Skip to content
All insights

AI and Institutions

AI Sovereignty for Middle Powers

Author
Sarthak Joshi
Published
Reading time
19 min
Length
4,199 words
Embedded Systems Series · Paper 2 of 5

Strategic dependency

AI Sovereignty for Middle Powers

Few countries can own the AI stack, and ownership is mostly beside the point. What matters is which critical functions keep running when a foreign supplier stops, how long they can wait, and who holds the switch.

In brief

  1. Treat sovereignty as a property of functions: the capacity to keep a critical function running, and to set its terms, when a foreign supplier withdraws or is compelled to. Framework
  2. Domestic or assured capacity is needed only where a critical function rests on a revocable flow that takes longer to replace than the function can wait, from suppliers few legal systems control. Framework
  3. India's cloud market has 3.5 to 6.1 effective firms but at most 2.4 effective jurisdictions. The three leaders answer to one legal system. Derived
  4. Continuity is affordable and the frontier is not. Hold the switches and the stocks, rent the storefront, and build dependence that runs both ways. Inference

How claims are tagged: Evidence Model Inference Hypothesis Framework

A mixed inventory

The programme's study of India's internet sovereignty opens part of its argument with an inventory. India holds the landing stations, the shutdown orders, the payment switch and the grid. It depends on others for the cable corridors, the repair ships, the software most people use and most of its fuel.1 The inventory is useful because it is mixed. It describes neither a sovereign state nor a dependent one, but a state that holds some of the points at which its critical functions could be stopped and not others. That is the ordinary condition of a middle power, and it is the condition in which AI is arriving.

AI adds new layers, and some are more concentrated than anything in the old inventory.

Figure 1

India's AI stack: who holds each layer

Apps and platformsAndroid 92.79% and iOS 7.15% of Indian mobile page views; US-parented apps carried at least about 78% of UPI volume (August 2026). The UPI switch itself is Indian.Switch held
Frontier modelsCome from a small number of developers abroad; no market-share source was found for this layer.Depends
CloudAWS 32.6%, Azure 20.8%, Google Cloud 11.5% of India's market (2024, a single estimate). Globally 28%, 20% and 15% (Q2 2026).Depends
AcceleratorsWidely described as dominated by one vendor; no reliable estimate of the data-centre market was found.Depends
Chip fabricationTSMC took 72.5% of top-ten foundry revenue (Q2 2026); Taiwan made over 70% of advanced-node chips (2022).Depends
Lithography toolsOne Dutch firm holds a monopoly on extreme-ultraviolet machines.Depends
Connectivity18 cables across 23 landings; Mumbai carries about 95% of international bandwidth; no Indian-flagged repair ship.Landings held
EnergyOver 88% of crude imported; 30–53% routed through Hormuz, depending on source and period.Grid held

"Holds" means India controls the point at which the layer could be stopped; "depends" means a foreign firm or state does.

The programme's facts log with its verification status; TrendForce; SIA and BCG; RAND Europe; Competition Commission of India; Synergy Research Group.2 Evidence

Two responses dominate policy debate, and neither is adequate. The first treats sovereignty as ownership and sets out to build everything at home: national fabs, national clouds, national models. For a middle power that target cannot be met at the layers that matter most, and spreading money across every layer underfunds the few where domestic capacity would change outcomes. The second trusts that commercial interdependence will deter anyone from exploiting concentration. Recent events do not support that trust unconditionally.

Figure 2

Eight days without service: a flow cut under someone else's law

18 JulyAn EU sanctions regulation lists Nayara Energy, an Indian refinery.
22 JulyMicrosoft suspends the refinery's email, collaboration tools and access to its own data.
28 JulyThe Delhi High Court issues notice on the refinery's petition.
30 JulyServices are restored, before the hearing. Counsel reportedly could promise nothing about the future.

The refinery's data sat where its data was supposed to sit. What it lost was access to a service, governed by the provider's obligations under a sanctions regime India is not party to. Keeping data at home settles where data sits, not whether the provider keeps serving.

The programme's verified record; the counsel's remark rests on a single report.3 Evidence

What sovereignty should mean

"AI sovereignty" is used in at least four senses: control over data, control over capability, control over the terms of dependence, and the state's control over its own population's use of technology. The last is a different thing, and the vocabulary of sovereignty is often used to justify it. India's record shows why the difference matters. Access Now counted 84 internet shutdowns in India in 2024, second only to Myanmar and the most of any democracy, and 65 in 2025.4 A shutdown exercises sovereignty over a network. It is not the kind this article is about, and a framework that cannot tell them apart ends up recommending more of the first in the name of the second.

