On 31 August 2026, the Ministry of Electronics & Information Technology (MeitY) notified Semicon 2.0 – Scheme for the development of Semiconductor Design and Manufacturing Ecosystem in India. While Semicon 1.0 in 2021 arrived as four separate notifications—the Design Linked Incentive (DLI), a compound-semiconductor/ATMP scheme, a display-fab scheme, and a silicon-fab scheme—Semicon 2.0 consolidates the whole thing into a single scheme with six pillars and ten categories, spanning design, materials, fabrication, packaging, R&D and talent.

I had analysed all four 2021 schemes here, here and here. This post revisits those concerns, sets out what has changed, and flags the risks that come with the new regime. My frame of reference remains public policy design, and my motivation is the Indian national interest.

What the Scheme Does

The stated objective is to build a “self-reliant and globally competitive semiconductor chip design and manufacturing ecosystem” that supports economic growth, strengthens national security through supply-chain resilience, and establishes technological leadership in critical sectors.

And to do this, the scope and scale of government involvement have increased with Semicon 2.0. The India Semiconductor Mission (ISM) is now the single nodal agency for the entire value chain, with C-DAC assisting only on the design and deployment categories. The scheme is open for applications for an initial three years, projects run up to six years, and a mid-term appraisal is due after three years. Approvals under ₹100 crore go to the MeitY Secretary; ₹100–500 crore to the Minister; and anything above ₹500 crore, plus all manufacturing categories, to the Cabinet.

The table below gives a one-shot view of all that has changed.

The Gory Details

Pillar 1: Chip Design

Chip design is front and centre of Semicon 2.0 and has three categories.

Category 1: Strategic & National Priorities

This category targets mission-critical silicon (defence, telecom, critical infrastructure) via competitive bidding managed by C-DAC. Crucially, the notification mandates that all resulting IP rights be co-owned by the applicant company and C-DAC.

The Positive: Secures sovereign control over critical IP, mitigating risks of export-control blockades and ensuring state-funded IP cannot be liquidated overseas.

The Risk: Co-ownership with a state entity is structurally incompatible with global venture capital, which requires unencumbered IP to underwrite subsequent funding rounds. This effectively restricts Category 1 to traditional defence contractors or heavily subsidised state consortia. Since C-DAC both runs the Category 1 bidding and co-owns the output, the old conflict-of-interest concern survives in a new form.

Category 2: Commercial Sector Fabless Design

Stated objective: “To position India as a globally competitive hub for fabless semiconductor chip design, fostering innovation, creation of intellectual property and product ownership by way of extending fiscal and chip design infrastructure support, as well as catalysing venture capital funding for domestic semiconductor chip design companies.”

This category offers centralised access to a national EDA tools grid, multi-project wafer fabrication, IP cores, and post-silicon validation. It proposes a hybrid capital stack: Seed Funding (up to ₹15 crore), and Equity Co-Investment or Royalty Financing if demand is higher. The previous strict 50% FDI cap has been removed, making Overseas Citizens of India (OCIs) eligible, provided the company remains “owned and controlled” by citizens/OCIs. Royalty Financing has been included to provide upfront support to firms that might not want to dilute their equity.

• The Positives:

  • Moving away from pure reimbursement grants to flexible equity/royalty structures allows startups to access critical early-stage risk capital. Centralised design infrastructure significantly lowers the initial capex barriers.

  • Allows Indian firms to bid in consortium mode with global companies, R&D organisations, and academic institutions

  • The rules for exiting the financing options are clear, so firms can choose to take the government off their cap table as they scale up.

• The Risks:

  • Equity co-investment and royalty financing turn the government into a quasi-VC holding stakes in private chip-design firms. This raises problems the reimbursement model avoided: how stakes are valued, how and when the government exits, whether public money crowds out or merely rides on private VC. It also puts the state on the cap table of companies it is selecting for incentives and regulates.

  • OCI founders now qualify, and the notification has no hard FDI percentage, which is good news. But the clause has not disappeared. The constraint has shifted from a capital-share line to a control test. A startup can raise substantial foreign VC or PE — the scheme even anticipates this, since Category 2’s equity co-investment beyond ₹15 crore is reserved for companies that have already secured VC/PE funding — but only up to the point where citizens or OCIs still own and control it. That reconciles the government’s understandable reluctance to fund a foreign-controlled firm, but it stops short of the reform I had hoped for - delinking ownership entirely and funding any entity registered in India, on the logic that design capability comes first and home-grown IP follows. Semicon 2.0 adopted half of that.

  • The policy mandates that all IP rights and design files remain strictly within India. This conflicts with the realities of global deep-tech financing, where VCs frequently require startups to “flip” IP to jurisdictions like Delaware or Singapore for cross-border M&A and global licensing.

  • The exit clauses for royalty financing require a steep 1.5x to 2.0x repayment penalty, punishing startups that need to pivot during industry down-cycles.

A Speculation:

Venture capital relies on a power-law distribution where the vast majority of early-stage startups fail, and only a rare few succeed. Conversely, standard Indian public finance rules (overseen by the CAG, CVC, and CBI) treat the loss of public funds as potential impropriety. Thus, bureaucrats negotiating equity valuations will likely be overcautious. Also, deep-tech startups frequently face dilution. State entities lack the legal and administrative agility to approve rapid equity write-downs or bridge rounds, forcing promising startups into bankruptcy.

To mitigate this, I think ISM might avoid direct stock selection (this will only be clear once the scheme guidelines are released). ISM might follow the Research and Development Incentive (RDI) scheme of the ANRF, where capital is routed passively as “automated matching funds” alongside SEBI-registered, private Alternative Investment Funds (AIFs). The private lead investor negotiates the valuation and conducts due diligence, while the state takes non-voting, pari-passu equity shares, shielding public officials from direct audit scrutiny.

Category 3: Deployment-Linked Incentive (DLI)

Offers a 9% reimbursement on net sales for 5 years to help startups offset the costs of commercialising their chips.

• The Positive: The aggressive 9% margin buffer allows nascent Indian chipmakers to compete aggressively on price against established global incumbents during early customer trials.

• The Risk: The DLI subsidises chip supply but does not address demand-side risk aversion. Indian OEMs prefer incumbent foreign vendors due to proven reliability. Without domestic procurement, the subsidy may struggle to generate actual market traction.

Pillar 2: Machines and Materials

This scheme is meant for raw materials and equipment used in a fabrication facility. For firms doing this, the scheme introduces a 30% capex support and a Production Linked Incentive (PLI) ranging from 2% to 10% on domestic BoM-sourced equipment (starting FY29).

The positive: This is arguably the most critical addition to the policy. A fab cannot be globally competitive without a localised, ultra-pure supply chain. Tying the PLI to domestic BoM sourcing aggressively promotes backward integration.

The risk: Another PLI to manage for ISM.

Pillar 3: Silicon and Display Fabs

• Category 5 (silicon wafer fabs: 300mm, 40,000 wafers/month, ₹20,000 crore minimum capital) gets 40% upfront capital support)

• Category 6 (compound/photonics/sensor/discrete fabs) get upfront capital support at 35%, plus an equipment PLI from FY29, and

• Category 7 (OLED/microLED/LCD display fabs) get upfront capital support at 35%

• Assessment: Not much change here apart from a minor change in the quantum of support. But it is not clear to me why display fabs are strategic. But we are well past the point of asking that question given how industrial policy has become the default in a mercantilist world.

Pillar 4: Chip assembly, packaging, and testing

Category 8 establishes a graded structure: 35% for Advanced Packaging (2.5D/3D, chiplets) and 25% for Legacy Packaging.

Assessment: This granularity is excellent. By offering a premium for advanced packaging, India aligns itself with the global shift toward heterogeneous integration, positioning the ecosystem for the next decade of semiconductor evolution. However, the ₹1000 crore minimum investment requirement may lock out agile, mid-sized domestic testing players.

This is a domain in which India has built up some base-level capacity, so it’s good to aim higher here, especially because packaging is where the action has moved in order to save Moore’s Law.

Pillar 5: R&D

Category 9: up to 75% of project cost for developing technologies such as advanced processes, silicon photonics and chiplet technologies for semiconductor companies, either on their own or in a consortium with academic institutions or R&D organisations, etc.

Assessment: A lot will depend on the guidelines because the qualifying criteria are quite vague. Nevertheless, focusing on R&D is overall a move in the right direction.

Pillar 6: Talent Development

Category 10: up to 75% for design-tools access, training infrastructure, and strengthening national nanofabrication labs.

Assessment: Funding opex alongside capex is a major victory. Historically, Indian institutions secured funds for expensive capital equipment but lacked sustained operational budgets to run them effectively. Success here will depend entirely on ISM’s ability to establish low-friction disbursement channels to academic labs.

Bottom line

Semicon 2.0 is a more coherent instrument than its predecessor. There is one agency, money on the table for people and research, and a move towards early-stage risk-taking. But it also expands the state’s role from paying costs to owning equity and IP. The parts that fund talent, research, and shared tools are the parts worth defending. The parts that put the government on the cap table and route billion-rupee decisions through the Cabinet are the ones to watch.

The single-most important leverage point here is ISM’s administrative capacity. Handing ISM the whole value chain — ten categories, equity management, royalty tracking and investing mechanics—is a far larger capacity ask than the 2021 split. The concern I raised about C-DAC now applies, at greater scale, to ISM.

India Semi Watch — September 2026

What’s new

The Indian PM laid the foundation stone for ASIP Advanced Packaging on August 1. This makes it the fifth OSAT plant under construction, with the three (Micron, CG Power, and Kaynes) already operational.

India has a real shot at becoming a competitive packaging hub faster than it will ever build leading-edge logic capacity, so the symbolism here is at least pointed in the right direction. Whether ASIP delivers on the timeline is the question that actually matters, and that’s worth watching closely.

The image from the tracker below provides more detail on ASIP.

On the design front, Marvell announced a $250 million R&D investment across its Bengaluru and Hyderabad centres on July 29. Marvell is not a household name outside chip circles, but it is a well-known firm that has pivoted to AI infrastructure, building custom silicon, interconnects, storage controllers, etc. for hyperscalers.

A commitment of this size to India design centres indicates that Marvell sees India’s chip-design workforce as crucial to its expansion. The caveat is that R&D investment announcements are easy to make and slow to verify. The key metrics to track are headcount growth and the complexity of design work shifting to India.

Timeline check

Delayed:

• Tata Electronics Semiconductor Fab (₹91000cr, Under Construction) — original target 2026-12-31 — formally flagged delayed

• Tata Semiconductor Assembly and Test (TSAT) (₹27000cr, Under Construction) — original target 2026-04-30, revised target 2027-03-31 — formally flagged delayed

On track:

• HCL-Foxconn JV OSAT (₹3700cr, Under Construction) — targeting 2026-12-01

• SicSem Silicon Carbide Fab (₹2066cr, Approved) — targeting 2027-08-01

• 3D Glass Solutions Packaging (₹1943cr, Under Construction) — targeting 2028-08-01

• ASIP Advanced Packaging (₹2500cr, Under Construction) — targeting 2027-08-01

Full facility details at fabs.pranaykotas.com; design-firm tracking at fabs.pranaykotas.com/design.html.

Neither the EU nor India can build their way out of dependence on foreign chips by 2030. The real questions are how each policy framework holds up when that dependence gets tested and how the two can work together to mitigate vulnerabilities. Both call this “AI compute sovereignty”, but neither means the same thing by it.

Hawkins, Lehdonvirta and Wu’s 2025 framework helps explain why. They break “compute sovereignty” into three separate questions — where the data centre sits, who owns the company running it, and who made the chips inside it. A state’s territorial jurisdiction over the data centre is automatic regardless of ownership or chip origin. But what varies is what gets layered on top.

Google’s upcoming data centre in Vizag is a good example of what this framework lays down. India can already regulate it because it sits on Indian soil. But Google is a US company, and laws like the CLOUD Act would apply to data access, regardless of where the servers are. So, apart from Indian jurisdiction, US jurisdiction also applies to the data centre. The third layer is the chips. A domestically owned data centre full of foreign-made chips remains exposed to that supplier’s home government’s export controls whenever further procurement is required.

As discussed below, the EU’s policies largely target the second question of ownership, while India’s target the first.

The EU is trying to enhance its domestic capabilities across the AI infrastructure stack by building legal control around chips and data centres. Meanwhile, India is focused on securing cheap access to compute built abroad for its developers. These two governments are solving for what each can afford, given very different starting resources. While the EU has capital and regulatory reach, India has a large captive market and talent. However, neither has significant fabrication capacity for AI chips and is dependent on foreign players, making both vulnerable to supply shocks.

