Chinese language open-source fashions, which had been already common amongst startups as a result of decrease prices, could now make their technique to massive enterprises trying to keep away from disruptions to top-dollar synthetic intelligence tasks. “The selection is pragmatic, not political,” mentioned Nipun Kalra, managing director and senior associate at BCG.
Functionality vs prices
The development is “pushed by price, robust efficiency and licences that permit enterprises self-host,” mentioned Kalra.
Final week, China’s Ziphu AI launched the GLM-5.2 mannequin, which sits fairly near Anthropic’s Opus 4.8 on efficiency benchmarks. The discharge has sparked a ‘DeepSeek-like’ second as a result of GLM is priced at a fifth the price of Opus.
One other Chinese language agency, 360 Safety Expertise, unveiled two AI safety instruments below the banner Yitian Tulong — named after Chinese language epic novel Tulongfeng — that routinely uncover software program 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 mannequin.
Reviews declare these Asian labs are head-to-head with Anthropic and OpenAI.
A brand new JPMorgan evaluation confirmed that Chinese language fashions now price 10-50 instances much less per token than frontier fashions.
These fashions from corporations together with DeepSeek, Alibaba, Xiaomi, MiniMax and Moonshot now dominate what it describes because the trade’s ‘intelligence-per-dollar’ frontier, the place efficiency is balanced in opposition to working prices, the report mentioned.
“We’re seeing actual momentum – however the shift is in direction of open-weight fashions, and by unbiased benchmarks a lot of the nicely performing ones immediately occur to be Chinese language,” BCG’s Kalra mentioned. He, nonetheless, mentioned western closed fashions nonetheless lead in frontier workloads. “Most enterprises are going multi-model and tiered, not switching wholesale.”
A current joint analysis by the Massachusetts Institute of Expertise (MIT) and Hugging Face revealed that Chinese language fashions accounted for 17.1% of worldwide downloads in 2025, surpassing the US share of 15.8%.
“On functionality, Chinese language open-weight fashions have gotten extraordinarily spectacular,” mentioned Nikhil Narendran, associate at legislation agency Trilegal and a tech coverage thought chief. “They’re far cheaper, can usually be downloaded and self-hosted, and the hole with frontier methods is narrowing quick.”
He defined that open-weights can run inside an enterprise surroundings and due to this fact have a privateness benefit.
Belief issue
However whether or not enterprises can belief Chinese language open-source fashions stays an open query.
“If India merely responds to American dependency by dashing into Chinese language dependency, now we have not achieved sovereignty. We now have merely modified our landlord,” Narendran mentioned.
This month, the US authorities restricted entry of Anthropic’s Fable 5 and Mythos 5 fashions to overseas nationals. Final week, it requested OpenAI to restrict the preview rollout of its GPT-5.6 household – Sol, Terra and Luna – to a small group of government-vetted companions.
“That is most likely the primary time that ethnicities have turn out to be a part of geopolitical technique in a means that determines who can entry vital applied sciences, and that might have far-reaching implications for R&D, hiring, funding and international collaboration,” mentioned Abishur Prakash, an creator and geopolitical strategist at Canada-based advisory agency The Geopolitical Enterprise Inc. “To me, the larger query is how this trickles into extra division throughout the international economic system and society.”
Consultants mentioned such restrictions underscore the necessity for India to construct a sovereign AI technique to make sure entry to newest frontier fashions and strengthen home capabilities.
“The true lesson of current weeks is focus threat: dependence on any single supplier can change by coverage in a single day,” BCG’s Kalra mentioned. “For India, the precedence is optionality and sovereign functionality.”