The sense used here is functional. A state is sovereign with respect to a function to the extent that the function survives the withdrawal of any single foreign supplier, or any group of suppliers subject to a common coercive instrument, for as long as substitution takes, and to the extent that the threat of withdrawal cannot be used to change the terms on which the function runs. Sovereignty is then a matter of degree, specific to functions, and measurable in principle by asking what stops, for how long, and at what cost. It separates who owns a layer from who can switch it off. Framework

Stocks, flows and flow tails

The first property the test asks about is whether a dependency can be revoked.

Figure 3

Three kinds of dependency

StockPossessed once supplied: chips installed in a domestic data centre, open-model weights downloaded and stored, data held locally, a trained workforce.Can only be degraded
FlowMust keep being supplied: a model reached through an API, rented cloud compute, updates and patches, licence checks, remote management, app-store access.Revocable at will
Stock with a flow tailInstalled, but still reachable: accelerators needing drivers and firmware, phones needing operating-system services, equipment with remote management.Reachable after installation

The two clearest precedents are flow tails. Google suspended Huawei's Android licence in May 2019, and by the second quarter of 2020 Huawei's overseas smartphone shipments were down 27%. On 24 February 2022 attackers remotely wiped Viasat modems, which also cut remote monitoring of 5,800 wind turbines.5 The question to ask about any installed technology is who can reach it after installation.

The criticality test

The test asks five questions of any dependency. Each can be assessed, at least roughly, from information a government can obtain.

  • Revocability. Is it a stock, a flow, or a stock with a flow tail?
  • Substitution horizon. How long would replacing the supplier take, at acceptable performance? Days for an open model already held on local hardware; months to re-host cloud work that was not built to move; years or decades for leading-edge fabrication.
  • Tolerance window. How long can the dependent function run on its fallback before losses become unacceptable?
  • Concentration by coercive jurisdiction. How many legal systems could compel the suppliers? Firms under one jurisdiction form one coercion group.
  • Mutual dependence. What would denial cost the supplier and its state?

How long can a function wait?

The companion study's loss model for a national connectivity outage treats manual fallback as a stock that runs down. On its stated parameters, fallback in financial services loses half its initial capacity by about the third day, and in public administration by about the fifth. Its central case puts the direct loss from a seven-day national outage at $17.1 billion (range $9.6 to 26.0 billion, or $24 to 34 billion with knock-on losses), with the seventh day costing 1.34 times the first.6 Model

Figure 4

The tolerance window, made visible

Direct loss each day, US$ billion

Cumulative direct loss, US$ billion

Low caseCentral caseHigh case

Each day that substitution takes costs more than the day before, until the fallbacks are spent and the daily loss levels off about a third above the first day's. The model describes a national connectivity outage, not the denial of an AI service. It shows the form of the loss curve; no tolerance window for an AI-dependent function has been measured.

The companion study's stated national outage loss model; its fallback parameters are judgements.

How many legal systems hold the switch?

Market-share measures count firms. For sovereignty, the relevant unit is the set of suppliers that one instrument, such as a sanctions regime, an export-control rule or a court order, can compel. Applied to India's cloud layer, the rule changes the count.

Figure 5

India's cloud layer: diversity by firm and by jurisdiction

Effective numbers are the reciprocal of the concentration index. The cited estimate does not break down the 35.1% held by others, so the bars span the possible cases. Jurisdictions reach 1.8 only if all the others form one non-US block, and fall toward 1.0 if some are US-parented, as some may be: the government's empanelment list includes the Indian arms of IBM and Oracle. A procurement policy spread across the three named providers has diversified across firms and not at all across the instrument that can compel them.

Computed from the 2024 shares. Assigning each provider only to its home jurisdiction overstates the diversity, since the Nayara case shows a third jurisdiction's instrument reaching a US provider.7 Derived

Mutual dependence cuts the other way. When US cloud providers stopped new sales in Russia in 2022 and withdrew in steps over some twenty-one months, Russia was, on the companion study's assessment, well under 1% of their business. India is a large and growing market for every one of these firms, and a major source of their engineering labour.8 That lowers the probability of denial. It does not remove the capability, and it gives no protection against a third country's instrument operating on a supplier who does not bear the cost.

