hello@tonishatagoe.com Abu Dhabi · London · Accra · New York
Digital Sovereignty

India Drew a Line Around Its Own Machines. Who Decides What Crosses It?

India is drafting a sovereign AI map that would place the most sensitive national workloads on government-owned, air-gapped hardware while leaving the rest on state-owned machines run by contractors. The tiering logic, not the hardware, is the governance instrument that matters.

India Drew a Line Around Its Own Machines. Who Decides What Crosses It?

India is drawing a sovereign AI map that will sort national data into tiers and decide, sector by sector, who is allowed to own the machines that think. According to reporting in The Economic Times on 20 October 2026, the most sensitive tier will be on-premises, entirely government-owned and air-gapped, with defence among the candidate domains. A second tier will keep state ownership of the hardware while private companies operate it. The plan is still being drafted and no timeline has been published.

The significance is not that a government wants its own computers. It is that a state has begun to draw a boundary inside its own economy and declare that some intelligence-bearing infrastructure may not be rented, shared, or reached from outside. That is a different kind of policy from a data residency rule. It is an attempt to name the point at which dependency becomes unacceptable, and then to build around that point before the dependency hardens.

Featured image: file picture used by The Economic Times with its report of 20 October 2026 on India’s proposed sovereign AI infrastructure tiers.

What the original report says

The Economic Times reported on 20 October 2026, in a piece by Subhayan Chakraborty and Surabhi Agarwal, that the Government of India is drawing up a sovereign AI map to identify critical sectors and data categories. People with knowledge of the matter said the most sensitive category of technology infrastructure, including GPUs, would be entirely on-premises, wholly owned by the government and air-gapped. Defence was named as a likely member of that tier. A second and less sensitive but still important category would be governed by a separate framework in which the state may own the hardware while private companies run and operate it. Officials did not disclose a timeline, and the plan remains at the drafting stage.

THE SCENARIO: A ministry has classified one of its AI workloads as critical. The vendor that has supplied the model for two years offers a regional cloud with local storage and a European data-protection addendum. The procurement team considers the matter settled. Then a legal adviser asks the only question that matters: if the vendor’s home government restricts that model tomorrow, can the ministry keep running the service in the same building, on equipment it controls, with staff it employs? Nobody in the room can answer. The contract was written to govern data location. It was never written to govern who holds the switch.

The Question

The question is not whether governments should own more computing hardware. Most cannot afford to own enough of it, and most would run it badly if they tried. The question is whether an institution can name the specific decisions that must remain inside its own perimeter, and then prove that they do.

India’s draft map is interesting because it proposes a graded answer rather than a single sovereign switch. Some workloads sit in an air-gapped room with government-owned silicon. Others run on state-owned machines operated by contractors. The rest continue on commercial cloud. That gradation is a governance instrument, not a procurement list. It is a statement that sovereignty is not one condition but a set of thresholds, applied workload by workload, with different controls at each level.

That matters because the alternative, an all-or-nothing argument about digital independence, is unwinnable and therefore useless. No serious economy builds every layer alone. The practical test is narrower: which layer, for which function, at which moment of stress, must answer to domestic authority rather than foreign discretion. The draft Indian map is an early attempt to answer that question in public. Whether it works will depend on two things the map cannot supply by itself: the discipline to define the tiers in operational language, and the institutional capacity to verify that a workload actually sits where the register says it does.

For leaders outside government, the same question arrives in a smaller form. Which internal systems, if access were withdrawn on a Friday afternoon, would stop the institution by Monday morning? That question can be answered this quarter, without a national strategy and without new legislation.

What Happened and Why It Matters

The Economic Times report describes a plan at drafting stage. Three elements carry the weight. The first is the existence of a map at all: an explicit register of critical sectors and data categories, which implies a classification exercise with named owners. The second is the air gap. Air-gapped infrastructure is not a hosting preference; it is a physical and architectural claim, and it is expensive to maintain. The third is the split between the top tier, where the state owns and operates, and the second tier, where the state owns and a private operator runs the estate.

That second tier deserves attention. Ownership of hardware without operational control is a partial sovereignty, and partial sovereignties fail in specific, predictable ways. The operator holds the credentials, the patching schedule, the telemetry, and the institutional knowledge. The owner holds an asset register. If the contract does not specify who can suspend a service, who sees the logs, and who can replace the operator without losing the workload, ownership becomes a legal fact with limited operational meaning.

The reported plan also arrives against a live capacity problem. The Hindu BusinessLine reported that the India AI Mission currently has access to about 30,000 GPUs against commitments of 45,000, with several private providers unable to deliver pledged capacity amid a sharp rise in equipment costs. The same report described government work with C-DAC on roughly 3,000 GPUs of installed capacity, and a separate arrangement with the United Arab Emirates under which an 8-exaflop system, equivalent to about 10,000 GPUs, would be deployed in India through G42 and Cerebras in partnership with C-DAC and the Mohamed bin Zayed University of AI, owned by the Indian government as a gift from the UAE government.

