The Hand on the Switch

The week began with an accusation and ended with a promise. Early in the week of 5 October 2026, Elon Musk wrote that "oligarchs" were keeping Starlink out of India, and called the delay "A Crime Against the People of India!" [1]. On 8 October he narrowed the target: "Is Ambani the real boss of India?" On 9 October he gave Mukesh Ambani a new title, "Prime Minister Ambani", and asked whether he would "consider allowing Starlink to compete" [1]. By the weekend he had moved to price, vowing that Starlink would be affordable in India because "if Starlink is too expensive in India, no one will buy it" [2].

The timing explains part of the anger. Reliance's Jio Platforms is preparing an initial public offering that targets a valuation of about $106 billion, which would make it the largest listing in India's history [2]. Jio built that value by making Indian mobile data some of the cheapest in the world, at around ₹7.74 a gigabyte [2]. Satellite broadband is the one market where Jio's towers give it no head start, and it is the one market where Musk's company leads the world.

The government replied in the flat, careful language of a ministry that does not want a fight. The Ministry of Communications described its framework as "fair and non-discriminatory". It said the three licensed satellite operators, Starlink and Jio's own satellite venture among them, were at "broadly the same regulatory stage", and that every one of them needed security clearance before it could be given spectrum [1][3]. The telecom minister, Jyotiraditya Scindia, said India allows no monopoly [2].

Musk's complaint is not empty. India decided to allocate satellite spectrum administratively, not by auction. That was the outcome Starlink wanted and Jio fought against [2]. Starlink received its operating licence in June 2025 and its space-regulator authorisation in July 2025 [4]. More than a year later it still cannot sell a single connection. A foreign entrant who has met every requirement it was given has some right to ask why the finish line keeps moving, and the question becomes sharper when the incumbent is about to list at a record valuation.

The finish line has moved for a reason, and the reason is not price. In June 2026 Bloomberg reported that India had frozen the final approvals Starlink needed to begin commercial service. Security agencies under the Home Ministry were said to be withholding clearance. The trigger was reports that Starlink terminals had been used in the Iran war, in a country where the service has no licence. Officials reportedly feared that India would be unable to control a US-based operator during a geopolitical crisis [5]. The ministry's language about "security clearance" is not a formality. It is the whole dispute.

But the quarrel, loud as it is, is about who gets to sell. It is not about the thing being sold. The government's own phrase gives this away. The condition every operator must meet is not a price, a spectrum fee or a market share. It is security clearance. To see why a satellite internet service needs one at all, and why every government that admits Starlink ends up asking the same questions, you have to look at the machine.

Satellite internet used to mean a single large satellite parked in geostationary orbit, about 36,000 kilometres above the equator. At that height a satellite takes exactly a day to orbit, so it seems to hang still in the sky. A dish can point at it and stay pointed. The cost is distance. A signal travelling to that height and back takes more than half a second for the round trip, which is enough to make a video call stutter and a game unplayable.

Starlink reversed the design. Instead of a few satellites very far away, it uses thousands of small satellites very close: most of them orbit at roughly 550 kilometres, about sixty-five times nearer than a geostationary satellite. At that height the delay falls to a few tens of milliseconds, comparable to a fibre line. The catch is that a low satellite does not hang still. It crosses the sky in a few minutes, so no single satellite can serve a home for long. The answer is numbers. As of July 2026 SpaceX had about 10,760 working Starlink satellites in orbit, roughly two-thirds of all active satellites on Earth. It reported about 10.3 million subscribers in 164 countries and territories [6].

Three components make the system work, and each one matters to the argument that follows.

The first is the user terminal, the flat white dish on a roof or a vehicle. It does not move to track satellites. It is a phased array: hundreds of small antennas whose signals are timed so that the combined beam can be steered electronically, jumping from one passing satellite to the next. The terminal is also a computer. It runs SpaceX's software, receives SpaceX's updates, knows its own location, and must be accepted by SpaceX's network before it will carry any traffic. A Starlink terminal is less like a radio and more like a phone tied to a single carrier, and that carrier is in Texas.