This creates a curious asymmetry of divergent strengths, a shared vulnerability, and two ways to hedge against it.

The EU treats sovereignty as an ownership test of who holds equity, whose personnel administer the system, and whether data can leave the jurisdiction. The Cloud and AI Development Act (CADA) proposal’s Union Assurance Levels (UALs) formalise this into four tiers, binding public procurement of cloud services to whichever tier a workload requires. Only the top two tiers gate ownership (public sector workloads involving classified or critical infrastructure data), and trusted foreign players remain eligible at the lower tiers where most workloads will sit.

For India, AI compute sovereignty is hinged upon accessibility, i.e. how fast and how cheaply local developers can access compute. Its rules follow that logic, from IndiaAI’s Compute Portal to the data centre tax holiday incentivising foreign hyperscalers. India is incentivising foreign players to build in the country, and this is an accessibility play in Hawkins et al.’s terms.

On chips, both run domestic chip programmes and have industrial policies aimed at advancing domestic capabilities to produce AI chips in the next 5-7 years. They differ in how each treats foreign participation. EU’s Chips Act 2.0 proposes a tiered, reciprocity-gated system within the policy framework itself, where foreign firms qualify for incentives only through strategic partnerships or by localising production as “domestic undertakings,” with access calibrated to how open their home country is to EU firms in turn. This is an ownership play in Hawkins et al.'s terms. ISM 2.0 has no such gating and largely welcomes foreign capital and technology.

The question to ask here is which pathway to AI compute sovereignty would survive supply chain shocks. Both blocs currently import the overwhelming majority of the advanced AI chips they use.

Let’s take the extreme case first: a full blockade on chips from today’s major suppliers. Could either bloc train and run inference at anywhere near the current scale on domestic silicon alone by 2030? Almost certainly not. The EU’s own foundry roadmap, as per the Chips Act 2.0 implementation timeline, doesn’t reach pilot production on advanced nodes until 2030–2033, and ISM 2.0’s 3nm/2nm ambitions sit on a similarly distant timeline.

The more useful question is the less extreme one. Can either diversify and set up significant domestic capacity within five years, given that both will remain import-dependent for at least the next two years? On current trajectories, partially, at best. The EU’s Gigafactories are being built on foreign chips while its chip manufacturing industry matures, while India procures and subsidises foreign chips. However, India’s data centre tax holiday also positions it as a favourable market for supplying cloud services to global consumers, thereby providing leverage against supply shocks. Diversifying suppliers, not replacing them, is the realistic five-year horizon for both.

These are the gaps that both the EU and India are trying to close, and this is where the EU-India Trade and Technology Council can play a significant role. The clearest path is formalising the link between India’s Semiconductor Mission design facilities and the EU Chips Act’s pilot lines around what each side is good at. Alongside that, cross-allocating EuroHPC and IndiaAI’s spare compute pool during demand spikes for suitable workloads and a standing early-warning consultation on compute supply chain shocks can help mitigate short-term vulnerabilities.

Subscribe now

Takshashila’s Space Power Course—an online four-week programme on space policy, law, and strategy—runs 5 September to 10 October 2026.

Apply now

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

Last week, the Mines and Minerals Development and Regulation (MMDR) Amendment Act, 2026 was passed. The Act seeks to reverse the Supreme Court’s nine-judge bench ruling in Mineral Area Development Authority v. Steel Authority of India (MADA), decided in 2024. The 2024 MADA judgement held that royalty paid by mining companies is not a tax, overturning a previous 1989 judgement.

Additionally, it held that State legislatures have absolute and exclusive authority to levy tax on mineral rights, subject only to a law enacted by Parliament. Further, it ruled that mineral-bearing land falls into the category of land, and that it is within the domain of States to tax these lands without any encroachment by Parliament.

The following key features of the MMDR Act merit discussion -

1. It invalidates any uncollected retrospective State taxation on mining companies since the 2024 Supreme Court judgement from the date of its commencement. It requires any tax, cess, or other such levy to be stipulated under conditions guided by the Union government. The act states that any retrospective collection by the States so far remains with them and is not due for a refund to mining companies.

2. It empowers the Union government to regulate the land underlying the minerals, reversing the Supreme Court judgement. More broadly, the amendment seeks to simplify mining taxation in India.

Mining in India seems to have a structure that resembles India’s pre-GST indirect tax architecture, i.e. multiple production-stage levies (royalty, DMF contribution, NMEDT contribution, state cess, auction premium, dead rent, transit fee, GST, etc), with their own base, own administering authority, own compliance process. This creates a cascading effect that raises the effective tax rate - up to 60-65% for major minerals. In light of the previous 1989 Supreme Court judgement, since 1992 the royalty rates had been steeply increased by the Union to compensate States, and the MMDR Act seeks to rein in the States’ ability to further increase the tax burden on mining companies, thereby lowering compliance and distorting costs.

The Laffer Curve shows that maximum revenue from a tax can be achieved at a specific tax rate, i.e., the optimum tax rate. Any rate higher or lower than this reduces the revenues. The current tax rates may be beyond the revenue-maximising range. If there were scope to increase revenue further, we would expect more exploration, more investment, etc. However, exploration remains minimal, output is flat, and reliance on imports is increasing. By reducing the scope for levying additional taxes, the bill aims to shift the sector to the revenue-maximising side.

Many of India’s major minerals are internationally price-linked commodities (iron ore, bauxite, base metals, etc.). Given elastic demand and inelastic supply (i.e., miners cannot simply shift their investments elsewhere), domestic consumers of these minerals can switch to cheaper imports rather than buy from high-cost Indian miners. Thus, the heavy tax burden falls disproportionately on the domestic mining companies. This explains the lack of investment in this sector.

That said, the sector’s growth depends on factors beyond tax rates. Given the market failures associated with mining (i.e., environmental externalities, information asymmetry), governments typically intervene through regulation. However, these regulations can be a source of government failure. The set of individually justified interventions (forest clearance, tribal consent, clearances from state mining departments, explosive-use clearance, etc.) can compound into extremely high transaction costs. These regulations take anywhere between 4 and 5 years in India, compared to 6 months in leading mining economies. Parallel, time-bound clearance can solve this problem while addressing externalities.

Reducing states’ scope to raise revenue from minerals in their jurisdiction is a genuine source of concern for them. The seemingly unilateral Union override can sow seeds of discontent between the Union and the states. More disputes will end up in courts, raising unpredictability in the sector and further dampening investment. A way out could be a GST-Council-style Minerals Council that enables negotiation between the Union and states in a harmonised manner. A stable, predictable tax architecture reduces uncertainty, increasing investment and raising revenue for the states as well as corporate tax revenue for the Union.​

By reducing the scope for taxing mineral land, the bill reduces states’ fiscal capacity to mitigate mining externalities. However, a mechanism already exists: the District Mineral Foundation. The DMF contribution from mining companies is meant to mitigate the negative effects of mining, improve the welfare of communities affected by mining, and ensure mining regions benefit from mineral extraction. However, recent CAG reports point to instances where DMF money was spent on villages not affected by mining, while no projects were implemented in some villages affected by mining. CAG audits in Chhattisgarh found poor planning, unfruitful expenditure and money diversion. Ringfencing the money geographically, ensuring greater control over fund utilisation by the affected communities, outcome-based spending, and a mining community transition fund that helps mitigate the long-term intergenerational impact of mining closure could be the way forward.

Further, the Union Government should prioritise the allocation of public goods such as geoscience data and the financing of merit goods such as infrastructure. Currently, less than 20% of India’s geological potential has been explored. This can de-risk private exploration and incentivise it.

Ultimately, the objective should be to create a mining regime that encourages investment and production, ensures that states have a predictable source of growing revenue, ensures mining communities benefit from mineral extraction, and allows governments to manage the externalities effectively. A stable and predictable tax regime, better-targeted DMF spending, and greater provision of public goods can help achieve this while expanding the revenue base for both States and the Union.

Subscribe now

The politics of the current age are of technology, for technology, and increasingly, by technology. Technology is at the centre of national power, economic growth, and societal change. This raises important questions for all of us:

• What will it take to strengthen India’s own research and development ecosystem?

• How can we build technology responsibly and anticipate its unintended consequences?

• Why are technology and national power increasingly part of the same conversation?

The Graduate Certificate in Public Policy (Tech and Policy) programme is designed to help participants answer these questions. Participants will gain the skills and frameworks to create technology responsibly, advancing the public interest alongside business interests.

Applications close on 23rd August.

Apply now

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

Every few years a new map of Indian innovation is published, and it looks the same. NITI Aayog’s India Innovation Index has put Karnataka at the top of the major states for several editions running. Tamil Nadu, Maharashtra, Telangana and Delhi sit close behind. Bihar, Jharkhand and Chhattisgarh sit at the bottom. The ranking barely moves from one year to the next.

The standard response to that map is to build. When a state lags, the demand goes up for a new IIT, an IISER, an NIT, a CSIR laboratory. The instinct is reasonable. Put a centre of excellence in a place, the thinking goes, and innovation will grow around it. Almost every state now has at least one national institute, and many have several.

However, the data does not support that premise as strongly as the politics around it assumes. It seems to correlate more with the size of the state’s economy, its per capita income, and the structure of the economy.

To understand this better, one must start with what actually predicts where patents are filed across Indian states. The first thing the numbers show is that patenting concentrates in the largest economies, and it does so faster than the economies themselves grow. A state with a bigger economy files exponentially more patents. Innovation, measured this way, is top-heavy by its nature. The biggest state economies pull away from the rest.

The second thing the numbers show is that two features apart from size determine the patent-intensive states. They are richer per head, and their output leans towards services rather than agriculture. This could be the footprint of the information-technology and corporate-research economy. The states that file the most patents for their size are the recognised technology hubs, led by Delhi, Karnataka, Tamil Nadu, Maharashtra and Telangana. A large industrial state that is not research-oriented, such as Gujarat, sits in the middle. The agrarian states trail.

The third finding is the one that should give pause to the build-an-institute reflex. Once you account for the size, income, and structure of a state’s economy, the number of flagship national institutes in that state adds almost nothing to the explanation of its patenting. A state with several national institutes but a small or farm-based economy does not become an innovation centre on the strength of those institutes.

This needs to be read carefully, because it is easy to misread. It does not mean that IITs or national laboratories are useless. They train people, produce research, and do work whose value never shows up in a patent count. The point is narrower and more specific. The simple presence of an elite institute does not, by itself, predict where inventions get filed, because in India patents are filed overwhelmingly by firms, not by the institutes. And firms cluster where the economy is large, prosperous and service-heavy.

This becomes more relevant as India’s patenting grows. The World Intellectual Property Organization’s latest indicators show India among the fastest-growing patent origins, with five straight years of double-digit growth, and resident applicants now accounting for more than half of all filings. The inventive base is becoming domestic. That makes the question of where, inside India, this activity happens a live policy question rather than a statistical curiosity.

The reading is that the geography of Indian innovation follows the geography of the economy, especially its richer and more services-oriented parts, far more than it follows the geography of elite institutions. The map of patents is, to a first approximation, the map of large service-sector economies.

For policy, this tempers a common hope. Locating a national institute in a lagging state may be worth doing for many reasons. It can widen access to good education, anchor a local skill base, and correct a regional imbalance that has political and developmental value of its own. But on the evidence, it should not be expected, on its own, to convert a state into an innovation hub within a planning cycle. The surrounding ecosystem has to be there for the institute to feed into: a base of firms that can absorb graduates, fund research, and file the patents that show up in the count.

That points to a less glamorous agenda than ribbon-cutting at a new campus. Broadening India’s innovative base is, in large part, a matter of broad-based economic development. It means raising state incomes, deepening the services and industrial sectors where research-performing firms operate, and improving the conditions that make a firm choose to locate its research function in a particular city. The patent map will follow the economy, and the way to change the patent map is to change the economy underneath it.

There is a federal dimension here too. The southern and western states that dominate the innovation rankings also dominate on income, urbanisation, and the services share of output. A state that wants to climb the innovation table is, in effect, being told to climb the development table first. That is hard, slow work, and it does not lend itself to a single announcement. It also means that the gap between the top and bottom of the innovation map is unlikely to close through institution-building alone.

