In the space of a single year, the world has watched the supply lines of artificial intelligence become instruments of statecraft. China has tightened the spigot on rare earths — the magnets, gallium and germanium that the entire electronics base depends on — turning export licences on and off as leverage. The United States has drawn and redrawn restrictions on advanced chips, controlling which nations and companies may buy the accelerators that train modern AI. And most strikingly of all, the controls have begun to reach past the hardware to the software itself: export rules have now been used to restrict access to frontier AI models across borders — treating a model, pure code deployed to millions, as something a government can switch off by decree.
Each episode has its own headline and its own half-life. But read together, they are not separate news stories — they are the same story, told at different layers of the same stack. For the first time, governments are treating every rung of the AI ladder — the raw materials, the machines, the chips, and now the models — as terrain that can be fenced, licensed, or closed at a border. And that should make every country building its future on AI pause and ask a simple question: how much of the AI technology do we actually control, and how much sits behind someone else’s off-switch?
This is not one country’s problem. It is everyone’s.
No single country owns the AI stack end to end. Not one.
The United States leads at the model layer and in chip design, but its leverage over the physical layers is borrowed — the only company on earth that makes EUV lithography machines is Dutch (ASML); the company that fabricates the most advanced chips is Taiwanese (TSMC). China leads no layer at the very top of capability, but it dominates the materials floor — the mining and, more decisively, the processing of the rare earths, gallium and germanium that everything else rests on. Korea holds critical ground in memory. Japan in materials and tooling.
Strip it down and the AI stack is a chain of near-monopolies held by different nations — each of whom can choke the others at a different point. That is not a comfortable picture. But it is a levelling one. Every nation pursuing AI-led growth is exposed somewhere. The question is not “are we dependent?” — everyone is. The question is where, on whom, and what we are doing about it.
So let me make this concrete for India — though the same exercise applies to almost any nation with ambition and a realistic view of its own position.
Where India stands, layer by layer
I find it clearest to see the whole stack at once. For each layer: where we are today, where a strategic alliance carries us, and where genuine ownership is achievable.
| Layer | What it is | India today | Through alliances | Path to ownership |
|---|---|---|---|---|
| Models | The frontier AI systems | Capable fast-follower; not yet at the frontier | Open-weight bases (Western & Chinese) give a strong start | Achievable — needs talent + compute, not scarce inputs |
| Compute / cloud | Infrastructure to train and serve | Dependent on imported accelerators | Sovereign-compute access via procurement & partnerships | Achievable over time; design-led |
| Chip design | Architecting the silicon | World-class today — ~20% of the world’s chip designers are Indian; a 2nm design taped out from India | — | Already owned |
| Packaging / ATMP | Assembly, test, packaging | Coming online now — first facilities operational in 2026 | Partner-led ramp | Achievable this decade |
| Fabrication | Printing the chips | Entering at mature nodes (28nm+); first silicon targeted late 2026 | Partnered with Taiwanese fab expertise | Mature nodes: yes. Leading edge: not this decade |
| Lithography (EUV) | The machines that print advanced chips | None | ASML ecosystem access (a Dutch monopoly) | Not realistic — but true for almost everyone, including China |
| Materials | Rare earths, gallium, processing | Reserves-rich, processing-poor — large ore endowment; <1% of output | Stockpile + overseas assets to bridge | Achievable in ~7 years — the gap is refining, not geology |
India’s two scarcest-seeming constraints are the two it already holds. Everyone assumes the hard parts are chip-design talent and raw materials. India has both: the engineers who design the world’s chips, and the geology beneath its own soil. The deposits are real — the missing piece is the infrastructure to refine ore into high-purity materials and magnets. That is a build problem, not a resource problem. You cannot manufacture geology; you can build a processing plant. India’s challenge is the buildable kind.
The genuinely hard walls — leading-edge fabrication and EUV lithography — are walls for nearly everyone. No nation is sovereign there. So they are not where the race is lost. They are where alliances do the work, and where the only real strategy is to refuse single-source dependence.
The honest difficulties
Three, plainly.
The materials catch. India’s rare earths sit mostly in monazite beach sands carrying radioactive thorium, and processing methods don’t transfer cleanly from other ore types. This is solvable engineering, but it is real work, not a press release.
The dependency trap. Even as we build processing, much of the mining and refinery machinery is Chinese-made. Build carelessly and you re-import the very dependency you set out to escape.
