Global businesses are increasingly switching from premium, closed-source US software – such as OpenAI’s GPT and Anthropic’s Claude – to cheaper Chinese open-weight models that offer near-frontier performance.
Since mid-June, the daily token volume of Zhipu’s GLM-5.2, which operates at about one-fifth the cost of Anthropic’s Claude Opus 4.8, had surged 50-fold on Vercel, the San Francisco-based cloud platform for AI web development reported on Tuesday.
Meanwhile, DeepSeek’s V4 Flash, a streamlined version of the firm’s flagship V4 Pro, had emerged as the single largest model by volume on the gateway, capturing more than 20 per cent of the platform’s traffic on Wednesday, up from about 15 per cent a month ago.
Open-weight models accounted for 29 per cent of token volume on its AI Gateway platform, nearly tripling their share since April, according to Vercel.
The rise of these cost-efficient powerhouses marks a growing shift in how companies worldwide source artificial intelligence. While proprietary, closed-source models require premium cloud subscriptions and charge by the token – the basic unit of data processed by an AI – open-weight models allow companies to download code for free and run it on local hardware.
The fast-improving capabilities of these free-to-download models have forced businesses to re-evaluate their tech spending. Until recently, enterprises chose to absorb the high cost of premium American models because open alternatives lagged too far behind – a gap that is narrowing fast.
In a research report published last week, Goldman Sachs analysts noted that Chinese open-weight models “are reaching a critical point of intelligence performance” relative to proprietary models, driving “a significant ramp-up” in enterprise adoption at home and globally.
The investment bank forecast that daily token consumption of these Chinese models would skyrocket from 350 trillion this year to 4,600 trillion in 2030, with international users expected to account for 55 per cent of the total volume.
The uptick in interest in open-weight models was emerging among larger global enterprises as rising AI token costs had become “a real concern for most organisations”, according to a UBS report published last month.
The investment bank cited the chief technology officer of an unnamed “large global bank”, who said the financial institution had begun hosting open models from Alibaba Group Holding’s Qwen series to manage token spending and “balance our use of premium models like Claude”. Alibaba owns the South China Morning Post.
Corporate giants in other sectors are taking a similar approach. Brian Armstrong, co-founder and CEO of Coinbase, the largest cryptocurrency exchange in the United States, posted on X last month that the firm was experimenting with setting open-weight options – such as GLM 5.2 and Chinese AI lab Moonshot’s Kimi 2.7 – as default models on its AI platform.
The strategy to “keep AI spend flat” would still allow internal engineers to choose different models depending on the task, he added.
The shifting preference for Chinese open-weight models has led some to question the future valuation of US AI labs and cloud giants. In a note last week, Apollo Global Management chief economist Torsten Slok wrote that if Chinese models continued to gain and token prices continued to fall, expected hyperscaler cash flows could prove too optimistic.
“The progress of Chinese open-source models demonstrates that the ‘raw intelligence’ provided by LLMs [large-language models] – in the form of inference or coding capabilities – will likely become increasingly commoditised,” said Dong Chen, Asia chief investment officer at Bank J. Safra Sarasin.
Companies producing LLMs would have to “compete strictly on cost”, meaning that the stand-alone business could become a low-margin industry, Dong added.
Still, many argued the global rise of Chinese open-weight models would have a limited impact on the US market. Chinese AI adoption among American businesses remained “very low” and was “concentrated among highly AI-intensive firms”, New York-based corporate spending platform Ramp said in an article last week.
Daniel Yue, assistant professor at the Georgia Institute of Technology’s Scheller College of Business, noted that Anthropic and OpenAI’s brands “have only strengthened over the last year”, adding that many companies strongly preferred to work with US providers through direct contracts and service agreements.
“US AI labs will still hold their ground despite Chinese competition,” said Dong of Bank J. Safra Sarasin. “Geopolitical fragmentation and Western regulations will likely provide a sufficient barrier against Chinese models in sensitive sectors.”
But in more cost-conscious segments of the market, Chinese models were poised to thrive, Dong stated. -- SOUTH CHINA MORNING POST
