Last week, Anthropic was forced to suspend Claude Fable 5 and Mythos 5—its most capable artificial intelligence (AI) models ever, pulled just days after public release—after US national security authorities flagged concerns about the system being jailbroken.
Every Indian senior executive and technology leader should read that news carefully to gauge what it reveals about the strategic risk of building an AI future based on someone else’s infrastructure and subject to someone else’s government.
India’s enterprise AI conversation has been dominated by a single assumption: that serious AI means access to the most powerful frontier models—GPT-4o, Gemini Ultra, Claude Fable and others. Boards often ask which model the company uses. Vendors compete on benchmarks. The implicit belief is that capability lives at the frontier, in hyperscaler clouds, or in San Francisco.
That assumption is now expensive, fragile and—for Indian enterprises—flawed.
The cost problem alone should have prompted a rethink. Running frontier models at any meaningful scale costs between $15 and $60 per million tokens.
For a large Indian bank that processes millions of customer queries, a logistics company routing hundreds of thousands of shipments daily, or a consumer products firm analysing distributor data across 600 districts, these costs compound quickly into a budget that is difficult to sustain—and nearly impossible to justify to a board that expects AI to reduce costs, not add a new category of them. For smaller firms the costs are unthinkably high.
Research from Andreessen Horowitz last year found that AI inference was consuming 60% to 80% of the total AI infrastructure spending at large enterprises globally. Indian companies, which must pay in dollars for cloud compute resources, face an additional foreign-currency burden.
Now add regulatory fragility. The American government’s action aimed at Anthropic’s Fable 5 and Mythos 5 is not an isolated incident. The Donald Trump administration has an ongoing lawsuit that aims to restrict US government agencies from using Anthropic’s tools. US export controls on advanced AI capabilities are tightening.
Any Indian enterprise that has embedded a US frontier model into a core workflow—credit underwriting, for example, or customer service or supply chain decisions—has effectively accepted dependency on the continued goodwill of a foreign government and an overseas company’s regulatory compliance. This is a sovereignty risk.
The capability gap between frontier proprietary models and the best open-source alternatives has closed decisively for most enterprise-use cases. Mistral, Microsoft’s Phi-3, and Google’s Gemma can handle summarization, classification, document extraction, customer query resolution and domain-specific reasoning at a fraction of the cost.
More importantly, they can be fine-tuned on Indian languages, regional datasets and sector-specific knowledge—something a generic frontier API cannot do well and will not prioritize.
India is not without its own answers here. Sarvam AI has built foundational models trained on Indian languages and domains, purpose-built for the subcontinent’s linguistic diversity. BharatGen, a government-backed initiative under IIT Bombay, is developing multimodal models rooted in Indian data.
These are early but significant signals that India need not simply choose between American proprietary models and Western open weight releases. It can and must invest in building its own model layer. For India’s 63 million MSMEs, models fine-tuned on GST data, vernacular interactions and local credit behaviour could outperform generic frontier APIs on every task that actually matters
Infrastructure logic reinforces this. India’s connectivity is uneven, cloud latency is real outside tier-1 cities and data localization requirements under the evolving Indian law make cloud-hosted foreign models increasingly difficult to justify in regulated sectors.
Edge inference—running capable models on devices or local servers—solves all three problems simultaneously. Modern chips can run models up to 70 billion parameters locally and at low cost, with no data leaving the enterprise or the country. For a rural micro- finance institution, district hospital or a state government department, this is the right architecture—not a compromise.
India has a genuine opportunity here that it should not squander by chasing the frontier of AI models.
The businesses and institutions that will lead India’s AI decade are likely to be those that invest in building proprietary datasets, fine-tune open models on Indian contexts, deploy it intelligently across edge and cloud, and then own the resulting capability. Those that had hoped to rent access to the world’s most powerful foreign model, in contrast, would find themselves in a bind.
Anthropic told the world that Fable 5 was too powerful to release. One US government order later, it was gone.
The question that we must now answer is: how much of our AI strategy lives in infrastructure we do not control—and what exactly would we do if it disappeared tomorrow?
The author is a chair of the Global Energy Alliance for People and Planet and a former chairman of Microsoft India.