The second is the gateway, a ground station with larger antennas that links the satellites to the ordinary internet. In the simplest arrangement, traffic goes up from your terminal to a satellite, down to a nearby gateway, and out into the wider internet. If the gateway is inside your country, your traffic touches your country's soil and its law.

The third component changes that. Newer Starlink satellites carry optical inter-satellite links: lasers that pass data from one satellite to another across space. With these, a satellite above a remote ocean, or above a country with no gateway, can hand its traffic along a chain of satellites to a gateway thousands of kilometres away. The laser links are a real engineering achievement. They are also what makes Starlink politically unusual. Physically, the network does not need to touch the ground of the country it serves.

Put those three parts together and you get the property that every national-security official notices at once. All three are run by one company's software, from one company's network operations centres, under one country's law. That company decides which terminals may connect, in which areas a beam is turned on, and where traffic goes down to the ground. It can make these decisions for a single terminal, a district, or a whole country, and it can make them within minutes, with no one on the ground touching any equipment.

That is a remarkable capability. Whether it is a public good or a threat depends entirely on whose hand is on the switch.

India has already seen the switch used. In late 2024, security forces seized Starlink terminals in two places where Starlink had no licence to operate. One seizure was in Manipur, linked to militant groups. The other was in the Andaman Sea, where a drug-smuggling crew had used a terminal to navigate. Musk responded publicly that "Starlink beams are turned off over India" [7]. The statement was meant to reassure, and it did. It also showed something. Whether Starlink works over Indian territory was, at that moment, a setting chosen by a company in the United States, not by the Indian state.

Ukraine has seen far more of it. When Russia invaded in February 2022, SpaceX shipped thousands of terminals to Ukraine within days. They became, in Musk's own later words, "the backbone of the Ukrainian army" [8]. In September 2022, Ukrainian forces asked for coverage near occupied Crimea to support a naval drone attack on the Russian fleet at Sevastopol. Musk refused. His biographer, Walter Isaacson, first wrote that Musk had secretly switched coverage off near the Crimean coast. Isaacson then corrected himself: coverage had never been enabled there, and Musk had declined to enable it because he judged that it could start a major war [9]. Musk's own account was that he had not deactivated anything, but had refused to activate it [9].

The difference between switching off and refusing to switch on mattered to Musk's reputation. To the argument here it matters very little. Either way, a private citizen of another country decided where a sovereign army could and could not communicate, on the basis of his own reading of the risk of escalation. He may even have been right. That is not the point. No one elected him to make that decision, and the government whose soldiers depended on it had no means to overrule him.

In March 2025, during an argument with Poland's foreign minister, Musk put the matter as plainly as anyone could. Starlink was the backbone of the Ukrainian army, he wrote, and their entire front line "would collapse if I turned it off" [8]. The same day he added that Starlink would never turn off Ukraine's terminals and would never be used as a bargaining chip [8]. It is to his credit that he said so. But a promise not to use a weapon is still a statement that one holds it.

Iran shows the reverse case. In September 2022, during the protests that followed the death of Mahsa Amini in police custody, Musk announced that he would seek sanctions exemptions to activate Starlink over Iran. By December he said nearly a hundred terminals were active there [10]. In January 2026, during another blackout, Iranians again used smuggled Starlink terminals to get news out. The government tried to jam the signals and seized terminals. President Trump said he would speak to Musk about restoring access, and Bloomberg reported that Starlink was offering service free of charge [11].

Many readers will sympathise with the Iranian protesters, and this essay does too. But consider the structure, not the cause. A foreign company, encouraged by a foreign head of state, provided communications inside a country against the explicit wishes of that country's government, and the government could not stop it. A liberal reader cheers when the government is Iran's. The same capability exists over every other country on Earth. It does not check the moral standing of the government below before it works.

So the switch works in both directions. It can deny a country a network it relies on, as near Crimea. It can also give a country's population a network its government has refused, as in Iran. In India's case, the beams were switched off because SpaceX chose to cooperate. The state had no technical means to make it.