None of this argues against research institutions. India needs more of them, and needs the existing ones funded better. The argument is about expectations. If a national institute is placed in a lagging state and judged, a few years later, by how many local patents it has produced, it will look like a failure by a standard it was never going to meet.

The map of Indian innovation is stable for a reason. It tracks where the prosperous, services-rich economy is. States that want to appear on it will have to build that economy, not just host the institutes that the economy, once built, tends to attract anyway. The order matters, and getting it backwards wastes both money and time.

Subscribe now

Takshashila’s Space Power Course—an online four-week programme on space policy, law, and strategy—runs 5 September to 10 October 2026.

Apply Now

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

Causal loop diagram (Author’s Own; Claude was used to brainstorm the diagram)

Takshashila’s State of AI Governance 2025 report reaches a blunt verdict: every major power chose innovation over safety. AI governance has been less about putting on the brakes to build carefully and more about industrial policy. Chips, data centres, and indigenous models now drive policy more than safety does.

The harder question is why this pivot is proving difficult to reverse despite public demand for restraint from civil society, regulators and even some governments. A causal loop diagram helps to untangle the different forces at play in the system.

Creative insecurity, a term coined by Mark Taylor in his book “The Politics of Innovation,” sits at the core of the situation we find ourselves in. It describes a scenario (like the current world order) when external military or economic threats force domestic political groups to overcome rivalries and heavily fund national science and technology innovation.

Creative insecurity powers two self-reinforcing loops. The first is a domestic Build Race, where insecurity drives industrial policy, which pulls in investment, which raises capability, which raises everyone’s insecurity again since the other side is doing the same thing. The second is the Sinking Floor of international governance rules: the same insecurity makes states unwilling to accept binding commitments, which weakens the strength of any agreement they do reach, which erodes the credibility of the regime as a whole. A regime no one trusts offers no assurance, so insecurity stays high. The AI Summits from Bletchley to Seoul to Paris to Delhi traces exactly this descent — broader declarations, thinner commitments, and the two most capable states walking away in Paris. Creative insecurity is the single engine driving two loops: it is simultaneously accelerating the domestic build-out and dissolving global coordination

Against both reinforcing loops stands one balancing loop – the Lagging Brake. Harm becomes visible, the public demands oversight, leading to regulation that bites. In principle, it self-corrects. However, it arrives late in practice because harms surface slowly, and it is actively weakened. Lobbying grows with investment and blunts regulation, while commoditisation of models scatters capable models across thousands of deployers, making it harder to target regulatory interventions.

This outcome is not set in stone. Reinforcing loops can be flipped, and the diagram identifies the levers: a focusing event that jumps the delay (as when Anthropic’s Mythos model exhibited capabilities to autonomously exploit software vulnerabilities, prompting governments to weigh pre-deployment oversight) and strengthening monitoring capacity that enables faster response.

Subscribe now

Takshashila's Space Power Course—an online four-week programme on space policy, law, and strategy—runs 5 September to 10 October 2026.

Apply Now

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

Critical mineral lists have expanded to the point that they no longer work as policy tools. In a frenzy to derisk from China, a trend globally is that these lists are bloating. India’s list runs to 51 elements, about two-thirds of all the elements mined for commercial use. The United States went from 35 minerals in 2018 to 60 in 2025; Russia now tracks 61, Canada 54, Australia 51, and the EU has grown from 33 to 49. If everything is a priority, then effectively nothing is. A label applied this broadly cannot tell a government with limited capital and state capacity where to act, how to act, and which minerals can wait.

In a world of complex interdependence where supply chains have gotten longer and longer the past couple decades, this bloating is understandable. Every dependence by every industry that contributes to exports, or may be deemed “strategic” - another amphibious term, can be considered “critical”. One also has to assume some self-interest that may have contributed to the problem - a place on the list unlocks government subsidy, fast-tracked permits, and the hope of benefits from industrial policy measures for private players.

The problem is that the “critical” label is binary while the problem is not and it needs unpacking to its constituent characteristics. A mineral can be hard to secure because it is geologically scarce, because its refining sits in one country, because it has no substitute, or because its price is weaponised. Each of these calls for a different response.

Tagging lithium as critical does not tell a policymaker whether the vulnerability is a lack of ore or a lack of chemical processing capacity, and so whether to fund exploration or subsidise refineries. Similarly, copper and ytterbium both carry the same label of ‘critical’, but they have little in common, and therefore require differing policy responses. A list only tells a government that a mineral matters. It does not tell it why, where the bottleneck sits, or what would fix it.

What we have done is take India’s own list of 51 critical minerals and pull it apart. We score each mineral on ten measurable vectors against defined, reproducible criteria, one of which is India’s position: how reliant India is on imports for that mineral and how far it has moved to secure supply. The shape of each profile then sorts the 51 into six groups, where a shared profile means a shared bottleneck and thereby, a shared response.

We did not roll these into a composite ranking, because a single score strips out the information that matters, and is exactly the binary problem we are attempting to address: copper, germanium and cerium can add up to the same total while their profiles look nothing alike.

The main takeaway from our work is that the similarity profiles from this exercise sort the 51 minerals into six groups.

Two findings bear directly on how the new partnerships should be designed.

First, a maximalist approach to self-sufficiency (an indigenous mine-to-magnet supply chain) is neither practical nor necessary. A new mine takes more than 16 years on average from discovery to first production, and India’s own mining policy regressed in the decade between 2012 and 2023. Interventions in processing, recycling and value-added manufacturing are more useful, and international partnerships are necessary to enable the actual capacity buildout.

Second, each of the mineral groups maps onto different policy instruments. The heavy rare earths (dysprosium, terbium and their neighbours) score high on refining concentration, reserve concentration, and extraction complexity at once, with substitution still some way off. The good news is that most are consumed in small quantities - only a few hundred to a few thousand tonnes a year - making them extremely amenable for stockpiling efforts. The high-volume base metals (copper, nickel, lithium) are the opposite case: their criticality comes from their high demand and a wide range of commercial applications. Another group of minerals that we call ‘A widening processing chokepoint’ have growing demand and a refining concentration that makes them more vulnerable to economic coercion going forward. Minerals like the Platinum Group Elements are expensive and already have well-developed recycling technologies. Setting up processing and recycling facilities, with offtake agreements selectively applied, are the high-impact policy levers here.

Our work focuses on supply diversification efforts, and does not cover demand-side policy instruments. But the important thing to note is that a complete decoupling from China is neither necessary nor possible. The aim, rather, is narrower: build enough capacity outside China’s control such that its monopoly stops working as a tool of strategic retaliation and leverage. Once that capacity exists, cheap Chinese minerals, propped up by subsidies Chinese taxpayers pay for, make for a good deal rather than imposing a risk.

The work needed is not as difficult as it sounds. And this prioritisation of action, mineral-wise, will be key. Buffers for the small-volume minerals can be assembled in months, while processing and recycling facilities take 3-5 years to be set up anecdotally. For a prioritised set of minerals, enough to achieve ‘de-risking’ from China, that is a reasonable timeline.

Do check out the dashboard here.

Subscribe now

Should India build semiconductor fabs or just focus on chip design talent? On AI, should India double down on its application layer strength or prioritise foundational models?

If you often think about these questions, then we think you may like to take the Technopolitik survey, designed by the Takshashila Institution. It aims to assess what people think about how India should navigate high-tech geopolitics.

Take the survey

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This edition is based on Takshashila Institution’s latest discussion document, “Navigating India’s Subsea Cable Governance”.

Subsea cables are the most critical, and least visible, infrastructure of the modern global economy. Far surpassing satellite networks in speed, reliability, and bandwidth capacity, these fibre-optic lines laid across the ocean floor carry roughly 95% of all transcontinental digital communications, underpinning close to USD 10 trillion in daily international financial transactions. The UN has formally designated subsea cables as critical communications infrastructure, but, India has not.

That omission is the starting point for mitigating the challenges that India’s subsea cable ecosystem faces, and the numbers behind it are stark. India hosts only 18 out of roughly 600 globally, as of 2024. 11 of these 18 were laid more than two decades ago and are approaching the end of their 25-year operational lifespan. As India builds out carrier-neutral data centres and AI infrastructure on the assumption of reliable global connectivity, the physical layer underneath that assumption is ageing without a clear replacement pipeline.

Another significant challenge is that of repair capacity, or the lack thereof. There are currently zero cable repair vessels that are Indian-flagged, Indian-owned, or permanently stationed in Indian waters. When a cable breaks, India depends entirely on foreign repair fleets mobilising from Singapore or the UAE — a dependence that produces average repair timelines of four to five months. In peacetime, that’s an inconvenience. In a period of genuine geopolitical tension, it’s a different problem entirely: foreign commercial vessels may simply decline to enter high-risk maritime zones. Under current arrangements, India’s ability to restore its own cables during a conflict or a coordinated hybrid attack is close to non-existent.

To add to that, most cable landing stations are concentrated around Mumbai. No unified crisis protocol connects telecom operators with the Navy and Coast Guard. No emergency legal framework governs how surviving bandwidth gets allocated once a cut occurs. Individually, each of these is a policy gap but together, they constitute a systemic failure.

Shown below is a causal loop diagram from my discussion document on “Navigating India’s Subsea Cable Governance” to illustrate how these vulnerabilities compound. It is built around a single node: an outage event. Everything else in the model either feeds into that event or responds to it.

India’s subsea cable ecosystem challenges as a causal loop (Author’s Own)

Two reinforcing loops show how the problem compounds. Cable concentration at Mumbai raises the risk of physical damage, which invites grey-zone sabotage, which raises exposure further, making the next outage more likely — a loop that tightens on itself with each cycle. In parallel, India’s dependence on foreign repair ships, combined with time-consuming cabotage licensing rules, slows down restoration — and each slow restoration entrenches that same dependence for the next outage.

Three balancing loops show where intervention can break the cycle, and they map directly onto the paper’s recommendations. Investing in an Indian repair fleet cuts reliance on foreign vessels at the source. A pre-approved Green Channel protocol removes the licensing delay that currently adds weeks to every repair. And declaring subsea cables as Critical National Infrastructure unlocks pre-cleared land for new landing stations outside Mumbai, addressing the concentration problem at its root rather than managing its symptoms.

The paper adds a set of complementary measures to this: distributed acoustic sensing to detect threats to cables before they result in a cut, ownership thresholds that ensure a minimum share of India’s cable bandwidth sits under Indian control, and formal integration between cable fault detection systems and naval and coast guard response — so that an outage triggers a coordinated security response rather than a commercial one.

None of these fixes work in isolation, which is the point of modelling this as a loop rather than a checklist. CNI designation is what gives the state the legal standing to fast-track development, mandate ownership thresholds, and treat a cable cut as a security incident. Without it, the other recommendations have little to no administrative footing.

Read the full paper, including the policy roadmap, here.

Subscribe now

Apply Now!

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

Accenture share price peaked at $417 in Dec 2021 but is now quoting at $137 in July 2026. It is not a stray case but shares of all IT services firms have fallen significantly though none as sharply as Accenture. Does this mark the end of IT services firms as we know them? Accenture, TCS, Infosys and others like them are the companies that global firms across the world have relied upon historically to transform their businesses but these firms themselves seem to have lost the confidence of the investors today. While these companies are still growing as on date both in revenue and profits, the fall in share prices reflects the belief that they may not continue to grow and thrive in the AI era.

Some like Vinod Khosla believe IT service companies may have outlived their usefulness but Nandan Nilekani and others believe AI will drive increased demand for IT services. One thing is certain — the current service delivery model will need to change dramatically. Currently these companies derive 50-60% of their revenue from application development and maintenance of legacy applications. They get another 20-30% from Infrastructure management and business process outsourcing. The balance is from high margin strategy and transformation consulting work. When enterprises needed large teams to implement complex technology, this distribution of revenue made sense and helped these IT service firms grow tremendously. With AI assisted small teams delivering faster and when clients are building these capabilities in-house, the IT services companies are threatened.

AI tools are seen to compress timelines and resource requirements by 30-50% and team sizes are shrinking. The time and materials contracts are fast disappearing and clients are willing to only pay for outcomes. Accenture has launched a mid market solution called Accenture Edge to target clients, in the $300 million to $3 billion revenue size, it earlier ignored and felt would not be profitable. These will be less customised and more platform based solutions at a price point acceptable to the mid market segment. It will definitely put pricing pressure on solutions Accenture offers to its larger clients as well and impact margins if not revenue.

