Mint Explainer | Why open-weight AI is gaining ground over proprietary models

New Delhi: On 2 July, Nasdaq-listed software company Palantir Technologies' CEO Alex Karp argued that some enterprises are moving away from proprietary artificial intelligence (AI) models in favour of open-weight AI models, citing high costs. Speaking to CNBC, Karp said many enterprise customers are frustrated with token-based pricing, the model used by companies such as OpenAI and Anthropic, where users are charged based on the amount of text processed by an AI model. Indian AI startups, however, have been using open-source models to build enterprise AI applications for some time. Mint decodes why. What are proprietary and open-weight models? Proprietary AI models are developed and controlled by companies, with their model weights, training data, and code kept private. Users access them through APIs and typically pay based on usage. Examples include models from OpenAI, Anthropic, and Google. Open-weight models make their trained weights publicly available, allowing developers to download, fine-tune, and deploy them on their own infrastructure. However, the training data and training process are usually not released. Examples include GLM (developed by Zhipu AI), Qwen (Alibaba Cloud), Kimi (Moonshot AI), and DeepSeek (DeepSeek AI). In terms of cost, proprietary models are generally more expensive than open-weight models. For example, GPT-5.5 costs $5.50 per million input tokens, while DeepSeek-R1 costs $1.35 per million input tokens. Input tokens are text, numbers, or code that are fed into an AI model. Illustrating the difference further, Additi Upadhyay, co-founder of AI startup, Noveum AI, said, "Take a frontier proprietary model like GPT-4o versus an open-weight model like DeepSeek V3. GPT-4o runs about $2.50 per million input tokens and $10 per million output tokens (generated by AI models). For the same amount of tokens, DeepSeek V3 is roughly $0.27 for input and $1.10 for output. That's close to a 9x difference for almost the same quality on a lot of everyday tasks." Why are Indian start-ups shifting to open weight models? For Indian startups, the shift is driven largely by cost. India's deeptech ecosystem attracts far less funding than markets like the US, forcing founders to build with limited capital. According to Tracxn, Indian deeptech startups raised $1.47 billion in 2025, compared with $179 billion in the US. With proprietary AI models, companies have to pay each time they use them through an API (a way for one app to connect to another company's AI over the internet). This works well when they're just testing things out, but if many people start using their product, costs add up quickly. Open-weight models work differently. Instead of paying per use, companies run these models on their own servers or cloud systems, so they mainly just pay for the computing power needed to run them. As a result, for small or early-stage use, proprietary APIs are usually cheaper. But as a product grows and more people use it, open-weight models tend to save a lot more money, experts said. "Startups that have raised ₹5 crore or ₹10 crore obviously cannot spend endlessly. We had to understand the technology and build our own systems more frugally because Indian customers also pay less. That was the whole model," said Kalyani Khona, angel investor and AI researcher. How does geopolitics play a role? Geopolitical tensions and tighter regulations have also boosted interest in open-weight models. As enterprises move AI into production, many are wary of relying entirely on APIs controlled by a handful of companies. The concerns were highlighted in June when the US temporarily restricted foreign access to Anthropic's Claude Opus 5 before lifting the curbs after additional safeguards were introduced. The episode showed that access to frontier AI models can be shaped by government policy, export controls and licensing decisions. Experts said enterprises are increasingly prioritising greater control over their AI infrastructure, data and costs. "Recent restrictions and reversals around Anthropic’s Fable and Mythos models showed enterprises that access to frontier models can be affected by export controls, licensing, safety decisions and government policy," said Kashyap Kompella, founder of RPA2AI, an industry analyst firm. “With open weight, a startup can host the model on its own infrastructure, on cloud infrastructure, or within a customer's private environment. It can fine-tune the model, optimize it for latency, route only certain tasks to it, and avoid being fully dependent on one proprietary API provider,” Kompella added. Will the trend continue? Most founders believe the future of enterprise AI will not be defined by a choice between proprietary and open-weight models. Instead, enterprises are expected to adopt a hybrid approach, selecting different models based on cost, performance and the complexity of the task. By fine-tuning open-weight models on proprietary datasets, startups can build domain-specific AI systems that outperform larger general-purpose models in specialised use cases while keeping costs under control. For many companies, the customised model itself becomes a key competitive differentiator. “Most start-ups will go hybrid: frontier closed models for the hardest reasoning, open-weight models for the high-volume, latency-sensitive, or privacy-sensitive work. The gap has narrowed enough that open models now deliver most of the capability for a fraction of the cost, so for many production workloads, "good enough, cheaper, and fully under my control" simply wins,” said Upadhyay of Noveum AI.