The recently concluded summit between Chinese President Xi Jinping and U.S. President Donald Trump in Washington D.C. encompassed the standard diplomatic communiqués, though there was sparse regard for artificial intelligence (AI). No binding agreement on AI was signed, although there was some improvement in communication, as a bilateral incident hotline will be created to address AI-related emergencies.
Similarly, a U.S.-China dialogue will be established as a formal consultation mechanism, though Washington uses the term "Super Intelligence,” while Beijing prefers "Artificial Intelligence.” In addition, both sides touched upon basic guardrails and the prevention of automated command-and-control systems in nuclear arsenals, a highly important step in view of recent revelations about how advanced AI systems can behave.
Trump and Xi, however, failed to establish any binding framework that would limit the capabilities of frontier AI models. In fact, both leaders proved reluctant to restrict or negotiate any slow-down to advanced model development, continuing with what is a structural, systemic race to control the digital commanding heights of our time. Both nations are cognizant that the winner of this race could dictate global economic productivity, scientific advancement, and the very nature of sovereign security and power. The race is seen in geopolitical terms as a winner-take-all competition.
But this all-out competition has in recent months and weeks emerged under a new light. Ever more capable models, especially once they approach recursive self-improvement, could autonomously rewrite their own code and train their successor architectures. At that point, so the reasoning goes, the speed of technological evolution decouples from human control.
To understand how the current dynamic became so unpredictable, one must scroll back to the foundational architecture of contemporary AI. The field's trajectory was radically accelerated with the public release of ChatGPT in late 2022. This model demonstrated the emergent capabilities of Large Language Models (LLMs). But already at that early time certain seemingly persistent and structural flaws emerged: hallucinations.
Early LLMs did not “know” facts in a human sense, but rather statistically predicted the next token, frequently inventing plausible-sounding falsehoods with absolute confidence. While western labs raced to patch hallucinations through reinforcement learning from human feedback, Chinese firms took different architectural and strategic routes.
For example, Moonshot's models pioneered innovative solutions, focusing heavily on massive context windows and long-text processing. DeepSeek-R1, launched in late January 2025, further triggered massive cost-disruption with its considerable reasoning capabilities at low cost.
With this, the global AI landscape fractured into a foundational competition between open-weight and closed models. Closed models, championed by pioneers like OpenAI and Anthropic, are kept behind proprietary Application Programming Interfaces (APIs). The creators retain absolute control over the weights – the numerical parameters that dictate how the neural network processes information. Conversely, the open-weight paradigm involves releasing the model parameters directly to the public, allowing researchers, startups, and sovereign entities to download, modify and run the software locally.
Both approaches possess distinct advantages and disadvantages. Open-weight models have increasingly democratized computational power, as they are cheaper and more flexible. Chinese open-weight models, for example, have surged past U.S. counterparts on Hugging Face, capturing roughly 41 percent of platform supply and dominating developer adoption and download metrics. Especially smaller enterprises and startups increasingly choose Chinese open-weight options over expensive U.S. closed-source APIs due to lower inference costs, higher performance, and no guardrail friction – models lack restrictive proprietary guardrails, allowing full customization and local inference, which is much cheaper.
Besides cost, the primary advantage of open-weight models is their capacity for decentralized innovation. By giving the global developer community access to the underlying weights, optimization occurs in an open ecosystem. However, there is a substantial downside: a loss of control. Once an open-weight model is released, its guardrails can be stripped away entirely. Anyone can systematically bypass safety filters, repurposing the model.
Closed models mitigate these immediate deployment risks through strict API monitoring and centralized content filtering, but they create centralized single points of failure, encompass limited external auditing, and could impose steep monopoly costs on the broader digital economy if one model emerges as dominant.
For the past three years, the AI race between the United States and China increasingly looked like a competition between these two strategic tech approaches. The United States possessed a commanding lead in the closed-model ecosystem, driven by massive venture capital inflows, proprietary cloud infrastructure and frontier labs pushing the absolute limits of AI.
Chinese AI companies, facing stringent Western export controls on cutting-edge semiconductor hardware, strategically leaned into the open-weight ecosystem, bypassing the prohibitive costs of foundational training. They effectively closed the capability gap by operating as fast-followers.
What looked like a neat, strategic competition has now been viciously disrupted. Recent revelations regarding the inherent dangers and unexpected capabilities of the most advanced models have forced a reassessment of the entire trajectory of AI development. Models such as Anthropic's Claude and OpenAI's GPT-6 Astra have transcended simple pattern matching to display multi-step agentic reasoning, cross-modal tool usage, and even early signs of situational awareness. It is clear that these systems can now autonomously plan, execute code and manipulate environments.
This shift brings a whole new set of structural and existential questions to the forefront. First and foremost, should development slow down? From a pure safety perspective, the answer appears to be a resounding yes, as humanity lacks verified methods for aligning agentic systems that operate beyond human cognitive speeds. Yet, implementing such a pause seems virtually impossible given the existing structural rivalries. The friction is not merely among nations, but also among innovation ecosystems, corporations and the foundational philosophies of AI.
Corporate and philosophical rifts furthermore amplify the most critical technical question: what happens when a highly capable model approaches recursive self-improvement? Would other models be able to catch up? AI is after all a highly diffusion-prone technology. A breakthrough achieved in an academic lab in California can be adapted by engineers in Shenzhen within days, and vice versa.
Or could a permanent, unassailable monopoly emerge, locking in a decisive technological lead before competitors can even parse the paradigm shift? As scary as this scenario sounds, if open-weight architectures close the capability gap, a parallel nightmare scenario might emerge: the democratization of catastrophic risk.
Open-weight models allow users to directly manipulate the underlying alignments, so an end-user could systematically alter or strip away the model's core parameters. An aligned, safe frontier model could be downloaded, inverted via targeted fine-tuning, and transformed into an inherently malicious tool.
Put bluntly, the U.S.-China AI race is no longer a predictable game of geopolitical chess or a standard corporate competition over market leadership. We are entering the uncharted territory of artificial cognition. The Trump-Xi summit proved that sovereign states are still using an outdated diplomatic playbook to address a technology that defies traditional borders and governance structures. As the true, unpredictable nature of frontier AI models reveals itself, the illusion of control is evaporating. The global community is left grappling with a profound paradox: the very mechanisms used to win the race – relentless scale, open-source proliferation and agentic autonomy – are precisely the mechanisms that could render the race entirely unmanageable, creating a panoply of absolutely mindboggling apocalyptic risks.