The current incentive structures at these firms are built around maximising billable hours. They are designed to increase the number of consultants who can be billed to clients and the rate charged per consultant hour. This structure does not incentivise building AI tools and assets that can be used across clients while reducing the billable headcount being deployed. New structures that encourage building AI assets and platforms that can be deployed at multiple clients will need to be established.

While there are signs that some of that is happening at these firms, the bigger challenge is the war for talent. Traditionally, the kind of people who are required to build these AI assets have not seen these IT service firms as the place they want to work. These firms have specialised in hiring tens of thousands of good programmers each year and deploying them in a factory model. They will now need to hire a much smaller number of brilliant programmers and while it appears to be easy on paper, it will not be easy in practice to attract such talent. The AI native startups have the edge here. An environment that attracts, nurtures and rewards brilliance in an extremely disproportionate manner will have to be created. These are people who may not be billable to clients but these are the people qualified to build the AI tools. The companies that make this shift will win. These companies are acquiring talent today through acquisitions but that model has a limitation. Instead of expecting the employees of the acquisition to assimilate into the Accenture or TCS culture, the acquiring team from these firms will need to assimilate with the acquired company’s culture.

Another big challenge emerges with these acquisitions besides the cultural change and structural change. The incumbent senior management at these firms has grown within these organizations. A study of the top leadership in these organisations shows that 70% of them have been with these firms for over two decades. These leaders may not be the right set of people to lead these organisations in the new AI era and these firms will have to overhaul the leadership teams without losing the client relationships and domain expertise. Most people in power probably believe they have the chops to run these companies in the new era and may not quietly step aside and that makes this the hardest issue to fix.

This is a classic case of how technological disruption offers new opportunities but incumbents, despite being the ones who could have predicted these disruptions, postpone changes to incentive structures and leadership as needed to make the firms successful. If they act, they make minor changes rather than wholescale institutional overhaul. Institutional change at this scale is hard and the lag period will be painful. However, those who do not adapt will go extinct.

Subscribe now

Apply Now!

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

On June 2, 2026, the Trump administration signed an executive order that creates a framework for a “voluntary” frontier-AI oversight system. Within a few weeks, the government had used that ‘voluntary’ system to shape the release terms of two of the most powerful AI models.

Executive Order 14409’s Section 3 instructs the national security agencies to build a classified benchmark measuring the cyber capabilities of AI models and then to design a framework through which developers may voluntarily ask whether their unreleased model qualifies, grant the government up to thirty days of pre-release access, and help select the “trusted partners” allowed early access. These resemble the components of a licensing regime, however, Section 3(c) declares that nothing in it a “mandatory governmental licensing, preclearance, or permitting requirement.” The problem is that “voluntary,” structured this way, was never going to stay voluntary, and the developments alongside the passing of this order illustrate the point.

The framework can drift toward de facto mandatory because the incentives do the coercing. The government helps choose the “trusted partners” who get early access, turning participation into a competitive asset. The benchmark is classified, so the only way to learn whether your model is “covered” is to engage. CFR’s assessment also notes that US frontier labs will likely participate, “if only to forestall more invasive regulation later.”

The counter-case is that the framework could instead become irrelevant if nothing forces a developer to come forward, and a lab betting the government will not escalate could simply decline. But that seems unlikely seen against the backdrop of recent developments. Anthropic has been labelled a supply chain risk after raising concerns about the use of their models for mass surveillance and autonomous weapons. In mid-June, the Commerce Department directed Anthropic to bar foreign nationals from its new Fable 5 and Mythos 5 models. Unable to reliably comply, the company pulled both of them from the shelf, then saw Mythos partially reinstated for a small group of approved cyber defenders. OpenAI also released its new GPT-5.6 models only to government-approved partners. The direction of travel is toward more control, and now extends to proposals for the government to take equity stakes in the labs themselves.

Over the past two years, the theme of federal regulation in the US has been consistently about helping the US win the AI race. Now it involves an oversight framework. Governance has been reframed from protecting the public from AI to protecting and strengthening the nation through it. The security guardrails have been strengthened while repealing or inverting the ethics requirements around bias, civil rights, and transparency. Recall the executive order on preventing “woke AI” in the federal government and the rebranding of the AI Safety Institutes to focus on standards and innovation. This oversight mechanism that gates access for frontier models, further raises the conversation about AI sovereignty and adds much regulatory uncertainty on who gets access to frontier models.

Subscribe now

Apply Now!

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

Earlier this month, the EU published a proposal for the Chips Act 2.0, intending to repeal and replace the original 2023 legislation, and significantly expand the scope of EU intervention across the semiconductor value chain. The Commission frames the Act in explicitly strategic terms, positioning it as Europe’s solution to supply chain fragility, over-dependence on non-EU suppliers, and the weaponisation of semiconductor access as a geopolitical tool. While the text states that Europe’s tech sovereignty remains “grounded in openness, partnership and fair competition,” it simultaneously introduces mechanisms that structurally privilege EU-based players and create tiered access regimes for foreign companies.

To achieve sovereignty, the Act relies heavily on the concept of “EU added value.” This metric essentially grants preferential treatment, public funding, and fast-tracked permits only to companies that can prove their operations heavily benefit the EU through local job creation, domestic R&D, and supply chain localisation.

Beyond subsidising local supply, the Act systematically drives demand toward European players through three core mechanisms:

• Public Procurement Restrictions - European contracting authorities in critical sectors can now demand exhaustive supply chain declarations from bidders. While the Act acknowledges international free trade agreements, it includes broad security carve-outs allowing authorities to exclude non-EU suppliers entirely.

• Grand Challenges - This mechanism funnels R&D funding strictly toward Commission-defined technological priorities (such as AI and energy-efficient processors) giving selected domestic consortia a state-backed head start.

• Demand Forum and Accelerators - These platforms coordinate early purchasing signals between European chip manufacturers and key downstream sectors like the automotive and defense industries, pre-empting the market before foreign competitors can bid.

These protectionist mechanisms carry a direct competitiveness cost that the EU acknowledges but understates. Forcing European end-user industries to preference EU-sourced alternatives or to absorb the administrative burden of supply chain declarations and security assessments, adds considerable friction to an industrial base already under pressure.

For foreign companies attempting to navigate this insulated market, the Act establishes a tiered access framework. Position in this hierarchy is determined by two factors: the extent of a company’s “EU added value”, and whether a foreign state holds influence over its strategic decisions.

• Full Participants - This tier is open to non-EU-headquartered companies, but it requires genuine localisation rather than a mere sales office. Foreign firms that build local fabs or R&D facilities can meet the “EU value add” criteria, granting them full access to strategic project status and state support.

• Strategic Partners - Foreign companies based in nations that hold formal “Strategic Semiconductor Partnerships” with the EU receive favourable status. Entities in these two tiers are permitted to join the Business-to-Business Semiconductor Supply Chain Platform, where trusted companies share aggregated data to forecast risks and are exempt from direct government data requests during the pre-crisis alert phase.

• Conditional Access - Foreign suppliers lacking a localised footprint or a bilateral treaty face a disclosure-heavy public procurement process. They are entirely locked out of the supply chain platform, must submit detailed ownership declarations, and remain vulnerable to exclusion at the discretion of local contracting authorities.

• Restricted Entities - The Act effectively cuts off meaningful market participation for state-linked or adversarial entities. By scrutinising ownership for decisive influence by foreign states and utilising restrictive IP protection clauses, the framework bars entities that are deemed geopolitical risks.

2026 marks a turning point as India’s first commercial packaging facility entered production in Sanand. However, its capacity remains concentrated in assembly operations and legacy nodes (28nm and above) rather than cutting-edge logic fabs. Additionally, cross-border design collaboration remains bottlenecked by legal friction, as India lacks an EU GDPR adequacy decision. This issue is compounded by a limited EU-recognised R&D presence since much of India’s deep chip design sector is embedded within the global operations of US-headquartered firms rather than autonomous Indian firms with direct EU ties. Hence, India cannot yet satisfy the Act’s “EU added value” criteria.

On strategic partnership, while bilateral cooperation exists under the EU-India Trade and Technology Council (TTC), the two regions have yet to formalise a Strategic Semiconductor Partnership.

The immediate opportunity for India lies in capitalising on its engineering talent and shifting focus toward software and data infrastructure. India controls roughly 20% of the global total of chip design engineers. Since Pillar I of the Act heavily targets research, design, and skills development, Indian engineering service firms and fabless design houses are prime candidates to partner with EU entities.

However, talent alone is not enough to secure institutional access. To secure a seat at the table before the Act solidifies, India must use the TTC to urgently formalise a Strategic Semiconductor Partnership.

Furthermore, as India and the EU share a significant overlap in their identified critical minerals, the TTC provides an ideal vehicle to build joint supply chain resilience. This approach is already in motion. In May 2026, the EU and India launched a ~₹169 Crore joint initiative under TTC Working Group 2, which is explicitly dedicated to developing advanced EV battery recycling technologies. It focuses on recovering strategic minerals like lithium, cobalt, and graphite, directly addressing mutual supply vulnerabilities.

Moving forward, India should expand this momentum within the TTC to pool capital for joint third-country mining acquisitions and to scale up domestic midstream processing. Concurrently, these approaches can position India as a strategic partner in this framework.

Subscribe now

Apply Now!

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

I had the opportunity to spend a week of April in Germany as part of the Auswärtiges Amt (Federal Foreign Office) Germany’s Visitors Programme on “Germany as a Technology Hub”. The programme included journalists, think-tankers, industry leaders, and business development – sixteen participants, each from a different country.

The Visitors Programme’s organises themed trips that enable participants to “see Germany with their own eyes and thus gain an authentic, up-to-date and nuanced insight into the country”. The programme itself seemed to be a great way to showcase the ecosystem, develop soft power, and attract investments and collaborations. I’m sharing some things that stood out for me in terms of what makes the German tech ecosystem work and lessons that it might hold for India.

Fraunhofer And Technology Diffusion

The Fraunhofer-Gesellschaft, Germany’s applied research network, runs 75 institutes, over 30,000 employees, and an annual budget of €3.6 billion. It focuses on Technology Readiness Levels 4 to 7, and makes state of the art research capabilities accessible to the private sector.

Diffusion By Design

Three design choices make this work. A funding model that prices the institute by the market. Roughly one-third of Fraunhofer’s budget comes from federal and state base funding, one-third from industry contracts, and one-third from competitively won public grants. The more an institute earns from industry, the more federal base funding it receives, up to a cap. Industry demand triggers the public match and institutes that cannot attract contracts shrink. This inverts the default in most national R&D systems, where institutional funding is unconditional and industry engagement is an option rather than a requirement.

Competition between the institutes themselves. The 75 institutes operate as profit centres and compete with each other to secure industry contracts and public grants. Underperformance has consequences and protected fiefdoms cannot survive.

Alignment with industry. Fraunhofer research efforts are aligned with the needs of industry. It takes technology to prototype stage and hands it off. While there are spin-offs from some of the research projects, it does not productise, manufacture, or compete with its own clients, which preserves industry’s willingness to pay for its services.

Spillovers: Startups, Returns, GDP

More than 480 spin-offs have come out of Fraunhofer since 2000. Over 7,000 active patent families generate €164 million in annual licensing revenue.

Fraunhofer’s own impact studies estimate that companies cooperating with it are 10.2 percentage points more likely to launch new products. Every €1 of public funding is estimated to return €4 in tax revenue, with GDP effects amounting to 18 times Fraunhofer’s revenue. The numbers come from Fraunhofer-commissioned research; even halved, they are remarkable.

Challenges in Replicating the Model

The model has been admired widely but is challenging to replicate. Two reasons matter for India.

Institutional maturity. Fraunhofer’s competitive funding mechanism depends on the willingness of the state to close or restructure underperforming institutes, and on institutes themselves to operate with accountability and without political capture.

The missing Mittelstand. Fraunhofer works because Germany has thousands of medium-sized engineering firms that need outsourced R&D but cannot afford in-house labs. In India, large conglomerates have their own R&D facilities and MSMEs cannot afford applied research contracts. Without a demand side, no Fraunhofer-shaped institution will find revenue at the one-third industry mark.

The lesson for India is to recognise that institutions of diffusion are as important as institutions of discovery, and to design them for the industrial base we actually have, not the one we wish we had.

On Clusters and Compounding

Saxony has several major semiconductor fabs in dense proximity – Infineon, GlobalFoundries, Bosch and Jenoptik. ESMC, a TSMC-led joint venture with Bosch, Infineon and NXP, backed by €5 billion in German federal funding, is expected to begin operations from 2027. Every third chip produced in Europe is Made in Saxony. The fabs are the visible part but the microelectronics cluster has 3,650 companies and employs 83,000 people across the value chain.

