI, Microsoft, and Anthropic have all accused Chinese labs of extracting their model outputs to train competitors. Some evidence is suggestive—DeepSeek outputs that identified themselves as ChatGPT, large-scale API extraction detected by Microsoft’s security team. But even if entirely correct, the policy implication is unclear. You cannot un-distil a model. The knowledge is already out. And knowledge distillation is a standard machine learning technique, not an exotic weapon. Treating it as espionage while simultaneously failing to prevent it is not a strategy but mere whining.
Four: export controls will fix it. Three rounds of semiconductor export controls since October 2022 were supposed to deny China the compute for frontier training. DeepSeek emerged anyway. Chinese labs found architectural workarounds, used the below-threshold Nvidia chips, and produced competitive models at a fraction of American costs. And yet, we are moving toward more export controls on advanced AI chips. The medicine is not working, and yet the US is increasing the dosage.
These four arguments form a sequence of rationalisations by a power that senses its advantage slipping.
A confident technological power would not respond this way. It would champion openness and trust its private players to make world-beating products—as it once did. Xi’s WAIC speech—“encouraging openness, collaboration and sharing”—could have been written by an American technology optimist circa 2010. Instead, it was delivered by the leader of a one-party state with an extensive censorship apparatus, to applause from 29 countries. When your response to a competitor’s openness is restriction and secrecy, you signal anxiety and not strength.
While Washington debates its stance, Nikkei Asia reported last week that Indian companies are increasingly switching to Chinese LLMs to contain AI costs. Chinese open-weight models cut costs by “an order of magnitude.” And so, this is a market outcome, and it vindicates what I’ve argued in this newsletter before: India’s interests in AI are orthogonal to the US-China framing. Chinese open-weight models are inputs, not threats. India benefits from a multipolar AI supply structure where competition keeps costs down and options open.
PS: A falsifiable prediction. The CPC will impose formal export controls on its frontier open-weight models before January 1, 2028. This could take the form of licensing requirements, delayed releases, parameter-count thresholds, or end-use restrictions. Xi’s WAIC speech was the high-water mark of Chinese AI openness. The Reuters report on Beijing’s meetings with Alibaba, ByteDance, and Z.ai about curbing overseas access is the leading indicator. Openness was the play when Chinese models were catching up. Now that they are competitive, the logic of giving them away collides with the logic of controlling them, especially as AI becomes important for narrative control. Nevertheless, the recommendation for India doesn’t change since there are other open models outside China.
Global issues relevant to India
—Pranay Kotasthane
(This post was first published on the Takshashila Blogroll)
ASML reported its second-quarter 2026 results on July 15. Net sales came in at €9.3 billion, well above the consensus estimate of €8.8 billion. Net income was €2.9 billion. Most significantly, the company raised its full-year 2026 guidance to €43-45 billion in revenue, a sharp jump from the €36-40 billion range it had guided just a quarter ago. CEO Christophe Fouquet described order intake as “extremely strong” and announced plans to expand capacity for both EUV and DUV tools by 30 per cent in each of the next two years.
These are impressive numbers. But ASML is also a weathervane for the AI infrastructure story as it is the sole manufacturer of extreme ultraviolet lithography systems and holds a monopoly on the advanced lithography equipment that every GPU used for training requires. Broadly speaking, its order book is thus like a real-time dashboard of two forces: the global AI infrastructure buildout and the US-China semiconductor contest.
At its previous Investor Day talks, the company projected that 5 to 10 per cent of global fab capacity by 2030 will exist because of geopolitics rather than market demand. That is, capacity is being built not because the market needs the chips, but because governments have decided that strategic resilience requires domestic production. Every one of those geopolitically motivated fabs still needs ASML’s machines. Fragmentation, paradoxically, is good for ASML’s top line, at least until the AI bubble lasts or until the next inevitable trough of the semiconductor business cycle arrives.
The China story first. China’s share of ASML’s system sales fell from 36 per cent in Q4 2025 to 19 per cent in Q1 2026. The Q2 earnings call confirmed that China-related business remains at approximately 20 per cent of total net sales for 2026, though that percentage now applies to a significantly higher revenue base than previously expected. Thus, ASML expects to continue selling the N-1 generation DUV machines to China.
Next, these earnings also indicate China’s progress in alternatives to ASML’s much-in-demand machines. It seems that at the EUV level, China is still far behind. The technology required decades of ecosystem development and cross-border collaboration that cannot be easily replicated under sanctions conditions.
At the DUV level, however, the picture is more nuanced. SMEE, the Shanghai-based equipment maker, has delivered early 28nm ArF immersion DUV systems, reportedly designed to avoid any US-origin intellectual property. But it hasn’t been productised yet. SiCarrier, a Huawei-linked entrant, is testing a domestically manufactured immersion DUV tool at SMIC. These are prototype-stage tools, roughly where ASML was in the early 2010s. Even ASML’s EUV took thirteen years to go from a working prototype to a consumer product.