Run the test

Figure 6

The criticality test as a decision path

Is the function it supports critical?
Markets sufficeBuy on price and quality.
Is it a revocable flow, or a stock with a flow tail?
Holding it sufficesA stock with no tail cannot be withdrawn once held.
Does replacing the supplier take longer than the function can wait?
A rehearsed substitution plan sufficesWrite the exit and practise it.
Do few coercion groups control it, at little cost to them from denial?
Diversify and use diplomacySpread exposure across jurisdictions; invest in the relationship.
Domestic or assured capacityHold the capacity at home, or secure access that another state cannot revoke.
Answer the questions, or pick an example above.

All four conditions must hold together before domestic or assured capacity is warranted. Framework

The test can also be written as an expected loss, which shows which lever a proposed measure pulls.

Figure 7

Expected loss = P × a × L(min(D, Ts)): each strategy acts on one term

Pchance that denial happens
ashare of the function one denial reaches
Lhow fast losses mount while it lasts
Tstime to substitute the supplier
Alliances and mutual dependencelower P
Diversification across coercion groupsshrinks a
Stocks and fallbacksflatten L
Domestic capacityshortens Ts

D is how long a denial lasts. Writing the loss this way shows whether a cheaper instrument would act on the same term: a held open model flattens L for a fraction of what a domestic frontier lab would cost to shorten Ts.

India's stack, layer by layer

The test was applied to India because the evidence is strongest there and because India combines a large market, deep technical talent and dependence at several layers. The placements below are ordinal judgements, stated so they can be argued with.

Figure 8

Where the test points, on the author's ordinal placements

FlowStock with a flow tailAbove the line: substitution slower than the function can wait
Select a layer on the map to see its assessment.

Points mark the centres of the ranges implied by each judgement. Being above the line is necessary, not sufficient, for domestic capacity: the test's other conditions, coercion groups and mutual dependence, are not shown. Leading-edge fabrication sits off the map to the right, because chips once installed are a stock and denial limits growth rather than continuity.

Ordinal assessments, not measurements. Inference

Chips. TSMC's share and Taiwan's advanced-node dominance point to a substitution horizon measured in decades for any country without an existing leading-edge industry. India's semiconductor mission has approved twelve units worth about ₹1.64 lakh crore: a silicon fab, a silicon-carbide fab, a display fab and nine packaging units, with production begun at three packaging sites in 2026, the first fab due in 2028, and a ₹1.275 lakh crore second phase approved in July 2026.9 Installed chips are a stock; the continuity risk sits in firmware, drivers and replacement. The sensible strategy is inventory and replacement planning for critical-function compute, capability where India can become indispensable (packaging and design), and verification of installed hardware, not a domestic leading-edge fab.

Compute. India's AI mission, approved in March 2024 with ₹10,371.92 crore, has empanelled 45,081 GPUs for subsidised access on an undated count; other counts differ in source and basis. Installed data-centre capacity was about 960 MW in early 2026, roughly 3% of the world's, over a quarter of it around Mumbai.10 Access to accelerators has been conditional on another state's rules. The US diffusion framework of January 2025 left India off its list of eighteen privileged destinations; its rescission was announced four months later, before most requirements took effect, and no replacement governing third countries was found.11 The terms of access were set and withdrawn by decisions in which India had no part. For critical public functions, the test favours inference capacity held at home on hardware whose control channels are audited.

Cloud and data. The government empanels 28 cloud providers subject to data residency. Residency secures the stock, not the flow. Moving 16.68 lakh government email accounts to a system owned and run by NIC, built by an Indian vendor, with its data centres in India and no replication abroad, is a model of selective re-homing.12 Payment data has had to stay in India since 2018, and a wider power under the 2023 data-protection law takes effect from about May 2027. Data is the layer where the stock is most nearly secured and where further localisation buys least.

Models. A proprietary model reached through an API is a pure flow. An open-weight model run on domestic hardware is a stock with a licence tail. Poland built its national models on open foundations from American and French developers, and a Portuguese study found small open models a competitive alternative for public-service functions.13 A held model is frozen at release, so its value for competitiveness decays as the frontier moves. Its value for continuity decays far more slowly, because a model a year or so behind the frontier is usually adequate for a function it was already performing. Continuity sovereignty is affordable. Frontier sovereignty, for almost every country, is not. Inference

Payments, connectivity and talent. The UPI switch is operated by India's national payments corporation while the apps users see are substitutable, though moving the users of the two largest apps could not be done within a week, and the time it would take has never been measured. NPCI has capped any third-party app at 30% of volume from 31 December 2026.14 India holds the switch and rents the storefront. A public service that runs its models abroad adds a cable dependency to its compute dependency; running inference at home removes most of it. And talent is where India is strongest and most reciprocal: about a fifth of the global semiconductor design workforce, and IT and business-process exports that NASSCOM estimates at $246 billion in 2025–26.15