The time-boxed window. The current calm in rare-earth trade is a negotiated truce, not a resolution — and it has a clock on it. The opening to build is now, not indefinitely.
The path forward
A coherent strategy falls out of the table almost on its own. It is not “build a fortress” — that is neither possible nor necessary. It is three disciplines held together.
Own what you can dominate. Design, talent, software and models, packaging, materials processing — the layers where India’s endowments give it a genuine claim. Press the advantage where it is real, with the right investments and focus.
Partner-diversify what you cannot build alone. Leading-edge fabrication, EUV access. The goal is not ownership; it is multiplicity — derisking the dependency. This is the central lesson of the moment: dependence on a single sovereign’s goodwill is the core vulnerability. Easier said than done — managing equitable relationships in today’s world is hard, but it has to be managed.
Stockpile and time-box against the one true adversarial chokepoint. Hold reserves and buy the years needed to stand up domestic refining, using the truce window deliberately rather than hoping it lasts.
This is a journey of a decade, not a quarter. But a path exists — and for India, a more favourable one than the headlines suggest, because the foundations it most needs, it already has.
The challenge beneath the table: power
The table maps who controls each layer of the stack. But there is a constraint that sits underneath all of it, and in 2026 it has quietly become the most binding of all: electricity. A chip that cannot be cooled and powered is inert, and the world is running short of power. India must solve the power problem as much as any other in the stack. This one has no external dependency — it requires unyielding domestic execution.
There is a silver lining. India’s monazite sands are a dual-purpose strategic endowment — the rare earths feed the materials layer of the AI stack in the near term, and the thorium in the same sands feeds a long-horizon bet on energy sovereignty (the Fast Breeder Reactor reaching criticality in 2026 being a real but early step toward a stage-three payoff decades ahead). The same geology that complicates rare-earth processing also underwrites the power story — dual leverage that is rare among nations, and a genuine, under-appreciated element of India’s position. The caveat is honest: this is long-term energy independence. The short term still has to be actively solved.
And there is no off-switch on the shortage itself
Power is the most visible case, but the same mismatch runs through the whole manufacturing stack — and it holds even in a world of perfect goodwill, where no one ever restricts anyone. There is simply not enough manufacturing capacity to meet demand, and there will not be for years.
This is not speculation. Through 2025 and into 2026, the chief executives of TSMC, SK Hynix, Micron, Samsung, Nvidia and Intel said the same thing in unison — demand for advanced chips, advanced packaging and high-bandwidth memory is rising far faster than capacity can be built. TSMC’s CEO described demand for its most advanced fabrication as running at roughly three times available capacity. The binding constraints are not even the chips themselves — they are the two least-visible steps: the advanced packaging that bonds chip to memory, and the memory itself. A handful of specialised facilities, each taking two to three years and tens of billions to replicate, govern the pace at which the entire world can deploy AI.
For a nation, this turns the argument from defensive to offensive. There are now two reasons to build sovereign capability, not one. The political reason: every layer you don’t own is a layer someone can switch off. And the physical reason: even if no one ever switches anything off, capacity is rationed by who built it and who contracted for it first. In a world short of compute, access flows to whoever owns the plant or signed the long-term agreement years in advance. A nation with no domestic capacity and no pre-committed allocation does not merely face a political risk — it stands at the back of a physical queue that money alone cannot jump, because the capacity does not yet exist to buy.
That is why the window matters. The countries and companies building capacity now, during the shortage, are buying priority in a decade when compute is the scarce input to growth itself. Waiting is not a neutral choice — it is a decision to arrive last.
The real lesson
Each of these disputes will be resolved in its turn — licences will be reissued, restrictions eased, access restored, the specific quarrels will fade. But none of them was an anomaly. Taken together they are a preview, not an aberration — the moment the world began to treat every layer of the AI stack, from the rare earths in the ground to the most abstract layer of all, the model itself, as something that can be controlled by whoever holds it.
For any nation betting its growth on AI, the takeaway is not fear. It is clarity. Every layer you do not own is a layer where someone else holds a switch.
Sovereignty was never about owning everything — no one does. It is owning what you can, partnering widely enough that no single hand can close the door, and knowing exactly where your remaining exposure lies.
That is not a wall. It is a map. And the work of the decade is to walk it deliberately.