Thus, China’s substitution efforts are not production-ready, and they are not close to matching ASML on throughput, yield, or overlay accuracy. But the direction of travel is clear. For ASML’s long-term earnings trajectory, the key determinant is whether Chinese firms have found substitutes or alternatives.
The next escalation that could impact ASML is not far away. The MATCH Act, currently making its way through the US Congress, would ban not just new DUV lithography sales to China but also the servicing of existing equipment. The servicing piece matters more than the equipment ban. Chipmaking tools degrade rapidly without constant maintenance. If servicing stops, China’s entire installed base of ASML DUV systems becomes a depreciating asset, regardless of whether Chinese firms can eventually build indigenous alternatives.
This brings us to the Netherlands’ position. ASML accounts for roughly a quarter of the total market capitalisation of the Euronext Amsterdam exchange. Together with Shell, these two companies make up nearly 46 per cent of the exchange. The Dutch government has committed €2.5 billion in public infrastructure through its Beethoven project to keep ASML anchored in Veldhoven.
Asking the Netherlands to aggressively restrict ASML’s China business is comparable to asking Saudi Arabia to leave oil in the ground. The Dutch trade minister has explicitly said that the Netherlands opposes the extraterritoriality of the MATCH Act. And yet, the historical pattern is clear that the Netherlands blocked EUV exports from 2019, restricted DUV exports from January 2024, and will likely comply again when forced. The operative word is “forced”, meaning that ASML will follow the letter of US restrictions while creating maximum room to protect its remaining business.
Share Anticipating the Unintended
Insights on current policy issues in India
—Pranay Kotasthane
Every time India announces a semiconductor assembly plant or a large data centre, a familiar objection makes it to the news: what about the water? These plants, we are told, will drink a parched country dry. Is it really so?
The objection has become a ritual. And like most rituals, it survives on repetition rather than arithmetic. So let us put the numbers together to analyse water consumption in India.
Agriculture consumes close to 90 per cent of all the water used in India, two to three times more than China or Brazil, according to the World Bank. Everything else, every factory, every water filter, every fab and server farm, fights over the remaining tenth. One kilogram of Indian cotton takes around 22,500 litres of water to grow, according to the Water Footprint Network’s estimate. The global average is 10,000 litres/kg of cotton. Similarly, sugarcane and rice are water-guzzling crops, incentivised by a system of MSPs, fair remuneration prices, and state-administered prices.
Now, place the newcomer with this background in mind. A large semiconductor fab draws 38 million litres of water per day, while an average hyperscale data centre consumes about 19 million litres of water every day. That sounds enormous until you compare it with other uses. Across a full year, it is the same as what 500 to 800 hectares of paddy consume in a single season. Punjab alone grows paddy on over three million hectares. So the entire fab, running every hour of every day, amounts to a rounding error.
Moreover, a fab needs ultrapure water, and producing it wastes a little raw water in the making. But withdrawal is not the same as consumption. A modern fab reclaims most of what it takes, treating and reusing two-thirds of it before discharge. Much of a fab’s water returns to the system.
I am not claiming the concern is imaginary. Water problems are local. A fab draws from one basin, one aquifer, one municipal supply, and in a dry year, it can strain that source. Taiwan discovered this in 2021, when a drought forced the island to divert water from farms and left TSMC trucking in supplies. But the answer to a local risk is siting and recycling, not prohibition. Building the fab where water is available, auditing water withdrawals, and making high reuse a condition of approval can solve problems. These are just permitting problems, not blockers.
If agriculture uses nine litres in every ten, then improving farm water efficiency by a couple of percentage points can free up nearly as much as India's entire domestic water consumption.
But we fixate on the fab and a data centre’s water consumption because they are new, visible, and made by firms. The tubewell is old, diffuse, and politically untouchable, spread across millions of voters that no party dares confront.
The demand should never be “build less of the good stuff.” India’s water crisis is a reality, but it is a crisis of allocation and pricing, sitting almost entirely in a farm sector that pays next to nothing for its biggest input. Let’s fix a sliver of that.
P.S.: There are other fairly good reasons to question these investments, especially if the government funding is involved. But water consumption isn’t one of them.
Reading and listening recommendations on public policy matters
[Book] Ameer Shahul’s Vaccine Nation: How Immunization Shaped India makes for interesting reading as it narrates the story of vaccine development in India over the last 150 years. The fact that the rabies vaccine came into being only in 1927 was especially relevant for me (Pranay) because I got bitten by a street dog a week ago! Imagine, just a hundred years ago, there would have been no solution for this.
[Survey] Takshashila does an annual survey to assess Indian public opinion on technology geopolitics. This is an anonymous survey with 15 questions and takes about 5 mins. Please give it a try and help us! Link: tinyurl.com/technopolitik
[Paper] China’s mercantilist squeeze on developing countries provides much-needed empirical analysis of the impact of Chinese overcapacity on other low-income countries.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.