By Lina Irawan 
Earlier this week, U.S. officials levied serious accusations against six prominent Chinese AI companies, alleging they found a questionable shortcut to rapidly close the gap with their American counterparts. These companies, it was claimed, bypassed traditional, resource-intensive development by purchasing bulk subscriptions to American AI rivals’ services and subsequently training their own models on the outputs generated by these advanced systems. This method, if proven, represents a significant breach of implicit, if not explicit, terms of service, and raises profound questions about fair competition and intellectual property in the burgeoning AI landscape.
On Tuesday, a joint statement from key U.S. cybersecurity and intelligence agencies – the Federal Bureau of Investigation (FBI), the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) – specifically named DeepSeek, Moonshot, and four other Chinese entities. They asserted that these companies have, since 2024, "extracted capabilities worth billions" through this controversial method. The agencies further claimed that this alleged strategy allowed DeepSeek, for instance, to significantly understate its actual training cost, publicly reported at $5.6 million, by essentially leveraging the massive investments and proprietary data of U.S. labs without incurring equivalent development expenses. Such practices could distort market dynamics, undermine innovation incentives for those playing by the rules, and pose a direct challenge to the U.S.’s technological leadership.
In response, China’s foreign affairs ministry swiftly dismissed the accusations as "groundless." A spokesperson emphasized that China’s rapid advancements in artificial intelligence are "a result of high-level scientific and technological self-reliance," underscoring Beijing’s long-standing strategy to reduce dependence on foreign technology and foster indigenous innovation. This denial highlights the deep ideological chasm between the two nations regarding the origins and methods of technological progress, with the U.S. often suspecting state-sponsored intellectual property theft and China asserting its independent scientific prowess.
The allegations offer one compelling, albeit controversial, explanation for the remarkable progress observed in Chinese AI models. Recent analyses suggest that Chinese AI models are performing neck-and-neck with their U.S. counterparts. A comprehensive report from Stanford University, published earlier this year, notably placed Anthropic’s top model ahead of DeepSeek’s by a mere 2.7%. This razor-thin margin surprised many in the industry, given the perceived lead of American companies in foundational AI research and vast computational resources. Such parity, if achieved through illicit means, would represent a significant intelligence and economic coup for China.
However, many analysts argue that there’s a more nuanced, and perhaps more legitimate, explanation for China’s ascent in AI: a distinct advantage developed by Chinese labs through necessity – the ability to squeeze significantly more value from limited computational resources. This perspective suggests that while accusations of intellectual property infringement demand scrutiny, they may not fully account for the ingenuity observed in Chinese AI development.
More Value for Fewer Tokens: The Ingenuity of Constraint
The core of this alternative explanation lies in a sophisticated technical refinement perfected by Chinese labs, specifically concerning the "attention" mechanism. This mechanism, first introduced by Google researchers in a landmark 2017 paper titled "Attention Is All You Need," is the foundational innovation underlying every large language model (LLM) today. At its heart, attention allows an AI model to weigh the importance of different words or tokens in an input sequence relative to each other. This is crucial for understanding context, nuance, and long-range dependencies within text, enabling the model to generate coherent and contextually relevant responses.
However, the computational cost of the attention mechanism scales quadratically with the length of the input sequence, or "context window." As models process longer texts, the calculations required to determine relationships between all possible token pairs become exponentially more expensive, demanding immense processing power.
Brendan Burke, a semiconductors and supply chain analyst at the tech research firm Futurum Group, explained that Chinese labs engineered a shortcut to make this fundamental mechanism significantly cheaper and more efficient. "Chinese labs found algorithms that reduce the complexity of those calculations by an order of magnitude, and then achieve better results because they’re able to summarize the most relevant tokens," Burke told Fortune. This breakthrough allowed them to process information more efficiently, requiring less raw computational power to achieve comparable or even superior contextual understanding.
This innovative approach was largely born out of necessity. The U.S. government, through a series of stringent export controls, severely restricted China’s access to Nvidia’s most advanced AI chips, such as the A100 and H100 GPUs, beginning in 2022. These chips are considered the gold standard for training and deploying large AI models due to their unparalleled processing capabilities. This forced China to pivot towards domestic alternatives, notably those developed by Huawei, which, while capable, often lagged behind Nvidia’s cutting-edge performance. This limitation meant Chinese AI developers had significantly less access to the highest-performing compute available globally.
The disparity in computational resources is stark. According to a White House report, the U.S. commands approximately 74% of the world’s available high-end compute capacity. Furthermore, American hyperscalers – major cloud providers like Amazon, Microsoft, and Google – are pouring hundreds of billions of dollars annually into expanding their AI infrastructure through vast data centers, ensuring a continuous supply of top-tier compute for their research and development. This allows U.S. labs to scale up their models and experiments with an abundance of processing power.
