A significant transformation is currently unfolding in the global open-source AI ecosystem. As Western AI companies slow down or restrict their open model releases, Chinese developers are rapidly filling that vacuum. The consequence is that control over the open-weight AI space is gradually shifting.
Western players such as OpenAI, Anthropic, and Google are facing intensified regulatory scrutiny, commercial expectations, and safety-related constraints. As a result, these companies are increasingly prioritizing API-based access rather than releasing fully downloadable model weights. In contrast, Chinese laboratories are openly publishing powerful open-weight models that can run efficiently on ordinary hardware.
This divergence is quietly reshaping global AI infrastructure.
What the Data Indicates About This Shift
A recent security study conducted by SentinelOne in collaboration with Censys analyzed 175,000 exposed AI hosts across 130 countries over a period of 293 days. The findings were striking: Alibaba’s Qwen2 model consistently occupied the second position in global deployment surpassed only by Meta’s Llama family.
Most notably, Qwen2 appeared in 52% of systems where multiple AI models were running simultaneously. In multi-model configurations, the Llama and Qwen2 combination was the most prevalent. This clearly indicates that Qwen2 has emerged as the default alternative to Llama.
Another significant observation was the stability of Qwen2’s ranking. Across every measurement methodology total observations, unique hosts, and host-days its position remained consistent. No substantial regional volatility was observed.
The implication is straightforward: this adoption is not experimental. It reflects serious operational utilization.
Why Chinese Models Are Being Adopted More Widely
The driving force is not ideology, but pragmatism.
Chinese AI laboratories are releasing large, high-quality models that are meticulously optimized for:
• Local deployment
• Efficient execution through quantization
• Smooth operation on commodity GPUs and CPUs
• Edge and residential environments
For independent developers, startups, and researchers who cannot afford expensive cloud infrastructure, these models have become a compelling alternative.
Western frontier laboratories, by contrast, prefer to provide API-based access rather than releasing open weights. While this approach enhances safety and monitoring, it reduces developer autonomy particularly for those who wish to operate AI systems on their own infrastructure.
For operators seeking to run self-hosted AI systems without subscription fees or imposed restrictions, Chinese open-weight models are increasingly the more practical choice.
Geographic Deployment Patterns
The study also revealed a clear concentration in deployment patterns.
Within China, Beijing accounts for approximately 30% of exposed hosts, followed by Shanghai and Guangdong. In the United States, Virginia a major cloud infrastructure hub represents approximately 18% of deployments.
China and the United States have most of the exposed AI systems compared to other countries. However, the momentum in open-weight model releases appears to favor Chinese-origin models.
If disparities in openness, portability, and release velocity persist, Chinese model families may become the default foundation for open deployments not due to ideology, but because they are readily accessible and operationally viable.
What Is Governance Inversion
This shift introduces a novel challenge referred to as “governance inversion.”
Centralized platforms such as ChatGPT operate under unified corporate control infrastructure, safety mechanisms, misuse monitoring, and access regulation remain consolidated within a single entity.
Open-weight models function differently.
Once a model is publicly released:
• Anyone can download and modify it
• Safety mechanisms can be removed
• Authentication can be disabled
• Monitoring becomes entirely decentralized
According to the study, 175,000 exposed hosts are operating outside commercial AI governance frameworks. There is no centralized authentication, no unified abuse-reporting infrastructure, and no emergency shutdown mechanism.
Moreover, ownership of approximately 16–19% of infrastructure could not be identified. In the event of misuse, accountability becomes ambiguous.
This ecosystem also contains a stable backbone approximately 23,000 hosts maintaining 87% uptime. These are not casual hobby projects, but serious, long-term operational systems.
Tool-Enabled AI and Escalating Risk
Nearly 48% of exposed hosts were operating with tool-calling capabilities. This means these systems are not limited to text generation; they can:
• Execute code
• Invoke APIs
• Access databases
• Interact with external systems
When AI transitions from merely generating responses to performing actions, the risk profile increases substantially.
If such systems lack password protection, a malicious prompt could:
• Summarize internal documents
• Extract sensitive information
• Access connected services
Approximately 26% of hosts were running “thinking” models optimized for multi-step reasoning. In at least 201 deployments, safety guardrails had been intentionally removed.
In numerous cases, even basic password protection was absent.
This effectively transforms AI into an execution layer capable of remote operation without centralized oversight.
What Western Labs Can Do
Western AI laboratories cannot fully control decentralized deployment. However, they can recalibrate their strategies.
They can:
• Monitor post-release adoption
• Track misuse patterns
• Study the safety implications of quantization
• Establish voluntary reporting channels
At present, decisions made by a limited number of model providers are influencing thousands of decentralized systems.
Ignoring this reality may diminish long-term strategic influence.
What May Happen in the Next 12–18 Months
Experts anticipate that the open AI layer will become more stable and professionalized. Tool utilization, autonomous agents, and automation are likely to become default features.
While hobby experimentation will persist, the core backbone will:
• Become more powerful
• Handle more sensitive workloads
• Operate with greater autonomy
Traditional governance mechanisms do not effectively extend to residential or small VPS deployments, resulting in uneven enforcement.
This is not merely a configuration issue. It represents the emergence of a global, unmanaged AI compute layer.
And there is no centralized switch capable of disabling it.
Geopolitical Impact
If the world’s unmanaged AI compute increasingly depends on models originating from non-Western laboratories, traditional influence structures may be reconfigured.
Western companies may maintain strict governance within their own platforms, yet if dominant open capabilities are emerging elsewhere, their real-world leverage may be constrained.
Open-source AI is becoming global, but its center of gravity is gradually shifting eastward.
This shift is not the product of grand strategy, but of basic economics and operational practicality. Developers gravitate toward solutions that are accessible, affordable, and capable of running efficiently on their existing hardware.
At present, Chinese AI laboratories are providing precisely that within the open-weight domain.
A Silent Structural Transformation
The 175,000 exposed hosts that have been mapped may represent only the surface of a deeper structural metamorphosis.
While Western laboratories prioritize safety and regulatory compliance, Chinese developers are consolidating their position in the open-source domain through availability, performance optimization, and portability.
The long-term consequences remain uncertain. However, one conclusion is evident:
The future of open-source AI will belong to those who release models that anyone can readily execute.
And at this moment, Chinese AI laboratories are advancing decisively in that direction.