Industrial agglomeration has many benefits: a deep labour pool, specialised suppliers, and knowledge spillovers. All three operate in Dresden. Workers move between GlobalFoundries, Infineon and Bosch, carrying tacit process knowledge in their heads. Equipment suppliers and industrial-gas providers operate at scale because there are enough buyers within driving distance. Fraunhofer institutes and TU Dresden train engineers who can walk to their next job. A “Made in Saxony” reputation signals quality to international buyers without any individual firm having to spend on the branding.

The result is non-linear. The second fab values the talent pool the first one created; the third values the suppliers the first two attracted; the fourth comes for the brand. Clusters are sticky in a way single firms cannot be. A fab can be moved with enough money but a cluster cannot, because what makes it work is a web of relationships that took decades to weave.

India has its own clusters. Tamil Nadu is a manufacturing hub for automobiles and electronics. Bengaluru, Hyderabad, Pune, Chennai, and Delhi-NCR have GCC clusters. Each generates agglomeration benefits at its own scale. It is fascinating to observe what cluster economics delivers when you let them compound over decades. From a talent and supplier ecosystem the compounding returns can evolve into a brand. Their value value grows non-linearly with each addition.

If there is a single thread running through both Fraunhofer and Saxony, it is that Germany’s advantage lies less in any individual lab or fab than in the institutional scaffolding around them. A funding rule that took decades to tune, a cluster whose relationships took decades to weave, both reward patience and time. For India, as we face the temptation is to fund the discovery and treat diffusion as something that follows automatically. The institutions that move technology from prototype to product, and the ecosystems that let one firm’s investment become the next firm’s reason to arrive, are the harder thing to build and the more durable thing to own.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out The Bio-Logic, a newsletter by Anisree Suresh that analyses the economic logic and policy pathways to build India’s USD 300 billion bioeconomy by 2030!

This newsletter is curated by Anwesha Sen.

Four years ago, the US decided the best way to win the AI race was to make sure China couldn’t buy the chips needed to compete meaningfully. However, this week in Beijing, while the press was focused on comments about old trees, Trump made a crucial concession. He acknowledged the unintended consequence of the denial strategy: China no longer wants American chips because it makes its own.

The 2022 embargo forced Chinese firms to achieve hardware independence, producing domestic chips like Huawei’s Ascend that today compete with US peers. Denial, as it has across decades of technological history, acted as a driving force. India was denied this denial. That missing catalyst might exactly be why Indian AI innovation is currently thriving in software, an already existing strength, but is yet to penetrate the historical hardware weakness India has suffered.

Denial as a forcing function

Huawei’s Ascend chips are real and getting better. DeepSeek’s latest model runs on domestic Chinese silicon. The chip denial regime that was supposed to freeze China in place instead gave it the urgency, and the political will, to build an alternative supply chain. The logic is painstakingly simple: restrict something that someone needs, and they will build alternative streams to acquire it, on priority. Pranay and I have made the same argument with the roles flipped when it comes to critical minerals: China is helping make China+1 a reality by restricting exports globally.

India knows this dynamic well. In 1987, the US refused to sell India a Cray supercomputer. Rather than keep asking, India set up C-DAC and gave it three years and a mandate to build something better. PARAM arrived in 1991, capable enough that India eventually exported it across the world, to countries like the UK and Germany.

India’s nuclear programme followed the same cycle: decades of restrictions from the Nuclear Suppliers Group pushed India to crack the full fuel cycle on its own, and the 2008 civil nuclear deal was, in essence, the US admitting that the denial had failed to stop anything.

The pattern is not subtle. When access to something critical is cut off, countries with enough talent and institutional will tend to build it themselves. The pressure is the point.

India’s incomplete arc

So what happened when India briefly faced similar pressure on AI chips?

In January 2025, the Biden administration placed India in a restricted tier for chip imports under the AI Diffusion Rules, with caps on how many it could buy. Indian industry raised the alarm since the restrictions would have pushed AI investment toward countries with unrestricted access. Trump scrapped the rules four months later. In that window, the government pushed harder on the IndiaAI Mission, which funded Sarvam: a large language model trained from scratch in India, on Indian data, specialising in Indian languages.

But it was built on Nvidia GPUs, using CUDA software, with Nvidia providing technical support throughout. Pranay has drawn the relevant distinction in his latest paper: India has supply access to the chip ecosystem, not capability ownership. Supply access depends on external (largely American) policy decisions, which have proven to be volatile. Capability ownership, on the other hand, once developed, cannot be revoked. India solved the easier software problem when the external force existed. When the threat vanished, so did any imperative to tackle the far harder hardware problem. If chip access were threatened again, India’s only move is to lobby for the threat to be removed, as it did in 2025.

China’s refusal to buy US-approved chips is what a completed arc looks like, sovereign models and sovereign hardware together. That is true sovereignty, and that is the north star India must aim for. Not to follow China’s path itself, which it likely cannot replicate, given the difference in state capacity and industrial base, but rather to replicate the urgency that made the speed of development possible. The question, after Beijing this week, is whether India needs to wait for the next crisis to finish what 2025 started, or whether it can find a way to manufacture that urgency before one arrives.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

If you look at the trajectory of global energy systems over the past few years, two competing pressures have become obvious: the need to decarbonise, and the increasing electricity demand from industries such as AI. With intermittent renewables unable to provide continuous baseload power, nuclear energy has re-emerged as an indispensable solution.

Historically, nuclear energy developments have meant conventional gigawatt-scale nuclear power. But traditional nuclear plants are plagued by high capital costs, on-site construction delays, and massive land requirements, let alone significant political barriers.

It is in this context that Small Modular Reactors (SMRs) increasingly seem like a promising solution to meet global energy demands. By definition, they are smaller and modular, which means they can be mass produced and carry a lower capital expenditure per unit. It is also claimed that SMRs are much safer than conventional nuclear power plants as they incorporate passive safety systems (spoiler alert: Gen. III nuclear power plants also already have similar systems in place).

However, pitting SMRs against conventional nuclear power plants is pointless. Both serve different purposes, for the simple reason that the two have vastly different electricity generation capacities.

In India, the power grid enters its most precarious hours when the sun sets. Solar generation collapses while demand surges. Gas-fired plants, which can ramp up and down faster than coal, would ordinarily help bridge the gap — but the ongoing conflict in West Asia has disrupted LNG tanker routes, driven up spot prices, and strained India’s import-dependent gas supply precisely when the grid needs it most.

Coal plants, meanwhile, take hours to throttle up or down. The result is a structural vulnerability that India's rapid expansion of renewables has not yet resolved. Moreover, approximately 38.3 GW of renewable capacity was cancelled between 2020 and 2024 due to tender failures and delays, and over 50 GW of projects remain stranded due to unfinished transmission lines.

This is where SMRs can come in. SMRs offer dispatchable, round-the-clock, low-carbon power that can flexibly complement intermittent renewables like solar and wind.

But the economics of SMRs carry a significant caveat. SMRs inherently lack the economies of scale that make large reactors cost-competitive per unit of output. To overcome this, they must achieve what analysts call high "learning rates" — cost reductions through mass production. First-of-a-kind units will likely cost more than domestic coal or large-scale nuclear alternatives, but the use case differs significantly. SMRs are better suited for captive power generation in industries and complementing renewables for baseload power during demand surges.

Even so, meaningful grid contributions remain at least a decade away.

India’s 2025-26 Union Budget announced INR 20,000 crore for its Nuclear Energy Mission, targeting at least five indigenous SMRs by 2033. The centrepiece is the Bharat Small Modular Reactor — a 200 MW pressurised water reactor being co-developed by the Nuclear Power Corporation of India and Engineers India Limited.

The BSMR’s use of Slightly Enriched Uranium is a strategically sound choice. Some SMR designs require High-Assay Low-Enriched Uranium, a fuel with few commercial producers and significant supplier leverage. SEU avoids those acute bottlenecks, leaving India exposed primarily to generic uranium price swings rather than the geopolitical chokepoints that could constrain rivals’ programmes.

On the manufacturing side, India already possesses the heavy forging capabilities, the turbine assembly lines, and the nuclear-grade electronics sector — with organisations like L&T and BHEL leading in this space.

India is building a diversified partnership ecosystem rather than relying on any single supplier. In March 2026, India and Canada finalised a CAD 2.6 billion, nine-year uranium supply deal with Cameco alongside a formal SMR cooperation agreement. A month earlier, India and France signed a declaration to co-develop and co-produce SMRs and Advanced Modular Reactors. The United States authorised Holtec International in 2025 to transfer SMR-300 technology to Indian firms including Larsen & Toubro and TCS. Russia is in parallel discussions on deploying its own SMR designs and floating nuclear plants in Indian waters.

The decade ahead

SMRs are not a solution to India’s immediate energy stress — that will require accelerating renewables deployment, fixing transmission bottlenecks, and securing more stable gas supplies in the near term. The technology’s commercial maturity is realistically a decade away, and India’s 100 GW target will demand large conventional reactors and massive grid investment running in parallel.

What SMRs offer is something longer in the making: indigenous engineering competence for a technology that low and middle income countries will eventually need. Nations that are too small and capital-constrained for a conventional gigawatt-scale plant are the market a mature BSMR could one day serve.

For a more detailed read, head over to my issue brief, “From Gigawatts to Megawatts”.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

The Nvidia-AMD China chip story shows that export controls are rarely simple or lasting. These controls depend on deals shaped by politics, national security, and business interests. In April 2025, Nvidia was required to obtain export licenses for its H20 chips, which caused a $4.5 billion loss in one quarter. The ban was lifted by July. After that, Nvidia and AMD agreed to pay 15% of their China revenues for access, in a deal reportedly made directly between Jensen Huang and the President, who first asked for 20%. By December 2025, the deal included the H200 chip, and the US share rose to 25%. As of early 2026, Nvidia has not sold any H200 chips in China because Beijing’s approval process is delayed. Both companies and policymakers are facing a changing situation, not a settled policy.

The Conceptual Shift: From Prohibition to Tollgate

In my earlier paper on Technology Denial Regimes, I looked at export controls based on two factors: supply chain control and working with other countries. The four parts of the Imposing Country Strategic Effectiveness Framework - Iron Gate, Fragile Bloc, Paper Wall, and Hollow Threat - all assume the main goal is to block access. The Nvidia-AMD deal challenges this idea.

The US is no longer blocking China from getting advanced AI chips. Instead, it charges for access and takes a share of the profits. This change shifts the approach from blocking access to setting up a tollgate.

Farrell and Newman’s weaponised interdependence theory describes two ways countries use their network power: the panopticon effect, which is about surveillance, and the chokepoint effect, which is about blocking access. The revenue-sharing deal adds a third way, called the tollgate effect, where the controlling country charges for access instead of blocking it. The idea of club goods changes too. Before, there were two groups: insiders with full access and outsiders with none. The Nvidia deal creates a third group: conditional insiders who pay a regular fee.

Who Benefits - And How

By taking 25% of H200 revenues, the US government earns several billion dollars each year without taking on business risks. It also gets information, since every approved Chinese buyer is identified and every shipment is tracked.

It can be said that charging for access lasts longer than outright bans for dual-use technologies. Bans push the targeted country to act quickly, but paid access lowers that urgency. If China can buy H200 chips at a 25% markup, it feels less pressure to speed up its own development. The Sputnik paradox, where a country denied access innovates faster, does not apply if access can be bought. The CFR notes that Huawei does not plan to make an H200-level chip before late 2027.

These benefits depend on certain conditions. First, it should be planned through a clear, multi-agency process, not created under business pressure. Second, it must follow the law - the Export Control Reform Act bans fees on export license applications, and the current tariff workaround is still being challenged. Third, it should involve allies, but the Netherlands, Japan, and South Korea were not consulted, leaving the coalition without a shared reason for their restrictions. Fourth, it should focus on technology that is already widespread enough that a ban would not work - a decision that should come before the arrangement, not after. The current deal meets few of these standards. It was reactive, legally improvised, announced by the US alone, and applied to a chip - the H200 - that the US had recently called a national security risk.

Options for the Targeted Country

In the Targeted Country Adaptive Response Framework, China is seen as a Resilient Transformer. It has strong local research and development and can form other partnerships. The option to pay for access changes how China responds to the situation.