That combination is instructive. A state that intends to hold critical AI capacity inside its own perimeter is simultaneously dependent on delivered hardware, foreign suppliers, and in this case a gift from another government. Sovereignty here is not autarky. It is the deliberate management of dependencies that cannot be removed, with clear sight of which ones can be substituted and which cannot.

The demand side is equally unsettled. Outlook Business reported IT minister Ashwini Vaishnaw saying the government estimates demand for 10,000 GPUs, with an initial tender covering 5,000 units, alongside empanelment requirements before firms can bid for departmental work and five-year supply agreements intended to give vendors demand visibility. Medianama reported that C-DAC issued a request for empanelment for up to 10,000 AI compute units, with conditions requiring all infrastructure, data, metadata, logs, backups and disaster recovery to remain within India, physical and logical isolation of C-DAC’s data, and no unauthorised access, use, sharing or transfer.

Those conditions are contractual, and contracts are where sovereignty is usually lost. A residency clause can be satisfied while control sits elsewhere. A logging requirement can be met with logs the customer cannot independently verify. An isolation clause can be honoured for the data and quietly breached for the metadata that describes how the model was used.

The defence dimension sharpens the argument. Moneycontrol reported on 23 October 2026 that Defence Secretary Rajesh Kumar Singh said India is developing its own military AI layer, with the computing infrastructure being secured under the IndiaAI Mission, to support functions including decision support, target acquisition, data fusion and what he described as the sensor-to-shooter loop. He said that in some key areas the work has to be done domestically, because certain technologies are unlikely to be shared or are too critical to leave with foreign companies.

That is the clearest possible statement of the boundary logic behind the draft map. The threshold is not secrecy alone. It is the combination of consequence and irreversibility: systems where a foreign decision, a foreign outage, or a foreign legal order would produce effects that cannot be undone inside the national perimeter.

KPMG India published a report on 23 October 2026 arguing that India has built public AI capacity quickly and that the question for the next four years is whether the country can control, scale and trust what it has created. It defines sovereign AI as the ability to develop, deploy and govern AI on India’s own terms with appropriate control over critical data, models, compute and rules, and it frames sovereignty as a set of measurable choices about who controls data, models, infrastructure and the ability to switch.

The ability to switch is the least discussed and most decisive element in that list. Air gaps protect the top tier. Switching capability protects everyone else, and it is the only sovereignty asset that can be built by an institution that will never own a data centre.

What happened, then, is not a procurement announcement. It is the publication of a tiering logic, and tiering logics invite imitation. Once a government declares that some workloads must run on infrastructure answerable to domestic authority, every board, university and hospital that serves that government inherits the question in its own operations. The classification exercise will migrate from the ministry to the supplier and then to the supplier’s other clients.

The Economic Times report is careful to note that no timeline has been disclosed. That absence is itself a finding. Classification exercises without published schedules tend to stall at the point where an agency must accept that its own favourite system belongs in a higher-risk tier. The test of the Indian map will be whether the published criteria survive contact with institutional self-interest.

For leaders, the practical issue is not to await the final tiers. It is to begin the same exercise internally, using the same logic, so that when the government register appears the institution can answer it with evidence rather than assurances.

The Sovereignty Risk

The first risk is that air gaps are declared and not maintained. An air gap is a discipline, not a purchase. It survives only while someone owns the rule that nothing crosses it, and while that person has the authority to refuse an exception requested by a senior official under delivery pressure.

The second risk is that tiering creates a false comfort below the top level. If the second tier is described as less sensitive, the organisations inside it will behave as though it were not sensitive at all. In practice the second tier, state-owned hardware operated by private firms, is where most of the governance complexity lives: shared operational staff, vendor-held credentials, monitoring data leaving the perimeter, and an ownership line that nobody can operationally enforce.

The third risk is that the definition of critical sector is drawn by the sector itself. Defence will always volunteer for the highest tier, because it understands consequence. Education, health and municipal administration usually will not, because their systems are procured under cost pressure and their data is treated as administrative rather than strategic. Yet a national assessment system or a public health triage tool can hold more sensitive inference about citizens than many defence logistics platforms.

The fourth risk is records. Air-gapped systems are hard to audit precisely because they are disconnected. Model versions, evaluation results, data lineage and deployment decisions can sit inside a sealed environment with no independent examiner on the other side. Sovereignty that cannot be inspected is indistinguishable from secrecy, and secrecy is not a governance model.

