The release of Kimi K3, the latest large language model from Chinese AI startup Moonshot, has reignited a question Washington may not yet have a good answer for: what happens when the AI competition shifts from who builds the best chips to who deploys the cheapest, most widely adopted models?
Writing in Storm Media, Simon Shen — an associate professor at National Sun Yat-sen University's Taiwan-Hong Kong International Research Center and a specialist in international relations — argues that the defining logic of the US-China AI rivalry is quietly but decisively changing.
Benchmarks That Unsettled Silicon Valley
For most of the past two years, the AI competition between the United States and China played out primarily in hardware. Washington restricted NVIDIA's highest-end GPU exports, calculating that constraining China's access to advanced chips would slow its ability to train frontier AI models. Beijing, in turn, pushed to build domestic chip supply chains and find ways to extract more efficiency from the hardware it could access. The underlying assumption — that whoever accumulated the most compute power would dominate the AI era — seemed self-evident.
Kimi K3 arrived as a challenge to that assumption. Based on public benchmark results, the model performs competitively with, and in select tests surpasses, OpenAI's GPT-5.5 and Anthropic's latest models — while reportedly running at a fraction of the inference cost. Developers began circulating the phrase "Kimi Moment" to describe the release, drawing comparisons to DeepSeek's surprise debut earlier this year that briefly rattled U.S. AI markets and sent investors scrambling to reassess the competitive landscape.
The comparison should be taken seriously — but not uncritically. An analysis document circulated online raised questions about whether portions of Kimi K3's benchmark results may reflect optimization for specific test conditions rather than broad generalized capability. The document's origins and methodology have not been independently verified. Whether the model's scores accurately represent its real-world performance remains a live debate in the developer community.
The Economics That May Matter More Than the Scores
Shen's more durable argument concerns not benchmark rankings but economic logic. Even if American models maintain a 10% performance edge, he writes, Chinese models capable of delivering comparable results at 10% — or even 1% — of the cost reshape the competitive calculus entirely. Global enterprises choosing between AI providers are not running benchmark suites. They are making budget decisions.
This is what made DeepSeek's appearance genuinely disruptive, and what gives the Kimi K3 conversation its staying power regardless of how the benchmark dispute is resolved. The shift Shen describes echoes the arc of China's electric vehicle industry: Chinese automakers did not need to build the world's most technologically sophisticated vehicle to reshape global automotive markets. They needed to build one that was good enough, and far cheaper. The price disruption alone was sufficient to threaten established industrial structures in Europe, Japan, and the United States.
AI models, Shen argues, may be following the same curve.
US Lobbying Runs Into the Open-Weight Problem
Against this backdrop, the Wall Street Journal has reported that OpenAI, Anthropic, and other major American AI companies are lobbying the U.S. government to impose stricter restrictions on Chinese AI models, citing national security grounds. The concern has substance. Large language models are increasingly embedded in critical infrastructure — government operations, military logistics, energy grids, financial systems, healthcare — and if a model's training pipeline, update mechanisms, and underlying data are controlled by a foreign state, the risks are genuine. Beijing has reached an equivalent conclusion from its own side, moving to restrict certain AI and model-related technology exports in recent months — a signal that it views AI not as a commercial product but as a strategic asset.
But restricting Chinese AI models may prove far more difficult than restricting Chinese chips. Blocking access to a cloud-based service — a website, an API endpoint, a government procurement category — is operationally manageable. Restricting an open-weight model is a different problem entirely. Chinese AI companies have been aggressively releasing open-weight versions in recent months, precisely to accelerate global developer adoption. Once model weights are publicly available, any developer, researcher, or enterprise anywhere in the world can download and run them locally — on a private GPU cluster, an in-house server, a company's internal infrastructure — with no connection to the originating company required, and no central license to enforce.
Any American attempt to prohibit downloads of Chinese open-weight models would face steep enforcement challenges and a politically charged collision with norms around academic freedom, open-source software, and the free flow of scientific knowledge.
A Contest Over Who Becomes Global Infrastructure
The deepest argument Shen makes is about what this competition is ultimately for. The new front in the AI Cold War is not about controlling the most chips. It is about whose model becomes the default infrastructure — for governments, for enterprises, for the global developer community.
Whoever achieves that kind of embedded adoption wins something more valuable than a supply chain. They acquire ongoing insight into how their model is used, by whom, and in what context. They shape the development tools, the APIs, the fine-tuning pipelines, that an entire generation of technologists builds on top of. In Shen's framing, this is a race for the algorithmic layer of the international order — and it may determine more about the shape of global power in the coming decades than any hardware embargo. （Related： Allora Labs Launches Forge to Let AI Models Compete, Improve, and Earn on Real-World Predictions ｜ Latest ）