China has logical reasons to accept the current situation. The H200 is almost six times more powerful than the H20, which was the best chip available before this regime. Paying 25% to the US government is cheaper than waiting two to three years for a domestic alternative. The sensible short-term move is to buy access while continuing local development.

China can also refuse. As of early 2026, Beijing has not approved H200 imports under US terms. This could be a negotiation tactic, using delays to push for easier approval. It might also mean China is concerned that agreeing would make the tollgate model permanent.

A third option is for China to act as a Determined Innovator: treat access as temporary and speed up its own development. This can happen alongside buying access.

Evolution or Collapse?

A key question is whether the Nvidia deal is an evolution of technology denial policy, shifting to a more advanced tollgate model, or if it shows the denial regime collapsing under political pressure and only looking like strategic innovation.

The idea of evolution uses the OPEC example. After the 1973 embargo, OPEC found that setting prices lasted longer than outright bans. Bans led other countries to find substitutes, while pricing brought steady profits. From this perspective, the US is learning a similar lesson.

The collapse view is this. The full export ban was lifted in seven weeks, after Nvidia lost $7 billion in one quarter. The percentage was decided not by strategy, but by talks between a CEO and a president. Since the Export Control Reform Act bans fees for export license applications, the deal had to be set up as an import tariff through US territory. This workaround shows the political motives behind the arrangement.

The collapse argument also highlights a major strategic problem. By setting 25% as the price to turn a ban into permission, the US has revealed its minimum price. Now China knows the chokepoint can be bought, not just used as leverage.

The evolution view also affects the coalition. The Biden-era export control group, which includes the Netherlands, Japan, and South Korea, was based on a shared goal: stopping Chinese AI compute for security reasons. That reason falls apart if the US starts selling access for money without consulting its allies. For example, why should the Netherlands now keep limiting ASML’s EUV equipment sales at a big cost?

Implications for Semiconductors and AI

This episode has four lasting effects on how technology denial works in semiconductors and AI.

First, the Iron Gate quadrant is now basically empty for AI chips. It requires both supply chain control and multilateral cooperation. The revenue-sharing model has weakened supply chain control by providing access and damaged cooperation by showing that US policy can change under commercial pressure. The semiconductor denial regime has shifted from Fragile Bloc toward Paper Wall.

Second, the pattern of raising fees is risky. The deal increased from 15% on H20 chips to 25% on H200 chips in just five months. The CFR estimates that if this approach is used for Blackwell-generation chips, China’s AI computing power could grow by over 1000% compared to using only local chips. Setting a percentage fee for each new chip generation matters more than any single deal.

Third, the policy now includes public choice issues. The US government gains financially from keeping the deal in place. This makes it less likely that future rules will become stricter, even if strategy suggests they should.

Fourth, for countries like India, which have historically been targets of technology denial in nuclear, space, and supercomputing, the monetised access model opens a new space for negotiation. A country offering supply chain partnership commitments could negotiate better terms, such as lower or zero revenue share in exchange for semiconductor manufacturing investment. This is a very different way of dealing with technology denial compared to the adaptive strategies I described in my earlier paper.

The Nvidia-AMD case does not settle whether this is evolution or collapse. A future administration with stronger legal backing, better coordination with allies, and the right institutions could create a true tollgate model.

Technology denial regimes only work if they are credible. If a company can negotiate the rules down in seven weeks because of a bad earnings report, the regime loses credibility.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

1. NDRC, not SAMR, blocked the Manus-Meta deal

The proposed acquisition of Manus by Meta was stopped by China’s National Development and Reform Commission (NDRC), not by the State Administration for Market Regulation (SAMR). This distinction is significant.

SAMR is the antitrust regulator. It blocks deals on competition grounds. China is known to use this instrument for national security reasons, but it often does so under the guise of anti-monopoly or consumer harm. You’ll find more about SAMR’s actions in this post.

Anticipating the Unintended

#323 Emotional Integration

Applications for the 43rd cohort of Takshashila’s Graduate Certificate in Public Policy are now open. If you enjoy reading this newsletter, doing the GCPP is the next logical step…

Read more

10 months ago · 25 likes · 3 comments · Pranay Kotasthane

NDRC, by contrast, is the macroeconomic planning body. Its intervention signals that China views this deal not as a competition problem but as a security problem. If NDRC is now the effective gatekeeper for tech mergers & acquisitions, that suggests Chinese AI firms are no longer commercial assets subject to normal market rules. They are national assets, and their disposition is a matter for economic planners, not antitrust lawyers.

Meanwhile, the NDRC had banned two co-founders of Manus from travelling abroad since March this year, citing this regulatory review. Tough times for a Chinese technology firm—foreign companies see you surreptitiously, and even if you manage to convince them, the CPC blocks the sale. This is exactly the problem that rare earth refiners in China have been facing due to the stringent controls over the last 12 months.

2. The wrong lessons from Mythos

The Anthropic Mythos model, whose cyber-offensive capabilities led Anthropic to delay its public release while American companies patched their software, has become the centrepiece of the case for chip export controls. The argument, most prominently made in the Dwarkesh-Huang exchange, runs like this: if China had more compute, it could have trained Mythos first. And that would be unacceptable to the US and the world because of the harm it would cause.

This framing has three problems.

The first is the frame itself. Mythos-type capabilities do not structurally change national power the way nuclear weapons did. Thinking of AI through the nuclear lens, where getting there first and then denying others is the only viable strategy, leads to bad policy. For countries like India, AI is better understood as a vital, general-purpose technology: something you want widely available, not concentrated in one place.

The second problem is the assumption of asymmetry. The argument treats Chinese offensive cyber use as a given while treating American use as defensive. Both countries have offensive cyber capabilities, and both have used them. The real problem is not Chinese malice in particular; it is the absence of credible governance mechanisms for AI deployment, a problem neither Washington nor Beijing has solved.

The third problem is the most important for India. The Mythos episode shows that if AI models can find thousands of unpatched vulnerabilities in major software systems, then every country running those systems is exposed, including India, regardless of who trained the model. The lesson is about defensive readiness and software hygiene, not about chip exports.

3. What made Taiwan think ITRI’s spinoffs should be private?

Karthik Tadepalli has a lovely essay on ITRI, the Industrial Technology Research Institute that seeded Taiwan’s semiconductor industry, in the Asterisk magazine. He finds that ITRI demonstrates how a focused, applied R&D institute can build a world-class chip industry by starting small, training people, and serving firms rather than competing with them.

One question worth thinking over: what led Taiwan’s planners to assume that ITRI’s spin-offs should be private firms? The default in most developing countries, and certainly in India, was the opposite. C-DOT had genuinely good technology, including telephone switches that were competitive by any standard. BEL had developed strategic IP. But neither institution could imagine letting go of the downstream functions. They did engage private players, but merely as “vendors” rather than partners. The IP stayed inside the public sector and did not compound.

Taiwan’s answer to this question is the crucial variable in the ITRI story. It was not just industrial policy; it was industrial policy with a particular theory of ownership and incentives. Getting that theory right is what allowed TSMC to become TSMC rather than a government procurement vehicle.

[Do not miss the essay]

4. Heavy rare earths outside China, for the first time in two decades

Lynas has begun heavy rare earth production at its refinery in Malaysia, reports the Wall Street Journal. Apparently, this is the first time in twenty years that a heavy rare earth refinement on a commercial scale has happened outside China.

This is a direct consequence of Chinese coercion. Beijing’s export restrictions on rare earth processing, imposed as leverage in the broader technology dispute with the US, accelerated the diversification it was meant to deter. China+1 is happening, not because of Western industrial policy alone, but because China’s own actions have raised the cost of dependence.

The Lynas development will not resolve the structural imbalance overnight. China still dominates processing capacity. But the direction of travel is now clear, and it is being driven as much by China’s behaviour as by any Western strategy.

5. Tech companies as military targets

My colleague Nitin Pai makes an uncomfortable argument in a recent essay that tech companies, which have become entangled in military operations, are now legitimate targets under international humanitarian law.

The legal standard is whether an entity makes “an effective contribution to military action.” Palantir’s AI targeting tools, Microsoft’s communications infrastructure for US-Israeli operations, and the sanctions imposed on Russia at Washington’s behest are not incidental. Palantir CEO Alex Karp has said explicitly that tech firms owe the US government their support, given the historical conditions that allowed them to grow. That is a coherent position, but it has consequences. You cannot claim both the returns of geopolitical alignment and the protections of civilian status.

The Indian implication is that the more India’s critical systems depend on foreign tech infrastructure whose owners have chosen sides, the more India’s infrastructure is exposed to the targeting logic of conflicts it has no role in. This is not a hypothetical. It is the architecture of the current moment. Nitin’s conclusion is to reduce dependence on foreign tech for critical systems and increase defence spending to protect what cannot be reduced.

HomeWork

1. [Issue Brief] My paper maps out the causal connections arising from Nvidia selling chips to Chinese firms, but from an Indian lens.

2. [Book] Farrell and Newman have a wonderful book that expands on a concept they developed a few years ago, and that has now become all the rage—weaponised interdependence.

3. [Paper] OECD is producing some fabulous reports on semiconductor fabs. The Dec 2025 report titled 'The Chip Landscape: Geographical Distribution of Wafer Fabrication Capacity' has several great insights relevant to India. My summary is here.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

The irony of a top-secret project on cybersecurity not remaining a secret because of a trivial human error is not lost on the AI community. In March 2026, Fortune magazine in an exclusive report pointed out a data leak at Anthropic. Claude Mythos, according to Anthropic, has been trained on over 10 trillion parameters and heralds a step change in AI performance. Apparently, it is the most capable model they have built to date. While it is being trialled by a few early access customers, Mythos is apparently so powerful that releasing it to the world at large is seen as extremely dangerous by Anthropic. Yet, the documentation on the product was available to the world at large through a data leak.

This product is said to be incredible and has found a zero-day vulnerability in OpenBSD 27 years after it had been released, and a 16-year-old flaw in FFmpeg a memory-corrupting vulnerability in a memory safe virtual machine monitor. Suddenly all software products and platforms appear vulnerable and traditional checks done over decades may seem inadequate.

The company says it has no plans to make Mythos available to public because a bad actor with Mythos can wreak havoc on the world by spotting new vulnerabilities faster than companies are able to fix them. Anthropic has made this product available to a few select customers and those customers have been asked to run vulnerability tests on their systems using Mythos. These early customers are the likes of Microsoft, Google, and Meta.

To demonstrate how powerful Mythos is, Anthropic mentioned a case where “Mythos autonomously came up with a web browser exploit that chained together four vulnerabilities to escape the renderer and operating system sandboxes, gained broad band access and sent an email to the researcher.”

Anthropic therefore launched Project Glasswing providing access to a carefully selected group of large companies to employ Mythos for defensive purposes before hostile actors use it to attack vulnerabilities.

While the intentions to restrict are commendable, there are four arguments that favour releasing the software.

1. Avoiding security inequality – By making Mythos available only to the likes of Apple, Microsoft, Google, etc., one is making security the right of only the largest and wealthiest companies and forcing the small businesses and developing nations to remain vulnerable.

   With much of the world’s infrastructure running on open source, Mythos could help resource scarce developers auto patch their legacy code.

2. Security through obscurity is dumb – It is a matter of time before others like Meta, or state funded labs in other countries develop something similar. It will be difficult to counter a similar model developed in secret by bad actors.

   Releasing the model will help the global white hat community to build defensive tools to match the speed with which threats emerge.

3. Arresting progress is wrong – Mythos is a 10 trillion parameter reasoning engine and has capabilities beyond spotting security vulnerabilities. Holding this back also means curtailing speed of medical research, and green energy research.

   Restricting access to Mythos risks standing by mutely when it can solve many global crises.

4. Hype or reality – Unless Mythos is released, how does the scientific community assess whether the capabilities claimed are real?

   It is also important to understand what it is capable of while establishing benchmarks or creating effective AI regulation.

Instead of completely restricting the release of Mythos, Anthropic could offer tiered access with identity verified, it could hard-code red lines for specific outputs and add other similar guardrails. Release with restrictions is the right answer.

Subscribe now

Technology is no longer just a tool; it’s at the centre of national power, economic growth, and societal change. This raises important questions for all of us:

• What will it take to strengthen India’s own research and development ecosystem?

• How can we build technology responsibly and anticipate its unintended consequences?

• Why are technology and national power increasingly part of the same conversation?

