Anthropic, Atom Bomb, Arundhati Roy
The Antithesis Trap
In May 1998, after India tested nuclear devices at Pokhran, Arundhati Roy published “The End of Imagination.” The essay was morally serious. It called the bomb “the most antidemocratic, antinational, antihuman, outright evil thing that man has ever made.” It asked why a country with four hundred million people who could not read or write, who did not have safe drinking water or enough to eat, needed a nuclear weapon. The question was as solid as the prose was. But the structural position of the prose — what it could see and what it could not —is worth examining now, twenty-eight years later, because the same position is likely to reproduce itself in a different domain.
Roy’s essay could not distinguish between two claims. The first: nuclear weapons are terrible and the decision to build them must be examined with moral seriousness. The second: India should not have built them. The first is a thesis. It survives any government. It makes demands on every country that possesses nuclear weapons, including the ones that already had them when Roy wrote. The second is an antithesis. It is organised against the BJP government that conducted the tests.
The distinction matters because of what happened seventeen days later. Pakistan tested its own devices at Chagai on May 28. The liberal objection to Pokhran treated the test as provocation — as if India’s restraint would have prevented Pakistan’s programme from reaching completion. It would not have. Pakistan had been building its bomb since the mid-1970s. A.Q. Khan had stolen centrifuge designs from the Dutch in 1975. By 1987, Pakistan had enough enriched uranium for a weapon. The programme was already selling technology to Iran, Libya, and North Korea before India tested at Pokhran. The question was not whether Pakistan would test. The question was when. India’s test provided the political occasion.
The liberal complaint about Pokhran was not wrong about the moral weight of nuclear weapons. The argument is intellectually sharp but structurally wrong. Roy and the tradition she spoke from could not engage the question of sovereign capacity the country surrounded by nuclear-armed neighbours and dependent on the fleeting goodwill of distant powers for its survival. Engaging that question would have required granting the BJP government a legitimate national security achievement.
The antithesis would not permit it. So, the complaint was framed as a universal moral objection to nuclear weapons and published in India, directed at India’s test, timed to India’s government, and heard by Indians as an objection to India having the bomb. The universalism was the derivation. The residue was: this government must not be allowed this.
The aftermath settled the question. It always does. India has not used a nuclear weapon. Pakistan has not used one against India. The deterrence holds. Again, the liberal position was morally serious and strategically vacant. The strategic vacancy was not a failure of intelligence. It was produced by a structural incapacity to engage sovereign capacity. When the wrong government was building it.
Last Thursday, the United States government ordered Anthropic to suspend access to its two most advanced AI models — Fable 5 and Mythos 5 — for every non-American on earth. The mechanism was an export control directive citing national security. No consultation with allied governments. No international framework. No process. A letter was sent in the evening, and by the next morning Fable was gone.
Most people encounter AI as a chat window to cook up a recipe, a spreadsheet, a letter. Fable could do that. But the model beneath it — Mythos — can read a software system and find its vulnerabilities. It can analyse code, find its vulnerabilities, then fix them. Meaning it can hack a security system. The export control was not about cooking recipies. It was about capability that governments classify the way they classify signals intelligence.
Sridhar Vembu, who built Zoho from Tenkasi into one of India’s largest software companies and stepped down as CEO last year to focus on AI research, posted his response within hours. “Technology is the ultimate weapon. National sovereignty, national security, all of it is now about technology. Globalisation is dead, and Bharat must find her own way ahead.” He pointed to Sarvam AI — a Bengaluru startup that has built 105-billion-parameter language models under the government’s IndiaAI Mission using domestic compute — as evidence of the possibility of sovereign AI models.
Vembu is right that India’s dependency is real and that alternatives are being built. Sarvam AI, founded by researchers from IIT Madras, has received government support, access to thousands of GPUs, and membership in NVIDIA’s Nemotron Coalition. These are not mere gestures. They are the early infrastructure of sovereign AI capacity. Vembu is clear about the math: India cannot match the hundred-billion-dollar compute budgets of American labs. Instead, he proposes a pragmatic detour through smaller, hyper-efficient models—even if that means building on Chinese open-source architectures.
Anthropic itself calls the suspension a misunderstanding and is working to restore access. The point is not that India is helpless. The point is that India is vulnerable.
But Vembu’s frame stops one layer above where the problem actually sits. He sees the technology dependency. He does not see what would prevent India from resolving it. The obstacle is not compute or capital or talent, though all three are scarce. There is one thing that breaks things in India, from copyright amendments to labor laws — absence of political will. The obstacle is that sovereign AI capacity requires sustained political will across electoral cycles. That political will is likely to be structurally undermined by the same dynamic that made Pokhran a moral catastrophe in the eyes of the people whose support is vital for the project.
Watch for the derivations.
The Left would argue that the government is spending on AI while the poor have no clean water. The framing will echo Roy almost exactly — the juxtaposition of technological ambition against unmet basic needs, as if sovereign capacity and public welfare were a zero-sum choice rather than a sequence. The framing was available in 1998 and it is available now. The attempt will be sincere. It will also be structurally incapable of acknowledging that a country whose knowledge infrastructure is controlled from Silicon Valley cannot possibly sustain welfare programmes on its own terms.
Liberals would naturally worry that the state is building a surveillance apparatus. They would argue that sovereign AI in the hands of this government means the government will control the knowledge corpus — will read your messages, will monitor dissent, will turn the models into instruments of majoritarian enforcement. Rahul Gandhi has already alleged, more than once, that the government tapped his phone using Pegasus. The concern about surveillance is not invented. Pegasus and the IT rules are real. A government that has demonstrated willingness to surveil its opponents cannot be granted new surveillance infrastructure without institutional safeguards.
But let us examine the structure of the objection. It does not say: build sovereign AI capacity with independent oversight, judicial review, and legislative accountability. It says: do not build it. The concern about surveillance, which is legitimate, becomes the reason to prefer that the knowledge infrastructure remain under the control of an American corporation subject to American export controls exercised at the discretion of the American government — a government that pulled the capability from every Indian user last Thursday without asking anyone’s opinion. The antithesis to Modi makes American dependency more tolerable than Indian sovereignty. Because, at least, the American company is not him.
What the objections miss, as always, is the cost of not building. The objections to sovereign AI on surveillance grounds, would not address the implications when the next export control directive arrives and India has no fallback. The objections on welfare grounds would not calculate what happens when the tools mediating our laws, medicine, and education are trained exclusively on Anglo-American data, optimized for Anglo-American norms, and subject to sudden withdrawal at Anglo-American discretion. The objections are against the building. They are never against the dependency. The dependency is invisible because it is not the enemy. The enemy is the government that would do the building.
The question is not whether India should build sovereign AI capacity. The question is whether India can summon the political will to sustain the building when every opposition tradition in the country is structurally organised to prevent the government from being seen to have built anything worth having.
Vembu is right that globalisation is dead. He is a builder, and builders see what commentators miss. But the question he does not ask is the harder one. It is not a technology question, it is political. Whether a country whose intellectual traditions are organised around opposition to the government can build the things that only the government can build.
The bomb got built because the decision could be taken in secret and executed before the objections arrived. AI infrastructure cannot be built in a desert, in secret. It requires a decade of sustained public investment, visible allocation, political defence. In other words, it requires a consensus that the country’s opposition traditions are structurally incapable of joining.
The foot reads the road. The road is: last Thursday, every Indian lost access to one of the most advanced AI models in the world because one government sent one letter to one company. The foot knows what that means. Whether the head can act on what the foot knows is the question this country has never been able to answer for long.
This essay explores the core thesis of the book The Antithesis Trap