The dynamics of competition in artificial intelligence are undergoing a significant shift. Perhaps years from now, people will recognize that the turning point was triggered by what initially seemed like a minor incident.
In recent days, a dramatic exposure has unfolded within the global tech community, resembling an "AI authenticity scandal." Cursor, a highly prominent AI programming startup from Silicon Valley, publicly launched its purportedly "self-developed" Composer 2 model. However, developers worldwide quickly identified a critical issue within the code—the model ID clearly indicated it was based on Kimi K2.5, a large language model developed in China. The fact that a Silicon Valley star company, seeking a valuation of $500 billion, had quietly built its core competitive advantage on a foundational model from China—whose developer is valued at only $18 billion—struck many as astonishing. The AI community reacted with widespread surprise.
Interestingly, just four days prior, the Chinese company behind this model had already made a notable impression on the Silicon Valley tech circle. To understand the significance, one must look at the underlying technology. In recent years, "large models" have brought AI into everyday life, allowing ordinary users to interact directly with AI for tasks like work assistance, video creation, and brainstorming. This capability stems from logical reasoning, which relies on a core component called the Transformer. At the foundation of this component lies a key logic known as "residual connections."
Residual connections are fundamental to modern large models, enabling stable training of deep networks and effective information transfer—whether for long-text comprehension, complex reasoning, or multi-turn dialogues. However, this approach has limitations. As models grow deeper, effective information within residuals tends to dilute, often burying critical details from earlier layers. This has been a major obstacle to enhancing large model capabilities.
On March 16, Moonshot AI, the Chinese company behind Kimi, released a technical report titled "Attention Residual," proposing the first major reconstruction of the residual connection component in the Transformer architecture since its inception a decade ago. Essentially, Kimi introduced a new conceptual pathway for the foundational structure of large models, opening up fresh avenues for evolution. This achievement caught the attention of Elon Musk, who described it as "impressive." Soon after, developers noticed something familiar: the Chinese model's name appeared in the underlying code of Silicon Valley's supposedly cutting-edge, self-developed Composer 2 model.
Musk confirmed the discovery and shared it on social media, settling the matter decisively. Amid the uproar, Cursor felt compelled to respond. Its co-founder, Aman Sanger, issued a statement conceding: "We systematically evaluated numerous open-source base models, and Kimi K2.5 proved to be the world's strongest." This admission unmistakably pulled back a "curtain of pretense," revealing that the foundation of Silicon Valley's proudly touted technological innovation is being quietly replaced by Chinese-made solutions.
Silicon Valley is more than a geographic location; it represents a global hub for top AI talent and a cornerstone of U.S. competitiveness in AI. Its reputation has been built over decades by generations of elites. Two constants have defined Silicon Valley: first, it must consistently lead in innovation, maintaining a technological edge that directs global progress; second, it must set standards—not only technical benchmarks but also industry paradigms, such as the commitment to "open-source collaboration" in AI.
However, Musk's exposure has shaken these foundational beliefs. "Can we still trust Silicon Valley?" some are asking on social media. Once doubt takes root, change becomes inevitable. Increasingly, observers are looking beyond Silicon Valley narratives to acknowledge China's advancements. This shift in attention reveals that beyond startups abandoning expensive proprietary models for open-source Chinese alternatives, major players like Airbnb and German industrial giant Siemens are already openly utilizing Chinese models.
The market is voting with its feet, and capital is following suit. Take Kimi, for example: its valuation quadrupled to $18 billion within three months, outpacing the growth rates of ByteDance and PDD during comparable periods. As a co-founder of Hugging Face, the world's largest AI community, stated plainly: "Chinese open source has become the strongest driver shaping the global AI technology stack."