The Graduate Certificate in Public Policy (Tech and Policy) programme is designed to enable participants to answer such questions. It equips technologists, lawyers, civil society, and policymakers with the knowledge and skills to understand the economics of tech platforms, navigate the political economy, find an ethical orientation, and lead the public discourse. Participants will gain the skills and frameworks to create technology responsibly, advancing the public interest alongside business interests.

Apply Now

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

A new parliamentary committee report, presented in March 2026, provides a good overview into the Indian government’s thinking about Artificial Intelligence (AI).

It refers to AI as a “kinetic enabler to leapfrog traditional development barriers and catalyse large-scale socio-economic transformation in the country”. In the tussle between competition and diffusion in a sector increasingly framed by geopolitical contestations, the Indian government’s priorities are clear.

Agriculture: The Kisan e-Mitra chatbot provides real-time scheme information in 11 regional languages, handling over 8,000 queries daily. Additionally, AI-based monsoon advisories have been piloted to assist 3.8 crore farmers across 13 states.

Education: Initiatives like Shiksha Setu in Assam have used AI to record attendance, successfully identifying and removing 4 lakh ghost students. Furthermore, the FutureSkills PRIME program has already trained over 1.68 lakh individuals in AI-related courses.

Employment: AI is projected to create 47 lakh new tech jobs by 2027.

The report provides several examples of how different government ministries and departments are using AI.

As for AI in the military, the report provides several use cases where it is under experimentation or deployed, such as - surveillance and intelligence gathering, language translation on the edge, equipment design, development and maintenance, specialized applications such as cyber threat detection and robotic systems.

But the following challenges are mentioned -

Strategic GPU Dependence: India currently has a strategic dependence on imported accelerators (GPUs) that are critical for AI, which the government aims to break through indigenous initiatives.

Data Constraints: There is a gap in the availability of standardized and “learnable” datasets necessary to build reliable military AI models.

Legacy Systems: Older platforms (tanks, aircraft, and artillery) were not built for AI integration, making upgrades complex and expensive.

Verification Risks: AI/ML techniques are not always amenable to verified decision-making, which may result in unintended outcomes in critical operations.

And specifically, on lethal autonomous weapon systems, the report says this -

One major challenge as submitted by the Ministry is the lack of consensus on how to define the autonomy of weapon systems. A standard definition that accounts for levels of autonomy could help guide an incremental approach to proposing limits. AI ML based techniques are not amenable for verified decision making and hence are likely to result in unintended outcomes as well. The Committee would like to call upon the Ministry to apprise them of the way-out to the challenges that are posed in this field.

Challenges and constraints identified for India’s adoption and implementation of AI are -

Infrastructure Accessibility: There exists a notable disparity in the accessibility of advanced AI infrastructure, particularly between urban and rural regions. Accessibility not only pertains to physical resources but also extends to user interfaces, which must be user-friendly and designed for individuals with varying levels of technical expertise.

Cost Barriers: The high costs associated with advanced computing resources hinder broader participation in AI initiatives, particularly among startups and small businesses that may lack the financial resources for substantial investment in AI technology.

Shortage of Skilled Workforce: A significant challenge is the scarcity of individuals who possess the requisite skills and expertise in AI technologies and data science. This shortage undermines the effective optimization and deployment of AI solutions.

Data Privacy and Security: The assurance of robust data privacy and security measures is imperative for establishing trust among users and for compliance with existing regulations. Navigating the complexities of data governance remains a significant challenge.

Scalability and Future Readiness: As AI technologies evolve and the volume of data increases, there is a pressing need to develop scalable AI infrastructures that are adaptable to future demands. This includes addressing the requirements for both computational power and data management practices.

Institutional Frameworks: The establishment of a comprehensive institutional framework for data governance, as proposed through initiatives such as the National Data Governance Policy (NDGP) and the National Data Management Office (NDMO), is still in its formative stages. These frameworks are essential for ensuring orderly data collection, management, and access, which are critical for maximizing the potential of AI.

Safe and Ethical use of AI: There is a need for responsible AI frameworks for wider adoption of AI in the Indian context and ensuring solutions developed are relevant linguistically and culturally to India and its citizens.

Some recommendations of the report are -

1. Fostering talent. Skilling initiatives to mitigate job displacement and social inequality feature highly. Setting up 570 Data and AI Labs in Tier 2 and Tier 3 cities to provide hands-on exposure to tools and datasets.

2. Preventing AI misuse in financial frauds, intimidation and deepfakes, with a specific mention to the recent attempts at platform governance that aimed to reduce time to action for flagged content from 36 hours to 3 hours, mandate explicit labelling and metadata embedding of synthetically generated content.

3. Exploring age restrictions for certain social media platforms to safeguard citizens from emerging AI-driven risks.

Expect to come across more news chatter around deepfake regulation, platform governance and “AI for India-Specific Regulatory Framework”, as well as more social media restrictions for the under-age, in the coming months.

Do check out the report in full.

And while you are here, do check out the Takshashila Institution’s brand new offering. An Expert Capsule Course on the Politics and Policy of Critical Minerals.

If you are interested in understanding the complexities at the intersection of geopolitics, technology and economics of critical minerals, then this course is for you. Early-bird application deadline ends April 12th.

Link here.

Leave a comment

The world has tentatively watched the recent escalation in West Asia. The war between Israel and Iran has profound implications for the people of Iran, the strategic calculus in the region, and the wider global energy and crude oil market. However, beyond their own material and strategic interests, nations have watched carefully as weapons systems are being battle-tested in real time. Most of the conflict has been restricted to aerial strikes using uncrewed aerial vehicles and ballistic and cruise missiles. As a result, it is often hard to confirm the exact numbers of vectors involved since it is in every country’s interests to portray their air defences as robust and impenetrable. However, we can write with some certainty about the kinds of systems used.

The United States initially relied on missile strikes in its onslaught on Iran. The PrSM (Precision Strike Missile) was fired in combat for the first time. The PrSM is a relatively newer system, known for being harder to intercept due to its speed. The US also used other missiles such as the Tomahawk Block V with an upgraded navigation system, and the AGM-158 JASSM-Er, an air-launched cruise missile. The US also operationalised various drones including the Low-cost Uncrewed Combat Attack System (LUCAS). The LUCAS is a class of UCAVs that was reportedly developed by reverse-engineering the Iranian Shahed-136. At around 35,000$ per-unit, it is a relatively cost-effective option in the United States’ arsenal and carries an explosive payload of 18kg. Subsequent to the significant deterioration of Iran’s air defences, the US has also resorted to using high-volume bombs instead of more expensive missiles.

In its initial strikes on Iran, Israel largely used the Blue Sparrow ballistic missile, which is launched from F-15s and then the LORA– which is a quasi-ballistic missile, to target Iran’s military assets. Israel has also used a range of drones (spanning from long-endurance strike and reconnaissance platforms, to smaller suicide drones).

Iran’s retaliatory strikes have largely relied on one-way attack drones and solid-fuel missiles. This makes sense, given that Iran’s strategy appears to be to try and overwhelm American and Israeli air-defences, in an attempt to force de-escalation perhaps. The missiles used such as the Fateh-110 and the Kheiber Shekan are known for their maneuverability. The UCAVs used by Iran such as the Shahed-136 and Shahed-137 are low-cost drones priced between 20,000$ to 50,000$, and are deployed in large numbers, making the costs of interception high. The LUCAS and Merops American systems are designed to be more affordable and to make these costs more affordable. Interestingly, although there have been no verifiable reports on this, Iran has also supposedly been using un-weaponised decoy drones, forcing enemy air-defences to waste their interception capabilities on cheap targets that cost less than 5000 dollars.

While more sophisticated missiles and fighter jets retain their place for precision strikes deep in enemy territory, drones, and cheaper, attritiable systems are becoming increasingly central to modern warfare. Many of these systems are also equipped with autonomous navigation systems, which makes them impervious to jamming. Their cost-effectiveness also makes them ripe for attempting to overwhelm enemy defences. The increasing use of these systems is also likely to impact the kinds of air defences that are required- perhaps there is going to be a need for air defences that account for the volume of drone strikes.

Subscribe now

In an attempt to understand the use of these various systems, I’ve been working on a tracker (thanks, Claude code!) that looks at drone and missile strikes across West Asia over the course of this war. While this is still very much a work in progress, you can use it to view the kinds of strikes, their origin, whether they were intercepted, and whether there has been an increase or decrease in the intensity of conflict.

This newsletter is curated by Anwesha Sen.

The Union Cabinet’s March 2026 amendment to Press Note 3 was a delayed but necessary correction to prevent domestic manufacturing bottlenecks. A complete blockade on Chinese capital harms domestic infrastructure ambitions more than it punishes Beijing. China invests roughly USD 192 billion as outbound foreign direct investment (OFDI) every year, but India sees barely a fraction of that money. Sensible economic cooperation with the world’s second largest economy is a practical requirement for Indian growth, and this importance is further exacerbated in an age of uncertain US policy towards India.

Press Note 3 was introduced in 2020 and required all investments from land-bordering countries to secure government approval, instead of the automatic channel that FDI inflow usually passes through. The net effect was that Chinese FDI inflow to India fell from USD 242 million in 2018, to just about USD 11 million in 2021. The blanket design treated all Chinese capital as a security risk, which resulted in unintended economic costs.

The new amendment softens this rule. Global funds with less than 10 percent Chinese beneficial ownership can now use the automatic route. In effect, this means that a Singapore private equity firm with minor Chinese backing will be treated like any other international fund, which it was not prior to this revision. The government has also instituted a 60-day approval timeline for critical sectors like electronics and capital goods. This is still a fairly modest step, as most Chinese owned companies as well as the Chinese state are still effectively paralysed from investing in India.

Three takeaways emerge from this saga.

First, the consequences of absolute exclusion are visible in public infrastructure. As we have examined in a Takshashila discussion document published in September 2025, the Bengaluru Yellow Metro Line is a relevant case study. In 2019, a Chinese state-owned firm (CRRC) won the contract to supply train coaches because it submitted the lowest bid. The plan aimed to manufacture the majority of the coaches locally. Restrictions such as Press Note 3 stalled the project completely. They halted the local factory plans, and Customs held up crucial parts. Engineers from China could not get visas to travel. The net cost of this delay was not borne by China or CRRC, but rather by Indian commuters who waited four extra years only to see a crippled service running a mere three trains when it launched.

Second, capturing a larger share of global capital is important for industrial growth. The domestic economy relies heavily on Chinese imports for high and medium-technology goods. Replacing imported finished goods with direct investments in processes is what creates know-how and generates employment. Meeting global manufacturing competition requires domestic upgrades, which would be accelerated with access to Chinese capital and technical knowledge. Building a competitive electronics ecosystem is far more expensive and time-consuming if the world’s largest component supplier is totally excluded.

Third, the recent amendment is an initial step that must expand to cover other high-growth sectors. The electric vehicle space is the most obvious next target. The government previously rejected a USD 1 billion manufacturing proposal from a leading Chinese electric vehicle maker due to security concerns. This move protected local incumbents but cost consumers cheaper cars and blocked critical knowledge sharing.

Supply chain security concerns are not insignificant, but policymakers must find ways to mitigate them while reaping the benefits of lower restrictions. When foreign companies build physical plants on Indian soil, they carry the financial risk. Smart policy shuts out foreign players from sensitive areas like telecom but leaves the door open for heavy industry. Plugging into the global supply chain requires plugging into the Chinese supply chain wherever beneficial to Indian interest.

Ultimately, easing regulations does not guarantee capital inflows. Deep structural disputes between New Delhi and Beijing persist. Moreover, Beijing views regional economic competition strategically and may actively restrict capital exports to a rising India. Chinese actors may simply refuse to participate in the Indian growth story. While amending Press Note 3 is a step towards removing self-imposed domestic barriers, it may not be able to erase the fundamental geopolitical rivalry shaping the continent.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!

This newsletter is curated by Anwesha Sen.

In the summer of 1987, members of the United States Congress gathered on the White House lawn and smashed Toshiba radios and television sets with sledgehammers. Their anger was visceral and, in a narrow sense, understandable. Between 1982 and 1984, Toshiba Machine Company and Norway’s Kongsberg Vaapenfabrikk had secretly sold nine precision milling machines to the Soviet Union for roughly seventeen million dollars - machines that could carve submarine propeller blades to tolerances measured in microns. Soviet submarines had been so notoriously loud that American submariners called them “roaring cows,” detectable from a hundred miles away. After those nine machines, Soviet submarines became twenty times quieter. Detection range collapsed to about five miles. The US Department of Defense estimated it would cost thirty billion dollars to compensate through new detection systems and additional nuclear submarines. Seventeen million dollars in, thirty billion out: a strategic disadvantage ratio of nearly two thousand to one.

