A former Meta Platforms product manager has triggered a significant debate over enterprise AI adoption, suggesting that American and European enterprises may gradually move away from proprietary models developed by OpenAI and Anthropic to adopt self-hosted Chinese AI alternatives. According to Xiaoyin Qu's post on X, Chinese open-weight AI models are gaining traction among enterprises because they provide greater control over infrastructure, data management and operational costs. The former Meta executive argues that models such as DeepSeek and Alibaba's Qwen can be deployed on a company's own GPU infrastructure, allowing organisations to keep sensitive information within their networks while meeting governance and regulatory requirements. Self-hosted models also enable businesses to fine-tune AI systems using proprietary datasets, helping create a competitive advantage that rivals may find difficult to replicate. Recent developments show this trend accelerating as soaring AI bills are pushing companies to prioritise affordability over marginal improvements in model performance, with cheaper alternatives increasingly competing with premium frontier models.
Chinese AI models are gaining ground on Anthropic and OpenAI after Z.ai's GLM-5.2 closed the frontier gap. GLM-5.2 now sits within a single percentage point of Anthropic's Opus 4.8 on a closely watched agentic evaluation, with the model designed with 750 billion parameters and a 1-million-token context window. The system runs entirely on domestic Chinese chips, a critical detail given ongoing United States export restrictions. Z.ai shares surged more than 30% in Hong Kong trading and now sit up over 800% since debuting in January, with JP Morgan projecting Z.ai revenue to expand by more than 534% this year. OpenRouter, a popular AI aggregator platform, now shows that Chinese models hold the top four positions among the most widely used systems globally, with DeepSeek, MiniMax, Tencent, and Xiaomi collectively passing every major US frontier provider by token traffic. China's Z.ai demonstrated bug-finding and vulnerability-detection capabilities approaching Anthropic's Mythos-class systems, with security researchers citing that the model can match Mythos in some software vulnerability benchmarks, even though it continues to trail leading models from Anthropic and OpenAI in broader reasoning and general-purpose tasks.
The Trump administration has implemented unprecedented government oversight of AI products, with both OpenAI and Anthropic restricting releases of their most advanced models at the administration's request. OpenAI's new GPT-5.6 Sol model is now accessible only to customers approved by the Trump administration, with the company stating it doesn't believe this kind of government access process should become the long-term default. Anthropic announced that the Trump administration has approved a limited release of its strongest cybersecurity model, Mythos 5, two weeks after the U.S. Commerce Department effectively banned it. The government lifted restrictions on Mythos 5 on Friday, enabling it to be redeployed to a small group of cyber defenders and infrastructure providers. This represents the latest in an unprecedented government vetting of AI products for cybersecurity risks, with the administration establishing a framework for vetting national security risks of advanced AI systems for up to 30 days before public release.
The cost advantage is the most damaging factor for US labs, with DeepSeek V4 Pro charging $3.48 per million output tokens compared to Anthropic's Fable 5 at $50 for the same output. As a result, enterprise buyers are now openly rethinking their entire AI vendor relationships. The competitive picture remains nuanced, with DeepSeek itself estimating that Chinese models trail leading US systems by 3 to 6 months in terms of pure capability. However, that gap matters less when access becomes the primary risk factor, and pricing determines whether production is viable or token economics are prohibitive. The release timing was anything but accidental, with GLM-5.2 launching a day after Anthropic disabled global access to its most advanced models, including Fable 5 and Mythos, while OpenAI moved to limit access to GPT-5.6 following a separate government request. Speaking to Rest of World, Tiezhen Wang, former head of the Asia-Pacific at Hugging Face, said many companies initially build products using proprietary models before moving to open-source alternatives as they scale, allowing firms to reduce spending on AI tokens substantially.
Despite the predictions of complete migration away from proprietary models, industry observers note that enterprises are increasingly adopting multi-model strategies, combining proprietary frontier models with open-source alternatives. According to Xiaoyin Qu's analysis, large customers often negotiate customised agreements governing the storage and use of their information, with OpenAI and Anthropic offering enterprise-grade services with contractual commitments on privacy, security and data handling. However, the former Meta executive questions whether companies should entrust critical business data to AI providers that continue developing their own commercial products, raising concerns around vendor dependence and data ownership. Businesses face mounting pressure to justify AI spending, with proprietary AI services becoming costly at scale, while open-weight models offer greater flexibility and potentially lower long-term operating costs. Open-source models allow developers to download, modify and deploy systems without relying on a single provider, while the approach may encourage innovation at lower costs, but can make oversight more difficult.