Take away the personalities and the problem is general. It has three parts, and any country that admits a foreign-run satellite network meets all three.

The first is availability. A network that can be turned off from abroad becomes, for a country that depends on it, a point of failure that a foreign actor controls. This hardly matters for a farmer's home broadband. It matters a great deal for a border post, a naval vessel, a disaster-relief team or a power grid's backup link. These are exactly the users satellite internet suits best: remote, mobile, and out of reach of fibre.

The second is jurisdiction. Every state requires its telecom operators to support lawful interception: the ability, under warrant, to monitor communications for the investigation of crime and terrorism. A network whose traffic can travel by laser to a gateway abroad can, in principle, carry a country's communications beyond its own courts. The company running it is subject to the courts of another country. American law can compel an American company to hand over data, and it can forbid that company from serving a sanctioned party. Neither power stops at the American border.

The third is uncontrolled access. The Manipur and Andaman seizures showed that terminals cross borders as easily as smartphones do. A network that works anywhere a terminal is carried also works for anyone who carries one, unless the operator chooses to stop it. The state's protection rests on the operator's cooperation.

India's response has been to write conditions into the licence. It has written a great many. Starlink has committed to keep Indian user data in India, to route Indian traffic through Indian gateways, and to build 20 gateway sites in the country. At the India Mobile Congress on 7 October 2026, the week the posts began, SpaceX's Lauren Dreyer said its operations had been "specifically built" for India's security and regulatory needs, with "controls at every layer for India" [12]. These are serious commitments, and the government deserves credit for insisting on them.

But conditions in a licence have limits that no licence can remove. A gateway in Gujarat forces traffic onto Indian soil only as long as the satellites are told to send it there. The satellites, the laser links, the network software and the terminals' firmware are all designed, updated and run from abroad. India can inspect what it is shown. It cannot audit the code it is not shown, and it cannot run the network if the operator walks away. Data localisation protects where the records are stored. It does not change who can turn the service off. The security clearance India demands is the right instinct. But the most any clearance can buy is trust in a partner. It cannot buy control.

Other powers have concluded the same thing, and they are paying heavily to avoid the dependence. The European Union has committed more than €10 billion to IRIS², a sovereign constellation of about 290 satellites in several orbits, because it does not want European government communications to depend on an American company [13]. China is building two constellations of its own, Guowang and Qianfan, which plan more than 13,000 and 15,000 satellites respectively [13]. Neither is close to Starlink's scale, and neither may get there. That has not stopped either power. They are not trying to beat Starlink at selling broadband. They are buying insurance against the day the switch is used against them.

India has begun to buy the same insurance, later than Europe or China and from a weaker starting point. At Reliance's annual general meeting in June 2026, Jio announced plans for a sovereign low-orbit constellation of about 1,600 satellites at roughly 650 kilometres [14]. In July the space regulator, IN-SPACe, working with ISRO and the Department of Telecommunications, judged the proposal "technically sound". It was a technical assessment, not yet a full authorisation, and orbital rights still had to be secured through the International Telecommunication Union [15]. The plan promises 4.5 to 5 terabits per second of capacity over India and 20 to 22 Indian ground stations. Starlink's approval covers 600 gigabits per second [15]. At the India Mobile Congress in October, Jio said it wanted the whole value chain in India, from design to launch to operation. It set out a timeline: a test flight in September 2027, 400 to 450 satellites by 2030 and the full constellation by 2035 [16].

Jio is not alone. Ananth Technologies of Hyderabad holds an IN-SPACe authorisation for a four-tonne geostationary broadband satellite, with service planned from 2028 [17]. Astrome Technologies of Bengaluru has spent a decade designing a constellation of millimetre-wave microsatellites, though it has launched none and is years behind schedule [18]. The industry that a constellation needs is also taking shape. Skyroot Aerospace's Vikram-1 became the first privately built Indian rocket to reach orbit on 18 July 2026 [19]. Dhruva Space now assembles satellites of up to 500 kilograms in Hyderabad [20]. And in October 2026 the Navy invited Indian firms to build the shipboard satellite terminals it currently imports [21].