The test applied to India's AI stackOrdinal assessments, to be argued with
LayerStock or flowTime to substituteTime the function can waitTest reading
Leading-edge fabricationStock with flow tailDecadesNot binding for continuityInventory, packaging, verification; not domestic fabs
Accelerators (installed)Stock with flow tailYearsDays to weeksAudit control channels; spares for critical functions
Cloud compute (rented)FlowDays to months (a design choice)DaysPortability mandate; domestic hosting for critical AI functions
Proprietary models (API)FlowMonthsDaysAvoid for critical functions without a held fallback
Open-weight models (held)Stock with a licence tailDaysDaysHold for continuity
DataStockn/an/aPartly secured (payment data; wider power from about May 2027)
Payment and service appsFlowMore than a week (unmeasured)DaysHold the switch, rent the storefront; measure switching time
Mobile operating systemsFlow tails on installed baseYearsVariesAssured access or audited flow tails for critical functions
International connectivityFlowMonths (repair)DaysDomestic inference reduces the dependency
TalentStock that decays without practiceYearsn/aThe main bargaining asset

Four strategies compared

StrategyWhat it changesMain weakness
Full-stack sovereigntyTries to shorten every substitution horizon by building everything at home.Spends most where horizons are longest and least binding, and creates new dependencies: a domestic fab would still need extreme-ultraviolet tools one foreign firm makes.
Selective sovereigntyBuilds domestic capacity only where the test is met.Misclassification and drift: functions become critical over time, so the test has to be re-run on a schedule.
Allied sovereigntyLowers the chance of denial by aligning with states that hold the hubs.Alignment is conditional. In August 2025 the US raised its tariff on Indian goods to 50%, half of it over Russian oil, and removed that half in February 2026; no US threat of technology restrictions was found.16
Diversified interdependenceSpreads exposure across coercion groups and builds reciprocal indispensability.Expensive where one jurisdiction dominates a layer, and slow to build. On this argument, the only strategy of the four that improves bargaining power as well as resilience.

The four are not exclusive. The test implies a mixture: selective sovereignty at the switches and for the inference capacity of critical functions; stocks in place of flows wherever conversion is cheap; diversification where it is available; alliances where it is not; and deliberate cultivation of dependence that runs both ways, in talent, design, packaging and exported public infrastructure. Framework

Objections, briefly

"Sovereignty duplicates what markets supply more cheaply."

Against undifferentiated sovereignty, yes, which is why the test exists. Its specific recommendations, such as portability terms, held open models, audited control channels and spares, are expected to be cheap relative to the losses they bound, though this has not been costed. Hypothesis

"India is too big a market to be cut off."

True as a statement about probability. Market size deters the supplier's own choice, not an instrument imposed on the supplier by a state that does not bear the cost, which is what happened to Nayara.

"Open models solve the problem."

They make continuity cheap for functions current models can perform. The pipeline of future open models is itself a flow that depends on foreign firms and their governments.

"Digital public infrastructure deepens dependence."

One critique argues that promoting India's stack abroad has expanded states' dependence on the largest technology firms.17 If so, a switch surrounded by foreign storefronts, clouds and operating systems can be held in law and hollowed out in practice. That is why the test asks who can reach a function's components after installation, not who owns its central node.

"Sovereignty talk justifies shutdowns."

The most serious objection, and the answer is in the definition. Sovereignty here is measured by the continuity of functions for the population. A measure that raises the state's power to switch off its own citizens' services while doing nothing for continuity under foreign denial fails the test.

What to do

Keep a dependency register

For each AI-dependent critical function: the layers it rests on, stock or flow, time to substitute, time it can wait, the coercion groups, and who can reach its systems after installation.

Convert critical flows to stocks

Hold open-weight models on domestic hardware for critical public functions, keep spares and firmware for their accelerators, and require escrow and perpetual licences where proprietary software is unavoidable.

Count jurisdictions, not vendors

Procurement rules that count suppliers should count coercion groups and shared technical dependencies instead.

Rehearse the exit

Make portability a condition of procurement and test it. A substitution horizon that has never been tried is a guess; a rehearsed exit turns it into a measurement.

Hold the switches

Identity, authorisation, payment, the orchestration of public services and the hosting of models used in public decisions. Rent the storefront, provided its flow tails are audited and switching time is measured.

