Dario Amodei knows his audience well. A reference to The Lord of the Rings is an easy way to get people in tech to listen. In his latest policy note, he invoked the indolence of Treebeard to urge quicker action on global artificial intelligence regulation. If only he also recalled J R R Tolkien's larger comment on unchecked power and environmental destruction. Nearly at the same time, economics writer Martin Wolf called for a US-China ‘AI disarmament treaty’ invoking Cold War language.
Amodei, in particular, seems to want to position himself as the Oppenheimer of our time. But unlike Oppenheimer’s change of heart after the atomic bomb, Amodei’s warning comes at a time when his company Anthropic seeks a trillion-dollar valuation in an upcoming stock listing.
Those calling most loudly for restraint on AI are the ones who built without it and will continue to benefit from it. One only needs to go back to 1968 when the nuclear non-proliferation treaty was hailed as an arms control achievement. The NPT limited proliferation, but entrenched the privileges of the existing nuclear powers. Decades later, the treaty’s Article VI commitments to disarmament remain unfulfilled.
Quite similarly, led by a handful of American companies, AI has acquired fearsome capabilities without much regulatory oversight. As this lead hardens into pre-eminence, the conversation is conveniently moving towards frameworks, guardrails and cooperation.
When Amodei makes a case for mandatory third-party testing of frontier AI, the proposal comes off as neutral. In practice, however, testing by bodies “authorised and inspected by the government” means that the US government holds influence over which AI systems can be deployed globally. That concern was brought into sharp focus when Anthropic itself, following a US export-control directive, abruptly disabled access to Fable 5 for users outside the US. This is what an effective veto looks like.
The wider recommendation Amodei makes for a “coalition” of democracies to manage AI supply chains is a palatable packaging of the chip export control regime. India already experiences this under the US’s AI diffusion framework where access to advanced chips such as Nvidia’s H100 and A100 is contingent on quantity caps and licensing requirements—not over productivity concerns, but because frontier compute enables capabilities from autonomous weapons to cyber operations.
The deeper problem is that while AI regulation is necessary, what is being proposed resembles a regulatory capture. Where is the democratic accountability from which any global regulation could feasibly derive authority? The plausible, even imminent, privatisation of public governance is the real danger. Amodei correctly identifies cybersecurity, bioweapons, automated research and development and loss of control of AI systems as risk categories that would trigger regulatory action. But what about other concerns developing economies have? The largest risk, of labour displacement, is left to voluntary corporate measures—entirely inadequate for countries without strong safety nets.
The nuclear analogy becomes interesting here. India rejected what Jaswant Singh called “nuclear apartheid”, endured decades of technology denial, built capability regardless, and was eventually included through the 2008 Nuclear Suppliers Group waiver. The lesson here isn’t that India should reject global rules but remember how a seat at the table is truly earned. It must soberly ask what there is to gain by accepting the terms of an existing framework.
Rather than pursuing the unlikely goal of autonomy as some have suggested after Fable’s pull-out, India should instead seek indispensability within the global AI stack. AI is not a single technology but a cascade of chokepoints, each dominated by different actors. The India Semiconductor Mission 2.0 emphasises the need to pivot towards design intellectual property, equipment, supply chain resilience and R&D.
India must prioritise natively built small language models in AI systems across domains such as agriculture, healthcare, governance, that others need to integrate with, on terms it sets. This is what the digital public infrastructure achieved in payments and identity. Something similar is achievable in AI, where India’s scale creates both the problem and the market.
The groundwork laid now could provide the demand-side infrastructure for future manufacturing breakthroughs. Current AI hardware, with its energy and funding limitations, is reaching its physical limits. India needs to have optionality on the hardware paradigms that follow. Like electricity far before it, AI will undergo multiple technological leaps before maturing. India should back a few calculated bets that could reshape the hardware foundations of AI. Photonics is one such bet, with IIT Madras and IIT Delhi already investing in the field whose technological trajectory remains open.
Amodei’s plea may only accelerate AI development in ways that he precisely wants to avoid. Language that describes countries even three years behind the AI frontier as mediaeval swordsmen versus US Marines will not engender calls for regulation as much as increase the determination to achieve some semblance of AI ‘sovereignty’. Responding to the Fable directive, he wrote: “We believe the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear and grounded in technical facts. This action does not adhere to those principles.” But he called for this power. He shouldn’t be surprised that the US government exercised it the first chance it got.
Indeed, as former diplomat Stephen Krasner points out, sovereignty is organised hypocrisy at its core. Global AI regulation, as currently conceived, is merely the latest iteration of this phenomenon.
Rohit Saxena and Nidhi Singh | Researchers at the Foundation for Advancing Science and Technology
(Views are personal)