None of this yet amounts to a working network. Jio has not said who will build its satellites or who will launch them. A 350-kilogram rocket cannot put 1,600 satellites into orbit at any useful pace. ISRO's own broadband satellite, GSAT-N2, went up on a SpaceX Falcon 9 in November 2024 [22]. If Starlink's beams over India are switched on in 2027 [23], they will arrive at the exact moment India's own constellation is due to make its first test flight. India will have admitted a mature network it can regulate but cannot run, at the moment its own is most fragile.

That would not be a new position for India. It would be the normal one.

Consider an ordinary Indian morning, from the phone up.

The phone runs Android, built by Google. In September 2026 Android ran on nearly 94 per cent of the smartphones in use in India, and Apple's iOS on most of the rest [24]. The first search of the day goes to Google, which handles about 97 per cent of Indian web searches [25]. Messages to family go through WhatsApp, owned by Meta, for whom India is the largest market in the world. So do messages to colleagues, the school group, the plumber and, increasingly, small businesses. News arrives through Instagram, Facebook, YouTube and X. At work, the laptop runs Windows or macOS, the email is on Outlook or Gmail, the documents are in Microsoft 365 or Google Workspace, and the company's accounts and supply chain run on SAP, Oracle or Salesforce. The servers are probably rented from Amazon Web Services, Microsoft Azure or Google Cloud.

Nearly every layer of that stack is designed, owned and finally controlled in the United States. A few layers are controlled in Europe. Almost none are controlled in India. This is not the result of any conspiracy. It is mostly the result of good products, network effects, and decades in which buying the best tool on the market was plainly the sensible thing to do. And for most of those decades, nothing went wrong.

It is easy to call that abstract. It stopped being abstract in July 2025.

On 18 July 2025 the European Union adopted its eighteenth package of sanctions against Russia. One of the entities it listed was Nayara Energy, which runs one of India's largest refineries at Vadinar in Gujarat. Rosneft, the Russian state oil company, owns 49 per cent of Nayara [26]. Nayara is an Indian company, incorporated in India, refining crude in India and selling fuel to Indian customers. India did not sanction it, and has never accepted EU sanctions as binding on Indian firms.

Within days, Microsoft cut Nayara off. Its staff lost access to their own email and Teams data, hosted on Microsoft's cloud [26]. Nayara went to the Delhi High Court, and Microsoft restored the service two days later [26]. In the same month SAP's Indian arm suspended support for the software Nayara uses to run its business, including its core enterprise system [27]. Nayara said this left it unable to install the updates needed to comply with India's own GST reforms [27]. On 21 September 2026, more than a year later, the Delhi High Court ordered SAP India to restore the services. It held that the contract was governed by Indian law, which takes precedence over foreign rules, and that the suspension was prima facie a breach [27].

Read slowly, the episode is remarkable. A foreign bloc took a decision about a company in India. India had not agreed to it. Two foreign software firms carried it out within days, by switching off systems the Indian company had paid for and depended on. The Indian company's remedy was to sue in an Indian court and wait. When the court finally ruled, the relief was interim. SAP's argument against restoring the service is worth noting. It was that serving its Indian customer might expose its German parent to criminal liability under EU law [27]. That is the dependence in one sentence: the Indian subsidiary's obligations to Indian customers come second to the parent company's obligations at home.

Nayara found an email provider quickly, and it was Indian: Rediff [26]. That is a small mercy and a useful reminder. Alternatives can be found. But a refinery could do it for its email. A country could not do it for its phones.

Defence is where the cost of dependence is counted in lives, and India has been learning that lesson for a long time.

