What the June 2026 export-control order revealed about dependency — and why India's real opportunity is to build the layer underneath, not just chase the frontier.
The day the off-switch became visible
In June 2026, a single government directive did something the AI industry had never seen before. The United States Commerce Department directed Anthropic to block its two most advanced models — Fable 5 and Mythos 5 — from foreign nationals anywhere in the world. The company could not separate those users from everyone else in real time, so it took the only option available to it: it switched both models off, globally.
To understand why this matters, look at how export controls on AI have moved over the last four years. They climbed the stack one rung at a time. First the chips. Then the equipment used to make them. Then the raw compute, and the cloud access that delivers it. Then, in early 2025, the weights of the most advanced closed models. And now, in 2026, the control reached the final rung: access itself — who is allowed to send a prompt to a model running in a data centre.
This last step was new in kind, not just in degree. No hardware moved. No weights were downloaded. Nothing was "exported" in the way the word normally means. Security researchers described it as unprecedented, and they were right. For the first time, the question was not what you could buy or own, but whether you were permitted to use a service at all.
Whatever you think of the decision, it made one thing impossible to ignore: the most powerful AI in the world can be switched off for you, overnight, by a decision you had no part in.
The real lesson is about dependency, not politics
It is easy to file this under "geopolitics" and move on. That would be a mistake. The deeper signal is structural, and it applies to everyone — companies, researchers, and entire countries.
If your product runs entirely on a single company's model, your roadmap is not yours. If your national digital strategy assumes uninterrupted access to systems built and governed elsewhere, your strategy carries a dependency you do not control. You are not buying capability; you are renting it. And the landlord can change the terms while you sleep, for reasons that have nothing to do with you.
To be fair, this is not a simple story of one government overreaching. Frontier models — the largest and most capable systems at any given moment — carry genuine dual-use risks, in areas from biology to cybersecurity. A government that restricts its most advanced systems out of security concern is not behaving irrationally, and pretending otherwise is not honest. The sharper, harder-to-argue point is this: concentration is the risk. When the most powerful capability sits in very few hands, a single decision becomes a single point of failure for millions of people who were never in the room.
Decentralisation is already happening — on two tracks
Here is what most of the coverage missed. In the same window that frontier access tightened, allied governments — Canada, the European Union, Japan — began openly raising the problem of concentration in the US AI supply chain and looking harder at their own alternatives. Restriction at the top does not stop decentralisation. It accelerates it.
The honest picture is that AI is decentralising along two tracks at once, and they are easy to confuse.
The first track is the closed frontier — the handful of largest, most capable models. That track is getting more controlled, not less, and it will probably stay that way, because that is precisely where the risks and the leverage are highest.
The second track is everything just beneath the frontier: open-weight models that anyone can run, and nations building their own sovereign capability. That track is opening up quickly. "Technology for everyone" is becoming true here, even as the first track narrows.
These two are not in contradiction. They are happening together. So when people ask, "Is AI really being decentralised, or not?", the honest answer is: both, on different layers, at the same time.
India has been making this exact case
India has spent the last two years arguing this on the world stage — and after June 2026, those arguments read less like idealism and more like risk management.
At the AI Action Summit in Paris in February 2025, which he co-chaired with President Macron, Prime Minister Narendra Modi said: "We must build quality data sets, free from biases. We must democratise technology and create people-centric applications." He also pressed for the global AI partnership to be made "truly global," more inclusive of the Global South and its priorities.
One year later, India turned that position into the agenda of an entire summit. Inaugurating the India AI Impact Summit 2026 at Bharat Mandapam in New Delhi — the first global AI summit hosted in the Global South — Modi was direct: "We must democratise AI. It must become a tool for inclusion and empowerment, particularly for the Global South." He warned against a future in which people are reduced to "raw material" for AI, and used a navigational analogy to describe the right relationship between humans and the technology: like GPS, AI can show us the way, but the final decision on which direction to take must remain in human hands.
He framed India's approach through a five-part vision he called M.A.N.A.V. — Moral and ethical systems, Accountable governance, National sovereignty, Accessible and inclusive, Valid and legitimate. Two of those five pillars — national sovereignty, and accessibility — are exactly the qualities the June 2026 episode showed the world were missing when capability is concentrated. He closed with a global invitation: "Design and Develop in India. Deliver to the World. Deliver to Humanity," under the summit's theme of सर्वजन हिताय, सर्वजन सुखाय — welfare for all, happiness for all.
This is no longer a slogan. It is a strategy with a clear logic behind it.
India's real opportunity is the layer underneath
So what should "India building AI aggressively" actually mean in practice? Not, primarily, trying to out-spend the frontier race head-on. That race demands enormous compute, capital, and concentrations of talent, and winning every front of it is neither realistic nor necessary. The smarter and more durable move is to dominate the layers where India is structurally hard to beat. There are three.
First, data and language. India has 22+ major languages and thousands of genuine regional dialects. Authentic, high-quality, low-bias data in these languages — the kind Modi himself named in Paris — is something no other country can simply copy. India's National AI Repository, announced to democratise access to datasets and models, points in the same direction: the foundation matters, and the foundation is data.
Second, talent and the application layer — building real products on top of models, where India's engineering depth is already proven and globally deployed. As Modi argued, a model that succeeds at India's scale and diversity can be deployed almost anywhere.
Third, and most overlooked, the human layer. Every model — frontier or sovereign, open or closed — needs human feedback to become useful and safe. Human-preference data, reinforcement learning from human feedback, careful annotation, multilingual evaluation, and red-teaming: this work is required no matter who builds the model, and no matter who is permitted to access it. It does not depend on a single company's permission, and it cannot be switched off by a single directive.
That last layer is the quiet centre of gravity in the whole field. A model is only as good as the human judgment poured into it — and for Indic AI, that judgment is best produced by people who actually speak the languages and live the dialects, not by approximations imported from elsewhere.
Resilient diversity, not fragmentation
One caution deserves to be stated plainly, because the opposite mistake is just as costly. Decentralisation is not automatically good. If every country tries to rebuild the same frontier model from scratch, the result is duplicated spending and weaker coordination on the safety questions that genuinely need global cooperation. The goal is not fragmentation, where everyone goes it alone and nothing connects.
The goal is resilient diversity: several strong, capable players, each standing on its own genuine strength, and able to work together through shared standards. Many points of strength instead of one point of failure. That is the difference between a healthy ecosystem and a brittle one — and it is the version of "decentralised AI" actually worth building.
Where TrainPlex stands
This is the conviction behind TrainPlex. We are not betting on which lab wins the model race; that bet is fragile by design. We are building the layer that stays essential regardless of who wins — the human layer behind Indic AI. A managed workforce of 20,000+ certified AI Trainers across 22+ Indian languages, producing the human-preference data, annotation, evaluation, and red-teaming that frontier and sovereign models alike depend on — in the authentic regional languages that are hardest to source anywhere else.
A single directive can switch off a model. It cannot switch off a country's decision to build its own foundations. June 2026 made that case more clearly than any of us could have. The work now is to answer it — not by trying to win every race, but by owning the ground that every race is run on.
— Vinod Parihar, Founder & CEO, TrainPlex (Cybdeer Network Pvt Ltd)