Build dependence both ways

Invest where others come to rely on India: talent, design, packaging and exported public infrastructure. Coordinate with other middle powers on shared compute reserves and common portability standards.

What this analysis cannot tell you

The test's inputs are assessed ordinally, and most could be measured only by exercises that have not been run; no tolerance window for an AI-dependent function has been measured. Several key figures rest on single commercial estimates or official statements not independently verified. The article treats denial as the main threat and leaves espionage, the other face of hub power, aside. It does not cost the measures it recommends. And it generalises from India on the argument that the structure of dependence is similar; countries without India's market size or talent would find mutual dependence weaker and alliances correspondingly more important.

Middle powers cannot own the AI stack and do not need to. They need the capacity to keep critical functions running when a supplier stops, and that capacity is built at a few points: the switches critical functions pass through, the stocks that can replace revocable flows, the portability that shortens substitution, and the dependence that makes denial costly for the other side.

Notes

  1. Laha et al., India's Internet Sovereignty (working paper P01, 2026), opening summary of Part II. ↩
  2. TrendForce, top-ten foundry revenue, Q2 2026; SIA and BCG (2024), a 2022 figure; Negele et al. (2025), RAND Europe, on ASML's monopoly; Competition Commission of India, market study on AI (October 2025), citing Market Research Future; Synergy Research Group (30 July 2026). StatCounter (August 2026); the US-parented UPI share is derived by the companion study and is a floor. Cable figures from Bhatt (2026), Carnegie India; energy shares from the Petroleum Ministry and trade reporting, which disagree on the Hormuz share. ↩
  3. EU Regulation 2025/1476 (18 July 2025); suspension (22 July); Delhi High Court notice (28 July); restoration before the hearing (30 July), all verified in the programme's record; the counsel's remark is from a single source (CIO.inc). ↩
  4. Access Now, #KeepItOn: Myanmar 85 and India 84 in 2024; Myanmar 95 and India 65 in 2025. ↩
  5. Federal Register and Reuters (May 2019); Canalys (Q2 2020); Viasat, the UK government and Enercon (February 2022). ↩
  6. The companion study's national outage loss model. Half-lives computed from its central parameters: fallback reaches half its initial level on day 1 + τ ln 2, about day 2.7 for finance and day 4.5 for public administration. ↩
  7. Firms: 6.1 if the other 35.1% is fully dispersed, about 4.9 if no unnamed provider is larger than Google Cloud, 3.5 if it were a single firm larger than AWS. Jurisdictions: at most 2.4, 1.8 if the others form one non-US block, 1.0 if all were US-parented. ↩
  8. Company statements (March 2022), reporting on renewals (August 2023) and the EU's twelfth sanctions package (December 2023). The Russian share and India's role as a source of engineers are the companion study's assessments; its facts log gives no source for the first and no figure for the second. ↩
  9. Press Information Bureau, via the India Semiconductor Mission. The split of the approved value between fabs and packaging was not found. ↩
  10. Press Information Bureau and the IndiaAI compute portal (45,081, undated); Syed et al. (2026) for 34,381; Press Information Bureau (February 2026) for 38,000 GPUs, 1,050 TPUs, a 3,000-GPU secure cluster and the data-centre capacity. ↩
  11. Bureau of Industry and Security: the framework (January 2025), the rescission announcement (13 May 2025) and guidance of 31 May 2026. Whether India fell under the per-country allocation also depends on its Country Group D:5 status, which was not checked. ↩
  12. Lok Sabha reply (1 April 2026) and Press Information Bureau (5 December 2025); ₹180.10 crore. MeitY cloud empanelment list (July 2026). ↩
  13. Kocoń et al. (2025) on PLLuM; Branco et al. (2026), which also served a 70-billion-parameter model from one machine with ten GPUs of a US vendor. ↩
  14. NPCI circulars OC-97 and OC-210. The judgement on switching time is the companion study's. ↩
  15. MeitY (22 July 2026) and Press Information Bureau (14 August 2026); NASSCOM Strategic Review (February 2026), estimates. ↩
  16. US Executive Orders 14326 and 14329 (August 2025) and later orders, and the Supreme Court's ruling of 20 February 2026, as recorded in the companion study. ↩
  17. Rikap (2026), Big Data & Society, abstract. ↩

Adapted for general readers from Working Paper P16 of the Embedded Systems Series (September 2026). The working paper carries the full argument, the provenance of every figure and the complete reference list. Placements on the criticality map are ordinal judgements, not measurements.