Open-weight AI models, where the underlying code is freely available for anyone to download and modify, are becoming the new frontier of technological power, with profound implications for countries like India caught between competing global AI ecosystems. At the BRICS Summit in New Delhi on September 13, 2026, China's President Xi Jinping proposed an "open source and inclusive AI initiative," signaling Beijing's strategy to build a unified AI ecosystem across developing nations. The proposal highlights a critical shift: the battle for AI dominance is no longer just about who builds the most powerful models, but who controls the entire technology stack beneath them.

Xi Jinping

What Are Open-Weight Models and Why Do They Matter?

Open-weight models are large language models (LLMs), the AI systems that power chatbots and language tools, whose internal numerical parameters are published freely for download. Unlike proprietary models from companies like OpenAI or Anthropic, anyone with the file can modify it, retrain it on their own data, and run it locally without relying on a company's servers. This fundamentally changes the economics of AI access. For developers in India and other emerging economies, open-weight models eliminate the need to pay per query to a cloud provider, reduce latency by running models locally, and enable customization for local languages and use cases.

The appeal is substantial. Chinese models now account for approximately 41% of all downloads on Hugging Face, the main public repository where AI models are shared, as of spring 2026. Of 178 Chinese model releases with 20 billion parameters or larger this year, 59% carried the Apache 2.0 license and 22% carried the MIT license, both of which permit commercial reuse. Meanwhile, DeepSeek's V4 Pro model costs just 87 cents per million tokens, a significant undercut compared to American alternatives.

How Is Meta's Llama Losing Ground in the Open-Source Race?

Meta's Llama family was once the dominant open-weight model, the foundation that most developers built upon. However, that dominance is eroding. The shift began in 2024 when Alibaba released Qwen 2.5, a Chinese alternative that offered competitive performance at lower cost. Today, the ladder of open-source development runs both ways, with some large American open releases now built on Chinese models rather than Llama as the base. This represents a significant strategic loss for Meta and the broader Western AI ecosystem, as the default starting point for new AI projects increasingly originates from Beijing rather than Silicon Valley.

The practical consequence is that developers worldwide, particularly in cost-sensitive markets, now have viable alternatives to Llama. This fragmentation of the open-source ecosystem means that Meta's influence over how AI models evolve globally is diminishing, even as the company maintains technical credibility in the field.

The Real Problem: Control Beyond the Model Layer

Here is where the geopolitical dimension becomes critical. Open-weight models are only one layer in a much larger technology stack. Below them sit computer chips (GPUs and specialized AI accelerators), cloud platforms that host the models, software frameworks that run them, and industry standards that determine compatibility. Above them sit applications and interfaces that users actually interact with. When China offers open models but controls the chips, cloud infrastructure, and standards, the appearance of openness masks a deeper dependence.

India's manufacturing experience offers a cautionary parallel. Cheap Chinese components transformed Indian electronics and solar industries, but they also created what analysts call the "inverted-pyramid problem." India now has vast downstream assembly capacity but minimal upstream capability. In solar, for example, India has 233 gigawatts of module assembly capacity but only 2 gigawatts of cell-making capacity and virtually no capacity to produce the raw polysilicon wafers. Similarly, an AI ecosystem built on open Chinese models but dependent on Chinese chips, cloud services, and standards would create the same structural imbalance.

Steps to Building Genuine AI Autonomy

• Develop Upstream Capability: India's IndiaAI Mission has onboarded over 38,000 GPUs at subsidized rates of 65 rupees per hour, but advanced processors remain overwhelmingly foreign-designed. Building indigenous chip fabrication capacity, particularly for AI accelerators rather than mature nodes, is essential to avoid dependence on imported hardware.
• Create Domestic Model Development: More than 60% of Indian startups seeking subsidized GPUs want them for inference, running existing models rather than training new ones. Shifting investment toward training indigenous foundation models and building the engineering talent to do so would reduce reliance on foreign model architectures.
• Build Dense Supplier Networks: China's own AI rise came less from replacing imports than from creating dense networks of suppliers, engineering talent, patient finance, and fierce domestic competition. Baidu went from zero model releases on Hugging Face in 2024 to over 100 in 2025. India needs similar ecosystem density, not just model access.

What Does This Mean for India's AI Future?

India occupies an unusual position in this geopolitical contest. In July 2026, 29 countries signed an agreement in Shanghai creating the World Artificial Intelligence Cooperation Organization (WAICO), a China-led body for AI coordination. India was not among them. In February, India joined Pax Silica, the US-led initiative spanning critical minerals, chip fabrication, and AI infrastructure. New Delhi then chaired the BRICS Summit where Beijing made its pitch, yet the New Delhi Declaration did not endorse it. For now, Xi's proposal remains a Chinese initiative, with China hosting BRICS in 2027.

The real question for India is not "Chinese or American technology?" but rather how dependence is distributed across the entire stack. Cheap access to open models can accelerate development and lower costs for Indian firms and startups. But without building capability in chips, cloud infrastructure, software frameworks, and standards, India risks reproducing the inverted-pyramid problem that has defined its relationship with Chinese manufacturing for two decades. Openness at the model layer does not decide who controls the hardware, cloud, standards, and institutions through which models are deployed.

The Parallel Shift Toward On-Device Intelligence

Interestingly, a parallel technological shift is underway that could reshape this entire dynamic. Rather than relying on cloud-based models at all, researchers are demonstrating that capable AI models can run entirely on personal devices like smartphones. A proof-of-concept architecture shows that a Gemma 4 model can run on an Android handset in airplane mode, generating roughly 40 tokens per second, fast enough to feel instant to users. This removes four structural dependencies at once: cost (no per-query fees), latency (no network round trip), privacy (data never leaves the device), and dependence (no remote provider can change or withdraw the service).

If on-device AI becomes mainstream, it could bypass the entire geopolitical contest over cloud infrastructure and model access. Users would own their intelligence locally, and the competitive advantage would shift to whoever builds the best software architecture to run models efficiently on consumer hardware, rather than whoever controls the largest data centers or the most advanced chips. This represents a potential third path for countries like India, one that sidesteps the need to choose between Chinese and American ecosystems by enabling genuine local autonomy.

For now, however, the immediate reality is that open-weight models are reshaping global AI development, Meta's Llama is losing its default status, and the question of who controls the layers beneath the model remains the defining strategic challenge for emerging AI economies.