Insights on current policy issues in India
—RSJ
Here are three different newspaper headlines from the past two weeks for you to read together.
“For1st time in a decade, India to import 10 lakh tonnes of sugar.” (Times of India, Aug 21, 2026)
“Consumers queue up for premium petrol as E20 concerns refuse to die down” (Mint, Aug 21, 2026)
“E20 mileage row: Govt says fuel alone cannot be blamed; cites driving, traffic, maintenance” (Times of India, Aug 15, 2026)
I confess I don’t shop for staples and vegetables at my home, so I had no idea that retail sugar prices had gone up by about 16 per cent in the past month. And India, which for years has had surplus sugar, has now opened a duty-free window for importing one million tonnes of raw sugar until October 31 to manage the impending festive season demand. Exports have been banned, and a series of stockholding restrictions have been placed on wholesale sugar buyers to prevent hoarding.
Mild nostalgia of the 80s crept into me reading all this. Back then I used to carry a ration card and a few textbooks to pass the hours standing in queue to buy sugar from the government “control” stores. Sugar was a “controlled” commodity, and we had a ration of five kgs per household per month or whenever there was stock in that store. What character-building endeavours those were.
India, to remind you, is the world’s second-largest sugar producer, and for much of the past decade the problem for the sugar industry has been surplus production rather than scarcity. So, what has brought us here?
While looking for answers, I also discovered a singular government organisation - the Directorate of Sugar and Vegetable Oils. This hallowed body has this responsibility (as per its website):
”In exercise of powers conferred under the Essential Commodities Act, 1955, Directorate of Sugar & Vegetable Oils monitors the production, sale, export and stock availability of sugar in the country to ensure sufficient availability of sugar for domestic consumption at stable prices. In addition, the Directorate is also a nodal Office to ensure sufficient capacity creation and production of ethanol in the country to meet the target of 20% ethanol blending with petrol by 2025 in India. The Directorate is responsible to implement various orders of the Government of India namely the Sugarcane (Control) Order 1966, the Sugar (Control) Order 1966 and the Sugar Price Control Order 2018.”
Kamaal hai.
After all this monitoring of every single leg of the sugar value chain, we are here! The government tracks acreage, production, recovery rates and stocks of sugarcane and sugar. It regulates pricing across the value chain. This is a commodity where the government is an active participant in the market.
And yet we have runaway sugar prices to deal with.
The immediate explanation offered by the government is an unanticipated drop in production. The government estimates current sugar production at around 30.6 million tonnes, against an earlier estimate of 34.3 million tonnes. This has meant a low opening stock projection for October (when the festive season peak starts) at about 3.5 million tonnes as against the usual 5-6 million tonnes. T
he drop in production volumes is being blamed on poor monsoon, lower recovery from the farms and disease in sugarcane crops in certain states. These are possibly good reasons. Although sugarcane acreage in July was 1.5% higher than in the corresponding period last year, despite deficient rainfall. That doesn’t necessarily mean more production, but it should give us pause before we blame poor rainfall for lower stock estimates.
One possible answer is related to the second and third news headlines that I started with.
India’s E20 programme has meant almost all petrol sold at retail pumps now has 20 per cent ethanol in it. This has created a second large market for sugarcane. The country has invested heavily in distilleries to meet the E20 target, and ethanol production capacity had reached around 1,822 crore litres a year across 499 sites by mid-2025. This is about 600 crore litres in excess of what’s needed for the domestic ethanol blended petrol programme. Sugarcane accounts for around 30-35% of ethanol feedstock, with the remainder coming from maize and rice.
I have written earlier about the E20 programme, its questionable benefits and the illogical move to not offer any real options for customers except to purchase E20 petrol. The underlying logic for imposing E20 petrol is that India imports a large share of its crude oil, and reducing that dependence is a strategic objective. But that should not mean that any solution that does so is a good solution regardless of its other consequences.
It is difficult to argue against the notion that E20 has changed the underlying economics of the sugar market. If the demand for E20 petrol continues to remain strong, since there is no other option available and that gives better realisation to the mills, it will mean more cane being diverted to the distilleries. The government has created an incentive to support higher E20 production and then discovered that players respond to that incentive, which creates a problem elsewhere. This is the old problem of not anticipating the unintended.
But there’s more. Sugar exports have contributed to reducing our trade deficit over the year. In 2021-22, sugar exports exceeded 12 million tonnes. Now we have temporarily banned exports, and we are importing sugar after slashing the 100 per cent duty we had on it (that itself is mind-boggling). The global sugar market has already factored in the absence of Indian sugar for the next few quarters. The international prices had moved from around $474 a tonne at the end of June to $552 on August 20. We will be buying sugar from the international market at a time when global prices are at their recent peak. On the one hand, E20 petrol reduces India’s crude import bill, but it might be leading it to import sugar at much higher prices.
