Synopsis
Chinese open-source models, which were already popular among startups due to lower costs, may now make their way to large enterprises looking to avoid disruptions to top-dollar artificial intelligence projects. “The choice is pragmatic, not political,” said Nipun Kalra, managing director and senior partner at BCG.Listen to this article in summarized format
Chinese open-source models, which were already popular among startups due to lower costs, may now make their way to large enterprises looking to avoid disruptions to top-dollar artificial intelligence projects. “The choice is pragmatic, not political,” said Nipun Kalra, managing director and senior partner at BCG.
Capability vs costs
The trend is “driven by cost, strong performance and licences that let enterprises self-host,” said Kalra.
Last week, China’s Ziphu AI launched the GLM-5.2 model, which sits quite close to Anthropic’s Opus 4.8 on performance benchmarks. The release has sparked a ‘DeepSeek-like’ moment because GLM is priced at a fifth the cost of Opus.
Another Chinese firm, 360 Security Technology, unveiled two AI security tools under the banner Yitian Tulong — named after Chinese epic novel Tulongfeng — that automatically discover software vulnerabilities, and Yitianzhen to automate cyber defence and incident response.
Japan’s Sakana AI launched Fugu, a multi-agent orchestrator system that works as a single model.
Reports claim these Asian labs are head-to-head with Anthropic and OpenAI.
A new JPMorgan analysis showed that Chinese models now cost 10-50 times less per token than frontier models.
These models from companies including DeepSeek, Alibaba, Xiaomi, MiniMax and Moonshot now dominate what it describes as the industry's ‘intelligence-per-dollar’ frontier, where performance is balanced against operating costs, the report said.
“We're seeing real momentum – but the shift is towards open-weight models, and by independent benchmarks most of the well performing ones today happen to be Chinese,” BCG’s Kalra said. He, however, said western closed models still lead in frontier workloads. “Most enterprises are going multi-model and tiered, not switching wholesale.”
A recent joint research by the Massachusetts Institute of Technology (MIT) and Hugging Face revealed that Chinese models accounted for 17.1% of global downloads in 2025, surpassing the US share of 15.8%.
“On capability, Chinese open-weight models are becoming extremely impressive,” said Nikhil Narendran, partner at law firm Trilegal and a tech policy thought leader. “They are far cheaper, can often be downloaded and self-hosted, and the gap with frontier systems is narrowing fast.”
He explained that open-weights can run inside an enterprise environment and therefore have a privacy advantage.
Trust factor
But whether enterprises can trust Chinese open-source models remains an open question.
“If India merely responds to American dependency by rushing into Chinese dependency, we have not achieved sovereignty. We have merely changed our landlord,” Narendran said.
This month, the US government restricted access of Anthropic's Fable 5 and Mythos 5 models to foreign nationals. Last week, it asked OpenAI to limit the preview rollout of its GPT-5.6 family – Sol, Terra and Luna – to a small group of government-vetted partners.
“This is probably the first time that ethnicities have become part of geopolitical strategy in a way that determines who can access critical technologies, and that could have far-reaching implications for R&D, hiring, funding and global collaboration,” said Abishur Prakash, an author and geopolitical strategist at Canada-based advisory firm The Geopolitical Business Inc. “To me, the bigger question is how this trickles into more division within the global economy and society.”
Experts said such restrictions underscore the need for India to build a sovereign AI strategy to ensure access to latest frontier models and strengthen domestic capabilities.
“The real lesson of recent weeks is concentration risk: dependence on any single provider can change by policy overnight,” BCG’s Kalra said. “For India, the priority is optionality and sovereign capability.”