For decades India has been one of the world's largest arms importers. In 2020–24 it took 8.3 per cent of all global arms imports, second only to Ukraine, which was at war [28]. Russia remained its largest supplier, though its share fell from 72 per cent in 2010–14 to 36 per cent in 2020–24 [28]. That fall is often presented as progress, and in part it is. But much of the business went to France, the United States and Israel, so India changed suppliers rather than escaping dependence. A fighter fleet, a submarine arm or an air-defence network cannot be swapped like an email provider. It needs spare parts, software updates, munitions and maintenance from the country that built it for thirty or forty years. Each purchase also buys a long-term relationship, and with it a certain amount of influence for the seller.

The lesson India remembers best did not involve a weapon. In 1999, during the Kargil war, Indian forces were fighting to dislodge intruders from mountain positions in conditions where knowing exactly where you stood could decide a battle. According to the account that has become part of India's strategic memory, India asked the United States for GPS data covering the region, and the request was refused [29]. Whatever the precise details, the lesson was clear. Satellite navigation, which every modern army now takes for granted, was a service India used at another country's pleasure. That country could withhold it at the moment it was most needed.

India's answer was NavIC, its own regional satellite navigation system, approved in 2006 [29]. It is a real achievement. Few countries have built one. But its story also shows how hard the alternative is. Atomic clocks failed on satellite after satellite. One replacement failed to reach orbit in 2017 and another in January 2025. By March 2026 only three NavIC satellites could provide positioning, one fewer than the minimum needed for an accurate fix [29]. Twenty-seven years after Kargil, the honest description of India's sovereign navigation system is that it is a work in progress.

The same pattern runs through the larger programmes. The Tejas light combat aircraft, India's most visible indigenous fighter, is powered by an American General Electric engine, and delays in engine deliveries held up the production of its newest version. India has made real progress. The guided missiles, artillery and air-defence systems built at home now feature in its own operations, and the government's push for indigenous production has made "Make in India" more than a slogan in the defence factories. But the progress is recent, it is uneven, and in the most demanding technologies, jet engines above all, it is not yet complete.

The pattern repeats across these programmes. When India builds its own systems, they are slower and less capable than what it could import at first. Over time they become something no supplier can take away. When India imports, it gets better equipment sooner and pays for it in something harder to measure.

The same pattern holds in the plainest part of the economy: moving goods.

Around 95 per cent of India's trade by volume travels by sea. Very little of it travels on Indian ships. In 1987–88 Indian-flagged vessels carried about 40 per cent of the country's export and import cargo. By 2019–20 that share had fallen to about 6.5 per cent [30]. The rest moves on ships owned and operated by foreign lines, among them Maersk of Denmark, MSC of Switzerland and CMA CGM of France. India's export and import freight bill is estimated at about $70 billion a year, and most of it is paid to foreign carriers [30].

On land and online the picture is more mixed, but the largest names are still familiar. Amazon is one of the two dominant e-commerce marketplaces in India, alongside Flipkart, which is owned by the American retailer Walmart. Both have built large warehousing and delivery networks of their own. The Indian merchant who sells online and ships abroad often relies, from the storefront to the hold of the ship, on infrastructure owned by companies whose home governments are not India's.

None of this was imposed. The foreign lines were cheaper and more reliable, and the foreign marketplaces had the capital to build what nobody else would. But shipping is where the cost of dependence turns physical. During the pandemic, freight rates multiplied and container shortages stranded Indian exporters while carriers sent their boxes to more profitable routes. A country that owns its fleet can, at the margin, decide what moves first. A country that rents it waits its turn.

The government has noticed. It has tried to reflag ships to India, offered subsidies for Indian-flagged vessels, and set long-range maritime targets. These efforts are welcome and they are late. Rebuilding a merchant fleet is the work of decades, as rebuilding a satellite navigation system turned out to be.

The best evidence that technological dependence can be used as a weapon comes from a system most people have never thought about.

SWIFT, the Society for Worldwide Interbank Financial Telecommunication, is a cooperative based in Belgium and owned by about 2,000 banks. It does not hold money, keep accounts or settle payments. It sends messages: the standardised instructions by which one bank tells another to move funds. It connects about 11,000 institutions in more than 200 countries [31]. Technically it is little more than a secure, very reliable system for sending formatted messages. In practice, almost every international payment begins with one of them.