The environmental question of increasing sugarcane acreage to cater to this demand is relevant too. Sugarcane is a water-intensive crop. Clearly, as E20 adoption is complete and we move to higher blends, there will be one of two consequences. Either we will have a sugar shortage for retail consumers as more of it is diverted to ethanol distillation, or there will be more area that will come under sugarcane cultivation. So, the cost of the ethanol blended petrol programme cannot be measured only in litres of imported crude that will be reduced. There are other considerations too, including sugar imported, groundwater being used, land being allocated and other crops or uses being displaced. These costs right now are outside the calculus.
The CEA, V. Anantha Nageswaran, has made the same point in arguing that India should remain at E20 and not move to higher blends until it has “thoroughly costed the food-versus-fuel trade-off”.
Let me bring the other headline into focus now. The government has said that E20 is safe and that fears about engine damage are unfounded. It has pointed to testing across the country and rejected the claims being made about the fuel. Who knows whether those tests are rigorous or not? The automakers have, with remarkable speed, declared their engines are safe with E20 petrol.
But here’s the revealed preference of the customers. Mint reports that the share of premium petrol in sales has risen from around 4% in March to 12-15% at some pumps. Consumers are paying more for higher-octane petrol (which has 20 per cent ethanol) because they believe it will help with concerns around mileage and engine performance under lower-octane E20, which is cheaper.
Consumers are spending more of their own money because their own lived experience and those of people around them don’t convince them of the government’s assurances. It can’t just be perception-based, on which customers are spending real money to safeguard their vehicles from E20. They should be seeing first-hand evidence of mileage calculations and concerns about long-term vehicle performance.
It goes back to my earlier argument on these pages. If E20 is the better product, consumers will buy it. If consumers prefer E10 or ethanol-free petrol and are willing to pay for it, that choice also contains useful information. Instead, India has moved towards a system in which E20 is the default, leaving consumers to trade up to expensive premium fuels. That they are doing so without a lot of protest suggests the government narrative of adopting E20 as a national duty of every Indian is working.
The Brazil comparison, which is brought into this, also deserves a bit more analysis. Brazil’s ethanol programme developed over decades, beginning with the oil shocks of the 1970s, within a different political economy (it was a military dictatorship), a smaller population (with fewer cars), and a different land-use pattern for agriculture. It took them over two decades since then, and that too in a world where alternative energy options were limited. India’s transition has been much faster, while we are simultaneously dealing with food security, water stress, a large and varied vehicle fleet and a sugar industry that is deeply entrenched in the political economy.
Once the immediate sugar stock crisis passes, the government should revisit the assumptions underlying the sugar and ethanol policies together. What is the full cost of the E20 transition? A tonne of sugarcane can become sugar or ethanol. Let the market make the choice based on the trade-offs made by millions of players in that market. Right now, the government is influencing that choice without fully appreciating the trade-offs involved.
This won’t be the last time we will have a sugar crisis on our hands.
Insights on current policy issues in India
—Pranay Kotasthane
Who says India doesn’t have a sovereign LLM? Not just a national LLM; many Indian states will soon have one each. And Maharashtra is leading this war of Local Language Mandates.
Of all the things that an RTO could do, the Mira-Bhayander RTO investigated whether auto-rickshaw drivers know Marathi. Then the state transport minister roared that auto-rickshaw and taxi drivers must know Marathi. With alacrity, his department began a 105-day campaign (read state-sponsored gundagardi) since May to teach Marathi. And now it has begun enforcing these rules.
See the Kafkaesque details courtesy Indian Express’s Zeeshan Sheikh:
The Transport Department, however, has maintained that the present exercise is about enforcing existing rules. Officials have also distinguished between permits and badges, arguing that Rule 24 applies to badges and that the 2017 judgment did not strike down the language requirement for badges.
The government subsequently made amendments to Rules 4, 78 and 85 to formalise Marathi proficiency as a condition for authorisation, permits and renewals. The draft rules were put out for objections and suggestions.
What happens to those who fail?
Drivers will not lose their licences immediately. Those found lacking the required proficiency will first be given a notice and one month to learn Marathi.
If they fail to comply, their PSV badge and driving licence can be suspended for three months. Repeated violations could result in permanent suspension.
The department is also checking for other irregularities, including allegations that fake domicile certificates were used to obtain permits and badges.
Good ideas travel far. Bad ideas travel further. So I expect many states to copy this idea to launch their own LLMs.
