US Delegation for Trump-Xi Meetings 

Jensen Huang’s last-minute addition to Trump’s Beijing delegation turned that subtext into the headline. According to news agency Reuters, Trump asked Huang to join the trip at the last minute, and Huang was seen boarding Air Force One during a refuelling stop in Alaska.  

It becomes evident that NVIDIA is no longer just a chip company in this geopolitical equation; it is the bridge, the bargaining chip and the constraint. For Washington, NVIDIA is leverage over China’s AI frontier. For Beijing, NVIDIA remains useful but increasingly risky if dependence slows domestic alternatives. That is why the meetings were not only about whether China can buy H200s. 

NVIDIA remains central to China’s AI trajectory, but US export controls have already pushed the country toward a more sovereign, non-NVIDIA stack. 

The Newly Cleared NVIDIA H200 Access 

During the Trump-Xi meetings, the US cleared around 10 Chinese companies, including Alibaba, Tencent, ByteDance and JD.com, to buy NVIDIA’s H200 chips, with Lenovo and Foxconn also approved as distributors. Each approved customer could buy up to 75,000 H200s under US licensing terms. 

The H200 clearance shows Washington is testing a more transactional model of AI-chip control, while Beijing’s hesitation to place orders shows China increasingly sees NVIDIA dependence as a strategic vulnerability. This is why the summit marks a transition – NVIDIA remains central but no longer uncontested. The question is shifting from “Can China get NVIDIA chips?” to “How much NVIDIA can China still use without slowing its own sovereign AI stack?” 

A short answer to that would be: China wants NVIDIA access today so it can need NVIDIA less tomorrow. 

Section II: The China AI Equation 

China AI now rests on five connected pillars – DeepSeek V4, non-NVIDIA compute, AI sovereignty, open model/agent ecosystem, and faster domestic innovation (as shown in the chart below). 

China AI Five Pillars 

Source: Counterpoint Research

DeepSeek V4 is the visible variable because it gives the market a concrete proof point – China can still move toward frontier AI under compute constraints. But the hidden story is non-NVIDIA compute. Once frontier models can run natively on Huawei, Cambricon and other domestic chips, AI sovereignty becomes less rhetorical and more operational.  

That sovereignty is also pushing China toward a more open model and agent ecosystem. Open deployment widens adoption, reduces the cost of experimentation and brings more developers into the stack. The early evidence is already visible. According to the latest OpenRouter data for May, Hermes Agent topped the rankings with 271 billion daily tokens consumed, overtaking OpenClaw. More importantly, Xiaomi’s open-source MiMo-V2-Pro has become the leading model powering Hermes Agent. This shows how quickly China’s open-model ecosystem can compound once strong models, agent platforms and domestic compute begin reinforcing each other. 

Global Ranking by Token Usage

Source: OpenRouter 

China's Open Models Leading Hermes Agent Usage

The intended outcome is faster domestic innovation. The Trump-Xi meetings sit in the middle as a transition window – NVIDIA access may still matter, but China’s trajectory is no longer defined by NVIDIA alone. 

The bigger picture is becoming clearer – better models validate domestic compute; domestic compute strengthens sovereignty; sovereignty accelerates open ecosystems; open ecosystems feed back into faster model and application innovation. 

Section III: What Happens Next 

We assign an 85% probability that China doubles down on non-NVIDIA compute, because H200 access may reopen tactically and remain conditional, political and reversible. Huawei, Cambricon and other domestic chip suppliers do not need to beat NVIDIA chip-for-chip in the near term. They need to become good enough for large-scale inference, agent deployment, model distillation and targeted frontier training. That is a lower but more commercially relevant bar. If DeepSeek V4 can run meaningfully on domestic chips, China’s most direct goal is to build a sufficiently dense domestic compute fabric across cloud, enterprise, government and consumer applications, rather than to find an NVIDIA alternative. 

China AI Trajectory Over the Next 6-12 Months

Source: Counterpoint Research 

We assign a 75% probability that open models and agent ecosystems will become China’s main deployment weapon. The US advantage remains strongest in closed frontier labs, premium developer platforms and advanced chips. China’s advantage is more likely to emerge through open deployment, rapid model adaptation and agent ecosystems. The more models are opened, forked, distilled and deployed, the faster the ecosystem learns. 

Our probability-weighted view is that China’s AI trajectory is entering a system-building phase. We assign a 90% probability that the AI race shifts from benchmark competition to system competition, because the next source of advantage will come from connecting models, chips, cloud infrastructure, agents and deployment channels into one scalable stack. Sovereign AI is not so much a slogan as the ability to keep improving, deploying and monetizing AI even when external compute supply is uncertain. 

In conclusion, the most likely outcome is a hybrid stack for AI compute. Chinese AI leaders will still buy NVIDIA chips when available, because performance matters and frontier training remains compute-hungry. But they will increasingly treat NVIDIA as a tactical accelerator, not the foundation of the strategy. Domestic chips will absorb more inference and specialized workloads. Open models will drive adoption, while agent platforms will become the usage layer. DeepSeek V4 is the proof point that this path is viable.