On 17 March 2012, to enforce European Union sanctions over Iran's nuclear programme, SWIFT disconnected 30 Iranian banks [32]. Nothing had been seized. No money had been frozen by this act alone. The banks simply could no longer send the messages, and so, for practical purposes, they could no longer pay or be paid across borders. Iran's oil exports depend on international payment. They became extremely hard to sell and harder still to be paid for. The disconnection did not end on its own. It ended when Iran agreed to the nuclear deal. When the United States withdrew from that deal and reimposed sanctions, SWIFT suspended Iranian banks again on 5 November 2018, the day American sanctions returned. It called the step "regrettable" but necessary "in the interest of the stability and integrity of the wider global financial system" [33].

Note the sequence in 2018. SWIFT is a Belgian cooperative governed by Belgian law. The European Union had not asked it to act. Europe was in fact trying to keep the nuclear deal alive. But American officials had pressed SWIFT, and its members faced American secondary sanctions [33]. The plumbing was European. The pressure that decided who could use it was American.

In 2022 the same tool was used against a major economy for the first time. After the invasion of Ukraine, seven Russian banks were removed from SWIFT with effect from 12 March 2022, and others followed [31]. That same month, Visa and Mastercard suspended their operations in Russia. Russian cards stopped working abroad, and foreign cards stopped working in Russia. Russia had prepared for this. After the sanctions of 2014 it had built its own card network, Mir, and its own interbank messaging system. Those preparations softened the blow. They did not prevent it.

The lesson for every other country is uncomfortable, and it runs through this essay. A shared system run by a neutral-looking cooperative turned out to be neutral only until a dominant power decided otherwise. The decision did not need a new law in the targeted country, the consent of its government or any physical act. It needed only that the targeted country depend on a system whose controllers answered to someone else.

SWIFT is a messaging network. Starlink is a messaging network as well, of a broader kind. Its owner has already been asked by one government to switch it on over another. It would be naive to assume that no government will ever ask it to switch one off.

One large country took the opposite path. Its example is worth studying, though not copying.

For most of three decades China treated foreign control of core technology as a strategic risk and built domestic substitutes for almost every layer of the stack. Where the world searched on Google, China used Baidu. Where the world messaged on WhatsApp and shopped on Amazon, China built WeChat and Alibaba. When it was cut off from Android's Google services, Huawei built its own operating system, HarmonyOS. China built BeiDou, a complete global navigation system that does not depend on GPS. It built CIPS, a cross-border interbank payment system that reduces its exposure to SWIFT. It is now building two broadband constellations of its own [13]. The approach was not mainly about technology. It was a choice to accept slower, rougher, more expensive products for a time, in return for systems no one else could switch off.

The costs were real, and some were ugly. Much of the substitution was made possible by the Great Firewall, which kept foreign competitors out and also kept Chinese citizens from information their government did not want them to have. It also meant shielding domestic firms from competition and absorbing years of inferior products. A democracy cannot and should not build its stack that way. The substitution worked, but it worked partly because the state was willing to do things India rightly will not do.

What matters for India is that the second half of China's approach does not require the first. Building one's own systems does not need a firewall. It needs the willingness to build, the patience to let a domestic system be worse for a while, and enough public demand to give it scale.

India has already proved that it can do this, in the very field the SWIFT lesson concerns. The Unified Payments Interface was launched in 2016 by the National Payments Corporation of India, a not-for-profit body owned by Indian banks. It is now the largest real-time payments system in the world by volume. A vegetable seller in Patna can be paid instantly by a phone scan through infrastructure that is Indian from end to end. RuPay, India's domestic card network, gives the country a fallback if Visa or Mastercard ever withdraw. Aadhaar, the identity system, and the broader set of public digital infrastructure built on it, were built in India and are now offered to other countries as a model. These are not small achievements. In payments, the exact layer where Iran and Russia were hit hardest, India is in a stronger position today than almost any developing country.