The fifth risk is the supplier position. A vendor that supplies both the model and the hardware, and that also operates the second-tier estate, accumulates a position that no single clause can dissolve. Every future migration becomes a negotiation with the incumbent, and the incumbent knows the switching cost better than the buyer.

The sixth risk is that sovereignty is measured by ownership rather than by capability. A state can own GPUs and lack the engineers to schedule them, the tooling to evaluate the models that run on them, and the procurement skill to replace a component when the supplier exits. KPMG’s framing is useful here: the layers are compute, data, models, platforms and digital public infrastructure, talent and intellectual property, and governance and trust, and each requires a different approach. Ownership answers only the first layer.

The seventh risk is time. Hardware ages. A tier built on chips that are three or four generations behind will be inefficient in both power and performance, a point Counterpoint Research’s Neil Shah raised in the BusinessLine report. An air gap defended with obsolete silicon is a sovereignty asset that decays into a cost centre.

The sovereignty test is therefore not whether the boundary exists on paper. It is whether the institution can state, for each critical workload, who owns the hardware, who holds the credentials, who sees the telemetry, who can suspend service, and how the workload would move if all of those answers changed at once.

The TEE Method Response

The TEE Method works through three connected capacities: Talent, Enterprise and Ecosystem. Talent is the internal ability to judge, operate and refuse. Enterprise is the institutional machinery that turns judgement into contracts, controls and decisions. Ecosystem is the set of external relationships, standards and suppliers through which the institution acts.

On Talent, a tiering regime exposes a specific shortage. Air-gapped infrastructure needs people who can administer isolated systems, evaluate models without external telemetry, and explain to a board why a capability gap is a governance risk rather than a technical inconvenience. Most institutions have cloud engineers and almost none have sovereign operations staff. The practical move is to name two individuals, not a team: one accountable for the critical workload register, and one accountable for verification of what actually runs where.

On Enterprise, the Indian draft map’s second tier is the clearest available lesson. Ownership without operational control is a contract problem, and contract problems are solved in four clauses: credential custody, telemetry access, suspension authority and exit assistance. Each must specify a named role, a maximum notice period, and a verifiable artefact the buyer receives. If the buyer cannot hold a copy of the artefact, the clause is decorative.

Enterprise also means procurement language that distinguishes an air gap from a private cloud. These are not the same control and should never be scored as such. A private cloud in a domestic region is a commercial arrangement inside a foreign governance framework. An air gap is an assertion that the workload cannot be reached from outside at all. Treating the second as an upgrade of the first is how institutions convince themselves they have done something they have not.

On Ecosystem, the lesson is diversification with memory. India’s position, in which committed GPU capacity has not been delivered and a gifted supercomputer from another government materially changes the national supply picture, shows that dependency management requires a register of suppliers by layer rather than a single vendor list. The institution equivalent is a dependency register that records, for each critical function, the model provider, the compute provider, the platform operator, and the fallback for each.

Together the three capacities answer the question in a specific order. Talent defines what must never leave. Enterprise writes the clauses that keep it in place. Ecosystem provides at least one credible alternative for each layer so that the clauses can be enforced without ceasing operations.

The TEE Method’s value in this case is that it prevents sovereignty from collapsing into a hardware conversation. Hardware is the visible layer and the easiest to fund. Talent and ecosystem are the layers that determine whether the hardware remains useful once the announcement is over.

Sovereignty Test Matrix

The matrix below scores the five sovereignty domains for an institution deciding where its AI workloads should sit. A score of one indicates a weak sovereignty signal: control is asserted but unverifiable. A score of five indicates a stronger signal: a named owner, a verifiable artefact and a tested fallback. Score each domain honestly, then apply the decision threshold that follows.

DomainWeak signal (1-2)Stronger signal (4-5)Score
PoliticalCritical-workload tiers defined by the supplier or by budget convenienceRegister signed by an accountable executive, reviewed on a fixed cycle, exceptions documented
EconomicSwitching cost unknown; incumbent renewal treated as automaticCost of exit modelled per workload, with a funded alternative identified
CulturalData treated as administrative; workforce cannot describe what the model doesStaff can explain model purpose, limits and escalation to non-technical oversight
IntellectualModel versions, evaluation results and lineage held only by the vendorInstitution holds its own evaluation record and can reproduce results independently
TechnologicalPrivate cloud described as air gap; credentials and telemetry vendor-heldPhysical isolation where required, credential custody defined, telemetry readable by the buyer

The decision threshold is simple and should be applied before any renewal or expansion. If any domain scores two or below on a workload that has been classified critical, the institution does not have sovereignty over that workload, regardless of where the data is stored. The correct action is not to renegotiate everything at once. It is to move that single workload to a higher tier of control, or to fund the fallback that makes the weakness survivable.

