Rising AI costs in US models spur India to build sovereign compute infrastructure as Reliance enters AI space

Rising AI costs in US models are signaling a broader need for India to build its own compute infrastructure and to ensure affordable power as a hedge against upstream shocks. The entry of Reliance into the AI space underscores a shift toward domestic capabilities as a strategic objective. India should prioritize compute-efficient research and affordable energy to sustain technological independence over the long run, reducing exposure to external disruptions.

AI has been widely adopted not only for its capabilities but also because pricing has remained relatively affordable. That affordability is under pressure; adoption scales usage, and usage costs for US-based AI models are rising. Price hikes reflect concerns about upstream dependency and the broader trend where external shocks create domestic challenges. As AI increasingly becomes viewed as national infrastructure, India should gradually build greater sovereignty over its compute stack, even if immediate risks are not evident.

Basic AI access remains affordable, but prices for advanced models are increasing. Data shows American AI software prices rising roughly 20% to 37% (December data from Tropic). Microsoft-owned GitHub announced shifts from flat-rate plans to more expensive usage-based models. Anthropic has proposed dropping Claude Code from its Pro plan while altering usage for Pro and Max users.

The surge in prices is closely tied to AI adoption intensity. AI does not read language like humans; it processes tokens. Each input is tokenized and sent to GPUs for computation. Modern AI systems require substantial compute, and compute needs scale with token volume and task complexity. When usage is heavy or complex, electric and water bills rise accordingly.

Firms like Cognizant have expanded compute access, helping lower token prices in some cases (for example, Grok 4.1 models charge about $0.20 per 1 million input tokens and $0.50 per 1 million output tokens). Yet token volumes are rising exponentially, so total AI bills trend upward even as nominal token prices fall. To illustrate, a user summarizing a 1000-page book involves tokenizing the document and processing it on GPUs, which raises costs with input size and output requirements. Anthropic’s Claude Opus 4.6 exemplifies higher usage costs: $5.00 input and $25.00 output per million tokens.

China’s DeepSeek counters with strong efficiency: its V4 Flash runs at roughly $0.14 per million input tokens and $0.28 per million output tokens, aided by sparse architectures like Mixture of Experts that activate only a fraction of the model per query. This efficiency highlights a key contrast with US models but remains a source of partial reassurance for India, given that the US currently dominates the market with roughly 100 million ChatGPT users.

The risks from rising subscription costs are not immediate for India, but domestic compute expansion is crucial to avoid ongoing exposure to upstream disruptions. If workloads continue to be processed largely on foreign-owned infrastructure, India faces potential currency and pricing shocks. A long-term path of domestic capability building remains the ideal solution, even as foreign alternatives and short-term cost pressures persist.

Even if India does not deploy frontier AI systems soon, substantial compute access will still be required to train domain-specific models. While compute infrastructure is expanding, a large portion remains foreign-owned, with India’s data-center sector attracting roughly 13-15 billion dollars of investment between 2020 and 2024, about 80% of which came from foreign entities. This means returns would largely flow abroad unless domestic capacity is expanded.

In this context, India sits at a crossroads. Staying with American models implies enduring cost shocks and geopolitical exposure; moving entirely to Chinese alternatives trades one form of dependency for another. Building domestic capabilities will take years, but it is the preferred path. Reliance is already pushing into the space, and India should maximize the use of available compute now by fine-tuning open-access base models on India-specific domains while investing in compute-efficient research, much as China is pursuing.

Even as domestic compute expands, data-center growth alone does not guarantee cheaper compute. Affordable AI depends on multiple factors, including low-cost energy, access to GPUs, and innovations in cost-effective compute utilization. India will require a holistic policy framework ensuring affordable power, adequate water supply, semiconductor access, and support for deep-tech startups, all while balancing climate goals. The overarching approach should deploy multiple policy levers in coordinated fashion to preserve long-term AI resilience. Amit Kapoor (Chair) and Mohammad Saad (Researcher) at the Institute for Competitiveness present this view, noting that the opinions expressed are their own and do not reflect those of the Economic Times.