Why did India succeed in payments but not in operating systems, search, messaging, navigation or engines? The answer explains the whole essay. In payments the state chose to build public infrastructure, with open rules that let private firms compete on top of it. In the other layers it waited for markets to produce a national champion, and the markets, quite rationally, bought foreign products instead. Dependence did not hold back Indian creativity by forbidding anything. It held it back by making the foreign option so convenient that building an Indian one never looked worth the trouble. The best Indian engineers built the products of American companies. In an open labour market that is often a sensible career. But the products they built belong to someone else.

The pattern traced in this essay is not history. It is happening again, now, in artificial intelligence. This time it is happening in months rather than decades.

In February 2026 OpenAI's chief executive, Sam Altman, said that India had 100 million weekly users of ChatGPT. That made India its second-largest market after the United States, and the largest in the world for student users [34]. Anthropic opened a Bengaluru office the same month and described India as Claude's second-largest market [35]. The adoption has not been left to chance. In November 2025 OpenAI made its ChatGPT Go plan free for a year to users in India [36]. A few days earlier, Reliance and Google had announced eighteen months of free access to Google's premium Gemini plan for Jio's subscribers [37]. Airtel had already struck a similar arrangement with Perplexity [36]. Within a single autumn, India's two largest telecom companies had become distributors of American artificial intelligence. One of them is the company now asking the country to trust it with a sovereign satellite network.

Enterprises and the state have followed. Cognizant is rolling Claude out to 350,000 employees. Air India, CRED, Razorpay and Swiggy have built on it. The Ministry of Statistics and Programme Implementation has worked with Anthropic on a government data server built to Anthropic's open standard for connecting AI models to outside systems [35]. Each integration is sensible on its own terms. Taken together they are the early stages of the same stack described in Part III, with a new layer on top.

India has not been idle. The IndiaAI Mission, approved in 2024 with a budget of ₹10,372 crore, has subsidised computing power for Indian developers. By July 2026 the government had backed 20 indigenous model proposals, 12 large language models and 8 smaller ones. Five had been released, among them Sarvam AI's models of 30 and 105 billion parameters [38]. These are serious efforts by capable teams. But they are being built for a market that has already been offered the world's leading models free of charge, through the phone plans of its two largest carriers. An Indian model does not compete with ChatGPT on quality alone. It competes with a free year.

This is how convenience works as a mechanism, and it explains every case in this essay. Nobody forbade India from building a search engine, an operating system or a container fleet. The foreign product arrived first, was better at the start, and was cheap or free. Each user who adopted it made it more valuable to the next. In artificial intelligence the effect is stronger still, because use itself is raw material. Every question asked, every document uploaded and every workflow built teaches the leading models more about Indian users, Indian languages and Indian businesses. That knowledge does not go to Indian developers. By the time an Indian alternative is good enough, the market has formed habits, signed contracts and built systems around something else. The alternative is not banned. It is simply never needed, and so it is never finished.

Artificial intelligence also carries its own version of the switch. The leading models run on servers abroad and can be withdrawn by contract or by sanction, as Microsoft showed with Nayara's email. The chips they run on are subject to American export controls. In January 2025 those controls placed India outside the circle of closest allies for advanced AI chips, until the rule was withdrawn that May. A country that builds its public services on another country's models is renting not just a tool but a dependency, at a point in the technology's history when its own capacity could still be built.

The argument of this essay leads to a conclusion India has rarely been willing to draw. Starlink should wait. It should not be admitted to sell satellite broadband in India until a domestic constellation is in operation and able to compete.

This is not a new idea, and it is not a fringe one. It is the argument that almost every industrial power used on its way up. Alexander Hamilton made it for the young United States in 1791. Friedrich List made it for Germany in the nineteenth century. South Korea, Japan and Taiwan practised it in the twentieth, and China in the twenty-first. The economist Ha-Joon Chang has observed that rich countries now tend to recommend free trade to poor ones, having climbed the ladder of protection themselves. He called this "kicking away the ladder" [39]. An infant industry protected for a limited time is not a closed economy. It is how most open economies began.