Two patterns usually emerge. Technological scores tend to be inflated because teams score hosting location rather than control. Intellectual scores tend to be depressed because almost nobody holds a reproducible evaluation record. That combination, high confidence about location and low evidence about behaviour, is the most common sovereignty failure pattern in institutions of every size.

What Leaders Should Do This Week

First, produce a critical workload register. List every AI system that would stop a core service within 24 hours if access were withdrawn. Name the accountable executive for each. If the list exceeds ten entries, the classification is too loose; if it contains nothing defence-adjacent, the register is incomplete rather than reassuring.

Second, answer the four control questions for the top three workloads on that register. Who holds the credentials. Who can suspend the service. Who can read the logs. Who can replace the operator. Write the answers down with names, not departments. Any question that cannot be answered in one sentence identifies a clause that is missing from the contract.

Third, stop scoring private cloud as an air gap. In the next procurement review, require the vendor to state in writing whether the environment is reachable from the public internet by any party, including the vendor’s own engineers. Record the answer verbatim. This single question changes the nature of the conversation more reliably than any generic sovereignty clause.

Fourth, hold one evaluation record internally. For each critical workload, store the model version, the test set, the results and the date. This is the cheapest form of intellectual sovereignty available and it survives vendor changes, mergers and contract expiry. An institution that cannot reproduce its own evaluation results has outsourced its judgement along with its compute.

Fifth, price the exit. Ask the finance team for the cost of migrating one critical workload to an alternative provider within twelve months, including retraining, re-certification and service interruption. The number does not need to be small. It needs to exist, because a switching cost that has never been estimated cannot be managed and will be discovered at the worst possible moment.

Sixth, watch the Indian register as a leading indicator. When a large government publishes tiering criteria, suppliers adapt their offerings and their contract language to match. Institutions that have already drafted their own tiers will be able to read those changes quickly. Institutions that have not will receive them as vendor improvements.

Seventh, assign one person to the second-tier problem. State-owned hardware operated by a private firm is the arrangement most institutions will actually end up with, and it is the one with the weakest control documentation. A single owner for credential custody and telemetry access across all such arrangements prevents the slow drift into nominal sovereignty.

Each action is small enough to finish this quarter and specific enough to assign. None requires a national strategy, and none depends on the Indian map being finalised. They depend only on an institution being willing to write down what it currently cannot answer.

The Question Revisited

The question was whether an institution can name the decisions that must remain inside its own perimeter, and then prove that they do. India’s draft sovereign AI map is an attempt to answer that at national scale, and its most important contribution is the tiering logic rather than the hardware. A boundary drawn in operational language, with named owners and verifiable artefacts, is a governance instrument. A boundary drawn as an aspiration is a communications exercise.

The answer the TEE Method supplies is that sovereignty is a capability, not a possession. Talent defines the perimeter by judgement. Enterprise writes the clauses that hold it. Ecosystem ensures there is somewhere to go when the clauses are tested. An institution with all three can operate on rented infrastructure without losing agency. An institution with none of them can own a data centre and still be dependent.

That matters because the next phase of AI governance will be decided in classification rooms rather than in parliaments. The question is not who buys the GPUs. It is who writes the criteria that decide which systems may be reached from outside, and whether anyone independent can check that the answer is true. Institutions that begin that work now will meet the coming register with evidence. Those that wait will meet it with a vendor’s assurance.

For boards and public leaders, the practical sequence is small and immediate: register the critical workloads, answer the four control questions, stop confusing private cloud with isolation, and hold one evaluation record that nobody else can take away.

Sources

The Economic Times, “Govt plans sovereign AI infrastructure with tighter control over data, GPUs”, by Subhayan Chakraborty and Surabhi Agarwal, 20 October 2026 (primary source; also the source of the featured image).

Medianama, “10,000 GPUs for India AI Mission; MeitY invites proposals”, October 2026.

The Hindu BusinessLine, “Centre fortifies India’s AI mission with Sovereign Compute Pool amid rising GPU costs”, October 2026.

Moneycontrol, “India developing its own military AI layer; cannot rely on foreign firms for critical technologies: Defence Secretary”, by Ishaan Gera, 23 October 2026.

Outlook Business, “Govt To Release AI Regulation Consultation Paper In A Month, Eyes Indigenous GPUs In 3-4 Years: Vaishnaw”, October 2026.

KPMG India, “Sovereign by design – Building India’s independent, inclusive and future-ready AI”, 23 October 2026.


Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.

This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework, see tonishatagoe.com.

Keep Reading

Related Articles

LEC Magazine

Join Our Community

Exclusive insights & inspiration

✓

Welcome to LEC!

Account created. Refreshing…

LEC Magazine

Join Our Community

Exclusive insights & inspiration

✓

Welcome to LEC!

Account created. Refreshing…