This episode - part Cold War thriller, part cautionary tale - sits at the centre of a question my recent research attempts to answer seriously: when do technology denial regimes actually work, and when do they quietly destroy themselves? The answer, as I argue in “A Strategic Assessment of Technology Denial Regimes” (Takshashila Discussion Document 2026-04), is more complicated and more ironic than policymakers typically acknowledge.

What Makes a Denial Regime Work, Or Fail

Technology denial regimes are as old as strategic competition itself, but their modern form - multilateral export controls, supplier cartels, sanctions regimes - took shape in the Cold War. The Coordinating Committee for Multilateral Export Controls, better known as CoCom, ran from 1949 to 1994, making it one of the longest technology control attempts in history. It was never a treaty. It operated as a gentlemen’s agreement, housed in a room in the US Embassy annex in Paris, unknown to most of the public, with no legal force and no enforcement mechanism. That structural informality proved fatal. When France sold ten machines similar to the Toshiba ones to the Soviets and faced essentially no consequences, the moral hazard was never addressed. If one member could defect without cost, why would any other hold the line?

My research identifies two axes that determine whether a technology denial regime succeeds or fails: the degree of supply chain control the imposing country exercises, and the depth of multilateral cooperation it can sustain. When both are high, you get what the paper calls the “Iron Gate” - the near-impenetrable barrier the Nuclear Suppliers Group constructed around uranium enrichment technology after India’s 1974 “Smiling Buddha” nuclear test. That test exposed a remarkable oversight: India had used plutonium from a Canadian research reactor and heavy water from the United States, both supplied for peaceful purposes, to build a bomb. The NSG was founded in direct reaction, and it works because both axes are genuinely strong - the nuclear fuel cycle is concentrated in a handful of countries, and major suppliers coordinate tightly.

When both axes are weak, you get what the paper calls the “Hollow Threat.” The 1990s attempt to restrict strong encryption is the defining example. The US treated robust cryptographic software as a munition subject to export controls. The effort collapsed almost immediately, because - as the paper puts it - cryptographic schemes are just mathematics, and any competent programmer can implement them. There were no chokepoints in the supply chain, no multilateral buy-in, and no realistic enforcement mechanism. Between these extremes sit the “Fragile Bloc” (strong supply chain control, weak coordination) and the “Paper Wall” (strong coordination, dispersed supply chains), each with its own characteristic failure modes.

For targeted countries, what matters, as the other framework in my research paper identifies, is indigenous capabilities of the country, which could be research and development (R&D) capabilities or manufacturing capabilities or any other which provides it ways to counter denial regimes. What is also important is the ability to form alternative partnerships which can help it overcome the impact of these controls, as we will see later in this essay.

And all of this will fail unless feedback loops are incorporated into the entire process. These feedback loops will assess a targeted country’s abilities and state of development, to recalibrate controls. This recalibration could also see the forthcoming of certain incentives to drive the expected behaviour from targeted countries. On the part of the targeted country, this means it should assess its own state periodically to understand whether it can achieve some amount of technological sovereignty which can help it overcome tech denial regimes. If it is not up to the mark, then there have to be strategic pivots towards more partnerships or negotiations.

The Paradox of Strategic Reversal

The most unsettling finding of this research is not that denial regimes fail - it’s the particular way they tend to fail, and who ends up stronger on the other side. The paper calls this pattern strategic reversal, and its most striking historical illustration predates CoCom by a year.

In 1950, the United States imposed the ChinCom embargo on China - a punishment for its entry into the Korean War that was even harsher than the CoCom controls applied elsewhere. US-China trade collapsed from around sixty million dollars in 1949 to less than a hundred thousand dollars by 1952: a 99.8% reduction in three years. The strategic logic was straightforward. China was technologically backward and had no serious industrial base. Isolate it and it would be contained. What actually happened is that the embargo pushed China directly and irrevocably into the Soviet Union’s arms. The Soviets responded with 156 major industrial assistance projects - factories, blueprints, trained engineers, entire knowledge systems transferred wholesale to Chinese institutions. When the Sino-Soviet split came in 1960 and Soviet support evaporated overnight, China had enough of a foundation to go it alone. By 1964, it had nuclear weapons. By 1970, it had put a satellite in orbit. ChinCom helped create the very threat it aimed to prevent.

There is a version of this paradox that recurs across eras with almost eerie consistency. When the Soviet Union launched Sputnik in 1957, the US responded by creating NASA, restructuring science education, and launching Apollo. The country that felt most threatened transformed itself more than the rival did. Today, anxiety about Chinese advances in AI and semiconductors has directly produced the CHIPS Act and a surge in federal research funding. The imposing country keeps getting changed by its own restrictions - often more profoundly than the targeted one.

India as Case Study: From Denial to Competitive Advantage

No country offers a richer empirical record of navigating technology denial than India. It has been on the receiving end of restrictions across nuclear, space, supercomputing, and defence sectors - simultaneously, for decades - and its adaptive responses across these domains are among the most instructive in modern technological history.

The nuclear case is sober. After the 1974 test triggered the NSG, India spent decades building the entire nuclear fuel cycle indigenously - uranium enrichment, heavy water production, reactor design, fuel fabrication - at enormous cost and opportunity. It worked, but slowly and expensively. By 2008, the US essentially acknowledged this reality by signing the Civil Nuclear Agreement, which brought India back into civilian nuclear commerce without it ever signing the Non-Proliferation Treaty. The denial regime didn’t stop India. It made India’s path longer and more expensive, and eventually the regime had to accommodate the reality India had created.

The space case is more optimistic, and more interesting. When the Missile Technology Control Regime cut ISRO off from key propulsion and launch technologies, India didn’t simply try to replicate what was denied. It innovated around the constraint. The result - low-cost satellite launch capabilities that now compete globally and undercut established providers on price - is a textbook example of what Clayton Christensen called disruptive innovation operating under constraint. India’s launch vehicles, developed initially from necessity, became a global competitive product. The limitation became the advantage. In my framework, India occupies the “Resilient Transformer” quadrant: a country with strong indigenous R&D that combines domestic capability with selective alternative partnerships to emerge from denial stronger than before.

The supercomputing case adds another layer. After the 1998 nuclear tests triggered American sanctions, India found itself largely cut off from international supercomputing capability at a moment when it needed it for weather modelling, scientific simulation, and defence applications. The response was slow, expensive, and - eventually - successful. India’s National Supercomputing Mission, launched in 2015 partly in response to these restrictions, has deployed 37 supercomputers totalling around 40 petaflops, with approximately 50% indigenous content. It is a long way from global leadership, but it is no longer dependent.

The Weapons No One Talks About

The conventional image of a technology war involves export control lists, licensing regimes, and customs enforcement - bureaucratic machinery designed to stop boxes from crossing borders. But the most consequential weapons being deployed today operate well beneath that radar.

Financial infrastructure is perhaps the most underappreciated. When Iran was disconnected from SWIFT - the Belgian cooperative that routes payment messages between banks globally - in 2012, it lost the bulk of its oil revenue almost immediately. Not because any technology was denied. Because the payment rails connecting it to global commerce were switched off. The same happened to Russian banks after the 2022 Ukraine invasion. The Committee on Foreign Investment in the United States operates with similar logic but more quietly. Expanded significantly by the 2018 FIRRMA legislation, it can now review minority stakes and joint ventures, operates in secrecy, rarely explains its decisions, and gives the President blanket authority to block transactions on national security grounds. It is, as the paper describes, one of the most quietly effective technology denial tools the US possesses - a scalpel that works without the political theatre of export control announcements.

Standards bodies represent another underappreciated battleground. The country or company that writes the technical standard - that gets its approach embedded into the international baseline - creates a form of de facto technology denial against anyone who wasn’t at the table. China’s aggressive entry into 3GPP, which sets 5G telecommunications standards, means technical requirements beneficial to its ecosystem are now embedded into the global baseline. Late movers face a binary choice: build to those standards or build a separate and incompatible ecosystem. Meanwhile, the Wassenaar Arrangement - which replaced CoCom in 1996 - has become a victim of its own consensus requirement. Any member can block updates to the controlled technology list. Adding new categories like advanced surveillance software has become extraordinarily difficult, meaning the control list perpetually lags behind the technology it’s supposed to govern. The architecture of multilateral cooperation, in other words, can itself become an asymmetric weapon for countries that benefit from institutional paralysis.

A Prediction on the US-China Semiconductor Confrontation

The current US-China technology confrontation in semiconductors and AI is, in several respects, unlike any previous denial regime. But history rhymes, and the frameworks developed in this research suggest a fairly clear trajectory.

On the semiconductor side, the US and its allies have correctly identified the most critical chokepoint: ASML, one Dutch company, makes the only machines in the world capable of manufacturing the most advanced chips using extreme ultraviolet lithography. By coordinating with the Netherlands, Japan, and South Korea to restrict ASML’s equipment exports to China, the alliance has created a genuinely effective short-term barrier. China cannot get those machines, and it cannot build them quickly. The tactical gains over the next three to five years are real.

But the semiconductor case is qualitatively different from nuclear technology denial. Semiconductors are not an end-use military capability with narrow civilian applications. They are central to virtually every dimension of modern economic life. This dual-use ubiquity imposes enormous economic costs on the denial regime - costs that will intensify as supply chains fragment and Western semiconductor firms bleed revenue from restricted Chinese market access. Public choice economics tells us that concentrated industrial interests usually outlast diffuse national security concerns in the policy arena, especially as strategic advantages become less visible and economic costs become more so.

And China, unlike most historical targets of denial, sits in what my framework identifies as the “Maximum Resilience” position: high indigenous R&D capability combined with high access to alternative partnerships. China is already pouring resources into neuromorphic computing, photonic processors, and quantum architectures - the next generation of computing that might render silicon-based manufacturing constraints entirely irrelevant. The US may have used its current chokepoint to catalyse the very technologies that make that chokepoint obsolete.

AI is even harder to contain. Unlike semiconductors, AI is largely software, algorithms, and the mathematical intuitions of researchers. Paul Romer’s endogenous growth theory captures why this matters: knowledge-based technologies have spillover effects built into their nature. You cannot put export controls on what someone learned during a PhD. When Meta released the weights of its LLaMA model publicly in 2023, those weights became permanently available to anyone with an internet connection and basic hardware. A single decision by one private American company effectively undermined a significant piece of US government AI containment strategy overnight.

What History Teaches

The overall conclusion of this research is genuinely humbling for anyone who believes technology denial is a reliable instrument of long-term strategic advantage. The most successful denial regimes in history are those that eventually transformed into cooperative frameworks - acknowledging the shifting distribution of technological capabilities and adjusting accordingly. The 2008 US-India Civil Nuclear Agreement is the clearest example: a recognition that the regime had not stopped India, but that bringing India into a cooperative framework still served both sides’ interests better than continued confrontation.

Technology denial works best when it is narrow, targeted at genuinely critical technologies, sustained by real multilateral cooperation, and built with adaptive feedback mechanisms that allow regular recalibration. It works worst when it is broad, unilateral, static, and aimed at technologies that are inherently difficult to control. The US-China confrontation in semiconductors and AI has elements of both - which is precisely why its trajectory is so difficult to predict, and so important to understand through frameworks grounded in historical evidence rather than geopolitical instinct.

The Toshiba scandal began with nine machines and ended with a thirty-billion-dollar liability. The lesson is not that technology denial is futile. The lesson is that the costs of getting it wrong are paid at a ratio of two thousand to one - and that the regimes which endure are those whose designers understood that lesson before they began.

Subscribe now

The Post Graduate Programme in Public Policy is a comprehensive 48-week hybrid programme tailored for those aiming to delve deep into the theoretical and practical aspects of public policy. This programme emphasises both academic knowledge and hands-on experience, blending conceptual learning with policy simulation exercises. Beyond the online webinars, students also get the added advantage of engaging with Takshashila’s esteemed subject matter experts and its vast alumni community.

Overall, this intensive programme will arm you with all the essential skills for policymaking, analysis, persuasion, and communication, preparing you for your chosen career path in public policy.

Apply Now!

And before you go-

Check out Grammar of War, a newsletter by Adya Madhavan, that looks at advanced military technologies!