Satellite broadband is a near-perfect case for the argument, for three reasons. The first is timing. Jio's first test flight is due in September 2027, and Starlink's Indian launch is targeted for early 2027. A network that arrives with 10,000 satellites already in orbit, on the same timetable as a rival that has not yet launched one, does not compete with that rival. It decides the market before the rival exists. The second is lock-in. Satellite customers buy terminals, sign multi-year enterprise and government contracts, and build operations around a particular network. Once those customers belong to Starlink, an Indian constellation must win them back from a position of weakness. The third is the switch. This is not a market for a consumer good. It is a market for national communications infrastructure, and the Home Ministry has already concluded that it cannot control the foreign operator in a crisis [5].

The history of this essay supports the same conclusion. India did not lose search, mobile operating systems, messaging or shipping because foreign firms were forced on it. It lost them because it admitted the foreign product first and expected a domestic alternative to appear later. One never did. Payments are the exception that proves the rule. There the state built the base layer first, the Unified Payments Interface, and foreign firms were allowed to compete only on top of rails India owned. Google Pay and PhonePe, which is owned by Walmart, now handle a large share of UPI transactions. But the system they run on is Indian, and it cannot be switched off from abroad. That is the right model for the sky: build the network first, and let the world compete on top of it afterwards.

Waiting has costs, and an honest argument must name them. The heaviest falls on the people Starlink would serve first: villages beyond the reach of fibre, island communities, and remote border posts. If Starlink is held back, many of them will wait longer for good connectivity, perhaps until 2030. The second cost is political. The Indian constellation that protection would shelter belongs to Reliance, and Musk will say, with some justice, that India is protecting its richest man. A third risk is that Jio does not deliver. A protected incumbent with no competitor and no deadline has every reason to move slowly.

These costs do not defeat the argument. They set the terms on which it is defensible. Four conditions would make it so.

The first is that the wait must be finite. It should end on a published date, or when Indian satellites reach an agreed level of service, whichever comes first. Protection without a deadline is not infant-industry policy. It is a subsidy to an incumbent.

The second is that protection must cover an industry, not one company. The spectrum, ground-station rights and government contracts that protection makes valuable should be open to every Indian applicant: Ananth, Astrome, Bharti's operators and those not yet founded. Jio's network, once built, should carry an obligation to sell capacity wholesale to other Indian providers. The answer to the charge of protecting Ambani is not to deny it. It is to make sure the protection does not belong to him.

The third is that the state should be the first customer. The armed forces, the railways, disaster response and BharatNet together amount to an anchor market large enough to give Indian constellations their first years of revenue. SpaceX itself grew on American government contracts long before it sold broadband to anyone.

The fourth is that the people who cannot wait should not have to. Remote communities can be connected in the interim through Indian geostationary capacity, including GSAT-N2 and Ananth's satellite from 2028, and through the satellite component of BharatNet. Where foreign capacity is truly needed, it should be leased wholesale by Indian operators, through Indian gateways, on contracts that end when Indian satellites are ready. It should not be sold retail to Indian customers under a foreign brand, on foreign terminals that answer to a foreign network.

The same conditions should guide the artificial intelligence that is now arriving. India cannot and should not ban ChatGPT or Claude. They are already part of the daily work of millions of people, and the time to protect that market has passed. But the state can decide what it builds on. Government services, public data and the systems citizens depend on should be built on models India can host, audit and keep running. Public procurement should favour models trained and served in India, as the IndiaAI Mission intends. Satellite broadband is the rarer case, a market the foreign leader has not yet entered. India still has the choice here that it lost in search, phones, messaging and shipping.

Juvenal asked, two thousand years ago, who would guard the guards themselves. The modern version of his question applies to every country that rents the systems it lives on. The hand on the switch always belongs to someone. In the sky above India, it has not yet been decided whose. For once, India can build its own network before deciding whose to let in.