Initially, Chinese AI firms were heavily influenced by Silicon Valley, maintaining a learner's posture toward its giants. But now, the gap on the track is narrowing. Followers are accelerating, closing in on the leaders. This is not mere speculation; beyond current developments, there is logical inevitability. Why did Kimi become the "world's strongest open-source base model" chosen by competitors, even at the risk of covert usage? While the incident may appear accidental, it reflects an underlying trend.
Before receiving recognition from Silicon Valley giants, Cloudflare—an $80 billion global internet infrastructure leader—announced deploying Kimi K2.5 in its production environment. The result: a 77% reduction in operational costs for its internal security audit agent, which processes 7 billion tokens daily, alongside notable efficiency gains. This improvement stems partly from technical advantages like "trillion-scale MoE architecture" and "native multimodal understanding," but more importantly, from the model's optimal balance of performance and cost.
Similar to "Made in China," Chinese large models possess distinct strengths: rapid iteration and lower costs. These advantages ultimately point to one critical factor: China's "infrastructure edge" in the AI era is becoming a decisive element in the competition.
Recently, another headline emerged: "Token出海" (Token going global), which has attracted significant attention. Reports indicate that over the past year, overseas paid usage of Chinese large models has exploded, with the trend accelerating in recent months. Two key terms are important here: first, "Token," recently assigned the Chinese term "元," refers to the basic unit of computational consumption in large model operations; second, "overseas paid usage" denotes individuals and enterprises outside China paying to use tokens generated by Chinese models. In simple terms, more people worldwide are spending money to access Chinese large models.
Why? On one hand, Chinese models are becoming more capable; on the other, cost plays a major role. China's advantage in electricity is translating into a computing power edge, allowing token pricing to remain significantly lower than international rates. This means not only that China will supply affordable computing power globally, but also that its AI industry now enjoys a "first-mover advantage" by being closer to the source.
Consider this: as Chinese large models serve global users with more cost-effective tokens, their growth will inevitably outpace that of Silicon Valley counterparts. This eastward shift is irreversible and likely to accelerate. The tipping point may not be far off, as signs are already visible: overseas usage of top Chinese models is multiplying; domestic models are narrowing the gap with leading international counterparts on key metrics; and China has produced innovators like DeepSeek and Kimi, who dare to challenge conventions and rethink foundational logic.
Notably, the aftermath of the Kimi incident has been remarkably calm. Instead of prohibiting Silicon Valley firms from using its product, Moonshot AI issued a gracious congratulatory message: "We are proud to see Kimi K2.5 provide the foundation." Such poised yet firm rhetoric is unlikely to comfort Silicon Valley, as it fundamentally challenges the valley's long-held authority to define future technology.
An analogy illustrates the shift: Silicon Valley has long been crafting luxury sports cars, setting industry standards. But if one day a new high-performance car is found to be powered by a Chinese engine, it signals that the time for China to redefine premium vehicles is near—especially as China builds highways worldwide. Silicon Valley giants may hesitate to acknowledge the rise of Chinese AI technology, because admitting that Chinese labs are advancing the industry more efficiently and openly, with superior and cheaper infrastructure, would undermine Silicon Valley's sacred status. And if AI enterprises worldwide turn their eyes eastward, how effective will the narrative—"We need $1 trillion to build AGI"—remain?
What Musk has revealed is not merely a cover-up but a fateful trajectory for Silicon Valley: global AI development is no longer confined to the "Silicon Valley invents, the world applies" model. A new path has emerged—"China innovates, the world benefits." Through companies like DeepSeek and Kimi, China is demonstrating that new quality productive forces represent not only more efficient tools but also a global infrastructure capability that can define costs, set rules, and export ecosystems. This includes the power to rewrite architectural standards, define cost efficiency and token pricing, govern open-source protocols, and champion a green, inclusive AI development path.
The "rise of Chinese AI" is not merely about GDP figures; it is about the systemic capacity to define the foundation of the next era. As more global intelligence operates on platforms where efficiency, cost, and rules are shaped by China, it becomes increasingly clear who will mold the future.
Comments