Image Source: ChatGPT
The gap between the best open-weight AI model and the most advanced closed (proprietary) one has narrowed to a handful of index points. On the Artificial Analysis Intelligence Index (as of 2 August 2026), the leading open-weight model, Kimi K3 Max, scores 60, just one point below the highest-scoring proprietary model, Claude Opus 5. At the end of 2024, the divergence was over 10 points.1 Across the three leading models in each category, open-weight models average 52.7 on the Index, against 60.3 for proprietary models, while their hosted Application Programming Interface (API) access costs approximately one-fifth as much. The input that once cost tens of billions of dollars to create can now be downloaded and repurposed.
Benchmark tables are useful for comparing model capabilities, while adoption rates provide a clearer picture of actual usage. The narrowing capability gap between open-weight and proprietary models is already producing a measurable shift in adoption. OpenRouter, the leading vendor-neutral AI model gateway, published An Empirical 100 Trillion Token Study in December 2025. The study tracked weekly token volume by category and found that open-weight models (primarily Chinese), which accounted for as little as 1.2 percent of token volume in late 2024, approached 30 percent of all token usage in some weeks during 2025.
OpenRouter’s study also showed that releases from DeepSeek, Qwen, Moonshot AI and OpenAI drove rapid and sustained increases in usage, suggesting that developers were moving production workloads rather than merely experimenting. Proprietary models still accounted for most usage, particularly in structured business and reliability-sensitive applications. However, open-weight models have moved beyond their role as a specialist alternative to become a substantial second pillar of the model economy. While OpenRouter’s study measured only inference routed through its own marketplace and did not capture models run locally, on private infrastructure or through other cloud providers, its findings serve as a strong indicator of developer adoption.
The contest is no longer just about who can stretch the frontier, but about who can do the most useful thing at the right price for the application layer on top of it.
The structural force here is open innovation itself, not the nationality of any one lab. Apart from China, the United States now ships capable open weights of its own, led by Meta, Nvidia, and OpenAI. What has changed for every developer, in Bengaluru as much as in San Francisco, is that frontier-class ability is no longer gated behind vendor pricing. Open models now run at a fraction of the cost of a leading closed model for the same unit of work. Model capability is, in effect, becoming a commodity layer. The contest is no longer just about who can stretch the frontier, but about who can do the most useful thing at the right price for the application layer on top of it.
India’s Opportunity Lies Beyond the Frontier
With frontier-class weights now downloadable, the highest returns on India’s limited compute and capital are likely to lie downstream. The IndiaAI Mission has committed over INR 10,000 crore and deployed more than 38,000 GPUs, with further capacity on the way. India’s sovereign models — Sarvam’s 30-billion- and 105-billion-parameter systems, and BharatGen’s 17-billion-parameter Param2 — were built on efficient Mixture-of-Experts designs using modest compute. Open weights can now drastically accelerate Indian sovereign AI development by helping domestic developers bypass the massive cost of pre-training base architectures through strategies such as transfer learning, adaptation, and fine-tuning.
The value in the next phase accrues to whoever turns a general open model into applied, reliable, domain-specific systems. Those who can access proprietary data and enterprise workflows from inside the customer’s premises can rapidly own the applied AI layer. An Indian company can take a frontier-class open base, fine-tune it on proprietary data for a bounded domain such as semiconductor design, law, healthcare, education, or public services, and deploy a customised model that reasons more accurately within that domain. India’s sovereign stack has already begun this move from general language models towards verticalisation, with agriculture– and healthcare-specific mandates emerging through 2026.
Open weights can now drastically accelerate Indian sovereign AI development by helping domestic developers bypass the massive cost of pre-training base architectures through strategies such as transfer learning, adaptation, and fine-tuning.
Private or hybrid cloud deployment is where this becomes India’s structural edge. Open weights can run entirely inside a customer’s defined infrastructure with security controls on access and data, which resolves the single largest objection enterprises have to using models from the frontier labs. Routing prompts and proprietary data through a hosted remote server exposes sensitive information to that provider’s jurisdiction, while deploying open weights on the customer’s infrastructure means the data never leaves the perimeter and governance is enforceable. The winning architecture is therefore efficient, vertical, private, secure by construction, and anchored within the jurisdiction.
India is also well placed to translate this advantage into practical gains. Its Global Capability Centres (GCCs) have become the layer through which global enterprises trial and act on technology decisions. India now hosts more than 2,100 GCCs, employing more than 2 million professionals and generating almost US$ 100 billion in revenue. The world’s densest GCC network gives Indian innovators a global market opening that the enterprise software era never offered. An Indian company can pilot an on-premises, open-weight-derived vertical model for a global enterprise through its India centre, customise it for proprietary data and workflows, secure sign-off in parallel, and convert a local pilot into a global contract. This is the practical shape of building for the world from India: private, applied AI, proven through the GCC, and then deployed across the customer’s global operations.
Building the Compute to Scale AI
The capacity to run and refine these models at home is still being built. India’s data centre capacity is projected to rise from 2.2 GW to 12 GW by 2030. Reliance Industries has committed about US$110 billion over seven years in data centres and adjacent infrastructure, while the Adani Group has announced US$100 billion investments. This data centre capacity will be crucial for Indian organisations to host and run large open-weight models domestically, fine-tune them for specific sectors, and serve them securely at population scale.
India’s AI advantage will be won in application, data, customisation and deployment, as convergence among open-weight models shifts value away from the base model itself.
India’s AI advantage will be won in application, data, customisation and deployment, as convergence among open-weight models shifts value away from the base model itself. IndiaAI compute and expanding data centre capacity provide the infrastructure to adapt and host these models domestically. Proprietary Indian data, vertical expertise and private, in-jurisdiction deployment can create durable advantages, while GCCs offer a ready pathway from domestic pilots to global contracts. The test ahead is whether Indian innovators can bring these elements together to build secure, scalable products for India and the world.
Nisha Holla is a Visiting Fellow at Observer Research Foundation and Research Fellow at 3one4 Capital.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.