These language training and assessments are just a ruse by incumbents to protect their turf from economic migrants. Making language mandatory for permits and renewals only licenses extortion by local goons and harassment by the state.
Despite the fundamental right guaranteeing the freedom of work, state governments try to get around the spirit of the Constitution in imaginative ways. Linguistic reorganisation on the principle of ‘one language, one state’, rather than the ‘one language, one state’ idea that Ambedkar recommended, has created states that assert monopolistic control over languages. The result is such absurd LLMs.
Despite their absurdity, these rules will find favour amongst ‘locals’, who have also internalised that the primary responsibility for developing their language rests with the state government rather than the society itself. Another case of economic freedoms being sacrificed at the altar of statism.
And I haven't even factored in the time, effort, and tax money that went into these training and assessment efforts. Doesn’t the RTO have other high-priority items to address related to road safety, traffic rules enforcement, and professional driving license tests?
It’s 2026, and we could easily use the other LLM to override our linguistic differences. But then, every state gotta have its LLM.
Big fish eating small fish = Foreign Policy in action
—Pranay Kotasthane
Why the Frontier Anxiety Is Overrated
Last week, I participated in a roundtable organised by the Council on Foreign Relations. The thesis I was responding to was that middle powers have no option but to fall in line with either the US or China in the AI era. I was invited to provide an Indian perspective. (Since it was held under the Chatham House Rule, I will stick to what I said.)
The discussion gave me a chance to articulate an Indian perspective on a question consuming the AI policy world: as access to AI models and training hardware becomes more contested, what should countries like India do?
I agreed with the diagnosis that access to frontier AI is becoming more securitised. But I disagreed that it would have disastrous strategic consequences for a country like India. I had five reasons in support of my argument.
The arms-race framing of AI in the US follows a logic that frontier access is scarce, falling behind is dangerous, and recursive self-improvement will make catch-up impossible for other countries.
But there are actually four distinct ways to frame AI geopolitically: AI as nuclear weapons, AI as a zero-sum arms race, AI as an innovation race, and AI as a general-purpose technology. The framing a country adopts determines the policy instruments it uses. Arms race implies export controls and compute denial. Nuclear weapons framing leads to non-proliferation and hardware monitoring. Innovation race framing creates support for subsidies and ecosystem-building. The GPT framing emphasises diffusion infrastructure and human capital.
India is betting on the fourth frame. Jeffrey Ding’s work has shown that historically, the countries that benefited most from general-purpose technologies were not the ones that invented them. Germany began the chemical revolution, but the US deployed it at scale. No country “won” electricity. Multiple countries found their place in the supply chain and captured value in different ways.
India knows this from its own experience as well. India didn’t build Java, Oracle, SAP, or cloud computing. What India built was the world’s deepest organisational capacity to deploy and integrate technologies invented elsewhere. The result is that India exports more in IT services every year than Saudi Arabia earns from oil.
Further back, the nuclear and space domains tell a similar story. After the nuclear tests in 1974, the world denied India reactor technology. The MTCR restrictions denied India cryogenic engines for the space programme. On both occasions, the prediction was permanent subordination. But India developed credible capability through self-strength backed by diversified partnerships.
The recursive self-improvement hypothesis also needs questioning. Real-world systems—enterprises, government departments, supply chains, defence systems—don’t change easily just because AI exists. The binding constraint for general-purpose technologies is always diffusion because organisations and processes take time to adapt. The question for India is therefore not “how do we get guaranteed access to the best model,” but where India’s place is in the AI supply chain, and how we make ourselves indispensable.
The “cut off” scenario applies to proprietary API models—where a company in San Francisco can revoke your access with an export control directive. It does not apply to open-weight models running on domestic hardware. Once a firm downloads the weights, no one can un-download them. The tap that can be turned off is the API. The weights are a static file.
But here’s the deeper point. The US has already tried blocking access at a much harder layer—chips. Since 2022, the US has progressively denied China access to frontier GPUs. A100s, H100s, even the downgraded H20. The result was not Chinese subordination. The result was DeepSeek, Qwen, and Kimi. China compensated for weaker chips with algorithmic innovation, engineering efficiency, distillation, and brute-force scaling. The controls restricted one layer, and the other layers adapted.
This was Jensen Huang’s ‘AI is a five-layer stack’ argument in practice. Strength at one layer offset weakness at another. If that’s true for chips, where there are essentially two suppliers globally, it’s even more true for models, where there are a dozen capable options from many different countries. If chip denial produced DeepSeek, why would model denial produce Indian subordination? The model layer is the most substitutable part of the entire stack.
