India faces a growing risk of an AI blackout if it continues to rely on foreign large language models (LLMs), global brokerage firm Bernstein has warned, calling for a sovereign AI stack and an India-made “DeepSeek” to safeguard economic and strategic interests.
Arguing that India “can’t build its AI future on borrowed models”, Bernstein analysts Venugopal Garre and Nikhil Arela cautioned that a strategy built on renting compute for foreign LLMs and earning data-centre rent could leave the country dangerously exposed. In a report the analysts flagged recent US export-control actions, including restrictions on access to frontier models for non-US citizens, as an early signal that AI is shifting from a globally shared technology to a tightly controlled strategic resource.
“Foundational models will no longer be SaaS products,” the report notes, likening cutting-edge AI to “fighter jets” in a new era where access to the best models is guardrailed and rationed. In Bernstein’s base case, AI settles into “a world of stratified access”, with advanced nations reserving bleeding-edge capabilities for domestic defence, intelligence and critical infrastructure, and exporting older, downgraded versions to emerging markets.
In the extreme scenario, India could “wake up and find that the most powerful applications in finance or military are no longer working, and the APIs and weights they rely on have vanished overnight”.
The concern is amplified by India’s deep technology dependence across the stack, from CPUs, GPUs and mobile SoCs to operating systems, cloud platforms, enterprise software and consumer internet rails. Across core compute, operating systems, browsers, public cloud, social media and productivity suites, India overwhelmingly relies on US and other foreign vendors, with domestic initiatives limited to pockets and lacking scale.
Bernstein warns that repeating this pattern in AI—letting global firms own India’s intelligence layer while local players stay in the application margins—could entrench dependence “for decades”.
The absence of an Indian “DeepSeek moment”, despite India’s vast data reservoir, is described as structural. India’s tech ecosystem has been “services-led”, without large-scale, consumer-facing platforms in search, social or messaging that generate rich, organized data for training frontier models. This has meant no sustained talent pipeline or academic depth around foundational models, while IT services have rewarded fine-tuning foreign software rather than building core platforms.
Bernstein notes that many Indian institutional leaders have argued India does not need its own LLMs and can focus on applications, calling these views “more reflective of the path India has taken” than a deliberate strategic choice.
Policy efforts so far are characterised as “too lumpy, too little” and “defocused, thinly spread”. The India AI Mission, launched in 2024 with an outlay of around INR 104 billion (about USD 1.2 billion over five years), identified compute, data, research, hardware, applications and foundational models as key pillars.
But allocations have been volatile, with a notable reduction in AI spending in the revised FY2025 budget, and only about USD 220 million—less than 20% of the total—earmarked for foundational models. Around 44% of the envelope is directed toward compute, including imports, which Bernstein suggests risks reinforcing dependence rather than building domestic capability.
The brokerage contrasts India’s broad-but-thin approach with more sequenced strategies in the US and China. The US leveraged access to compute and deep private capital to push rapid advances in model performance, while China focused on model efficiency and semiconductor capabilities under export pressure.
Over time, both ecosystems expanded across the stack from an initial base of concentrated investment, whereas India is trying “to do a lot from less resources” and across multiple layers simultaneously. Bernstein also reminds investors that India has historically been subject to Western export controls in nuclear, space and defence, arguing AI is likely to follow the same pattern of sanctions and restricted flows.
The report underlines the strategic risk of India’s AI stack sitting “at the mercy of someone else”. If core enterprise, defence, space and financial systems were to run on foreign LLMs, a geopolitical disruption could curtail access “overnight”, halting critical applications. Even without a hard cut-off, India could be locked into N‑1 or N‑2 models, operating one or two generations behind, with local firms competing against global startups bootstrapped by “vibe coders” who have direct access to frontier systems. “India cannot afford to operate on someone else’s trapdoor,” Bernstein warns, calling deep AI capabilities a “necessity, not a luxury”.
Against this backdrop, Bernstein outlines why it wants India to pursue its own “DeepSeek” and a sovereign AI stack. The core recommendation is to pivot from horizontal dependence on foreign models to owning and controlling high-quality, domain-specific datasets in sectors such as Industrials and healthcare. These vertical datasets, if retained and shielded from unrestricted access by global platforms, can underpin smaller, specialized LLMs built by Indian technology firms, on which defensible applications are layered. Bernstein sees the opportunity extending beyond software into the physical economy, including training industrial robots and next-generation humanoid systems where industrial-context data may matter more than bleeding-edge GenAI.
Policy levers, in Bernstein’s view, are not straightforward but two stand out. One is to “restrict or pace access to global AI models” while channelling capital and talent into India-native LLM capabilities. The other is to mandate or incentivise localisation—requiring foreign firms to build and operate India-based AI stacks insulated from geopolitics. Both options are imperfect, but they frame the crucial trade-off between access and autonomy. Bernstein suggests the government’s role should be primarily policy-driven—akin to China’s “great firewall”—as long as Indian companies are willing to participate in the LLM race; if no one wants to build models and everyone focuses on applications, even strong public investment will struggle to change the outcome.