The compute scarcity argument assumes today’s architecture is permanent. But three things are happening simultaneously.
First, inference is splitting from training. Training frontier models will stay expensive and concentrated. But inference—which is what deployment requires—is moving toward specialised ASICs, custom silicon, and edge devices. Models themselves are getting smaller through distillation, mixture-of-experts, and quantisation. Last year’s frontier capability runs on this year’s commodity hardware.
Second, the empirical pattern is that open-weight models close the gap to the frontier within months, repeatedly. DeepSeek emerged despite chip denial. Llama, Qwen, Mistral—every quarter, the next capable open model arrives.
Third, Sakana AI, a Japanese lab, launched an orchestration model called Fugu. Fugu doesn’t train frontier models. It coordinates other models, routing each task to the right specialist, delegating sub-problems, and verifying outputs. On ten out of eleven major benchmarks, Fugu Ultra matched or exceeded Anthropic’s Fable 5 without training a single frontier model and without even having those restricted models in its pool. Their launch pitch was explicitly about “frontier capability without the risk of export controls.”
Thus, multiple pathways are available to capable and motivated middle powers.
Middle powers are not supplicants negotiating for access if we were to look at the actual market dynamics.
India’s AI usage of ChatGPT and Claude is among the highest in the world. OpenAI and Anthropic are not doing India a favour by providing access. India is providing them revenue, usage data, and the demand signal their investors need to justify $100 billion valuations. The biggest buyer has leverage too.
And then there’s talent. Most frontier AI labs are building or expanding engineering centres in India. Google’s largest engineering office outside Mountain View is in Bengaluru. If the US were to “cut off” India from frontier models, it would also cut off American firms’ global capability centres. The AI labs need Indian talent at least as much as India needs their models. Where will they find their next 10,000 ML engineers?
At the India-US Forum I attended earlier this year, this point came up in a different way. Those in India’s technology ecosystem insisted that, despite tensions at the political level, existing projects and agreements haven’t been adversely affected. The relationship is more symbiotic than the “cut off” frame suggests. India is not a passive recipient of American AI but a participant in the ecosystem that produces it.
Nobody won electricity, but specific countries found specific leverage points within the supply chain. Switzerland dominated turbine engineering. Sweden dominated high-voltage transmission. Similarly, the Netherlands didn’t win semiconductors. ASML found a single bottleneck, and that one company gives a country of 17 million people more leverage over the chip industry than most countries that manufacture chips.
India needs to find its equivalent. I see three candidates.
The translation layer: Nvidia’s CUDA locks every AI workload to Nvidia GPUs. Indian engineers are among the world’s strongest PyTorch contributors. If India supported open-source projects that let any model run on any GPU—AMD, Intel, Huawei Ascend, Nvidia—that project can become the anti-chokepoint. As inference moves to dozens of specialised chips, this translation layer will become even more valuable.
Sovereign inference chip co-development: India has a growing fabless chip design ecosystem. The opportunity is not to make cutting-edge training GPUs. It’s to co-develop inference-optimised ASICs that run capable models cheaply and at scale. As AI deployment moves to edge devices and domain-specific hardware, this becomes an increasingly large market.
Domain-specific data: Indian agriculture, Indian healthcare, Indian legal corpora, twenty-two official languages: these are datasets no American or Chinese lab can replicate. Fine-tune open-weight models on these, run inference on Indian soil, and we can have practical sovereignty without a government-directed foundation model.
I offered a falsifiable prediction. By 2030, if the gap between the best open-weight model and the best proprietary model is still less than six months, and India is capturing significant value in AI enterprise deployment, then the frontier anxiety was premature, and the diffusion frame was correct.
But if frontier access has been formally tiered, India is stuck on N-2 models, and the capability gap has compounding economic effects, then the opposing perspective would have been more prescient.
My bet is on the first.
Reading and listening recommendations on public policy matters
[Idea] Given its importance, we should have a Scientific and Tech Survey of India tabled in the Parliament every year on the lines of the Economic Survey of India.
[Tool] If you have a public policy question, use this RAG search tool based on 350+ editions of this newsletter. Type any policy question and get the relevant frameworks and what we've written about it, pulled straight from six years of the archive.
[Dashboard] The income tax filing season just finished. If you want to know where your hard-earned money went, check this dashboard by Sarthak Pradhan.
[YouTube Show] The next Watching the Wheels episode discusses frameworks to understand the World Disorder. Don’t miss the Infinite Game and the Trivia I Love sections too.
[Puliyabaazi] We have a blockbuster episode ft. Gurcharan Das talking about all things freedom.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.