"Because they had less compute to work with, they found that computationally efficient method instead of just throwing more compute at an inefficient technique, as U.S. labs initially did," Burke observed about China. This highlights a fundamental divergence in development philosophy: while U.S. frontier labs, flush with resources, could afford to be "token hogs" – designing AI systems that were highly exploratory and perhaps less optimized for efficiency in their initial stages – Chinese labs were compelled to innovate within tighter constraints. Their focus shifted to algorithmic breakthroughs that maximized utility per unit of compute, a strategic advantage in a resource-constrained environment.
The tangible trade-off between these two approaches is quantifiable. Ameya Kanitkar, cofounder of the AI measurement platform Larridin, revealed to Fortune that in the enterprise workflows tracked by Larridin, Chinese models such as GLM 5.2 and Kimi 2.6 and 2.7 are capable of handling approximately 75% of typical engineering tasks "reasonably well," but at roughly a fifth of the cost of their U.S. counterparts. While U.S. frontier models still retain an advantage on the "most complex tasks," Kanitkar noted, "Chinese open-weight models are becoming more than capable enough for the majority of everyday enterprise engineering work."
This cost differential is becoming increasingly critical as AI spending consumes a larger share of corporate budgets. A recent McKinsey survey indicated that 20% of business leaders cited AI-related costs, particularly the expense of buying tokens (the fundamental units of text processed by AI models), as a significant constraint on their broader AI adoption. The ability to achieve similar outcomes at a fraction of the cost presents a compelling economic argument for businesses globally.
U.S. Enterprises Warming Up to Chinese Models
Beyond mere cost-efficiency, Chinese AI models are gaining traction among U.S. enterprises due to their increasing accessibility and flexibility, particularly through open-source initiatives. DeepSeek’s R1 reasoning model, for instance, was made available for download via popular platforms like Hugging Face. This open-source strategy empowers companies to download, run, and adapt versions of the model themselves, fine-tuning them to meet specific needs without being solely dependent on a closed, proprietary model from a single vendor. Furthermore, the option to run these models through U.S.-based cloud providers like Amazon Web Services (AWS Bedrock) mitigates concerns some businesses might have had about sending sensitive data to China-based companies.
This blend of flexibility, cost-effectiveness, and localized deployment options has significantly warmed U.S. businesses to DeepSeek and other Chinese models. Data from Hugging Face itself underscores this trend, reporting that Chinese open-source models accounted for an impressive 41% of total downloads last year, surpassing the share of U.S.-developed models. This indicates a growing appetite among developers and enterprises for these alternatives.
High-profile endorsements further validate this shift. Andy Fang, CEO of DoorDash, publicly stated that using Moonshot AI’s Kimi model resulted in outcomes that were both "cheaper" and of "better quality" without compromising the integrity or performance of code generation. Similarly, AI coding startup Cursor adopted Kimi to help build its Composer 2 coding agent, leveraging its efficiency. Major corporations like Airbnb and Siemens are also actively experimenting with models from Alibaba and DeepSeek. Airbnb CEO Brian Chesky, commenting on Alibaba’s Qwen model, described it as "fast and cheap," signaling a pragmatic approach to AI adoption driven by performance and economic factors.
The utility of Chinese models is even extending to more specialized and demanding applications. Thomson Reuters, a global information services giant, revealed it built an in-house model called Thomson-1 by adapting Alibaba’s open-source Qwen model. This custom solution is now handling complex document-review work, a task previously performed by Anthropic’s Claude, demonstrating a direct replacement of a leading U.S. model with a Chinese open-source alternative.
This enterprise shift is becoming increasingly visible in broader market data. Ramp’s AI index, which tracks corporate spending on AI platforms, showed a notable increase in the share of businesses paying for platforms with access to open-source and Chinese-developed models. This figure rose from 4.5% in January to 6.1% of total AI-spending businesses in July, indicating a clear upward trend in adoption.
Despite this growing acceptance and undeniable progress, U.S. models still hold a crucial edge. Industry experts, including Mike Finley, Chief Technology Officer of enterprise AI analytics firm AnswerRocket, contend that U.S. AI companies remain "months ahead in performance" on the absolute frontier of AI capabilities. Finley argues that the output of U.S. AI companies often serves as the "existence proof" for Chinese labs, providing the benchmark and conceptual framework upon which they innovate. "The work they do would simply not be possible without the frontier labs blazing the trail," Finley stated, suggesting a symbiotic, albeit competitive, relationship where foundational breakthroughs in the U.S. inform and enable subsequent efficiency-driven innovations in China.
The escalating AI race between the U.S. and China is therefore a multifaceted phenomenon. It encompasses serious allegations of intellectual property infringement and unfair competition, alongside undeniable advancements driven by strategic innovation under constraint. While U.S. agencies raise alarms about alleged shortcuts and the geopolitical implications of China’s rapid rise, the market is simultaneously witnessing a pragmatic embrace of cost-effective and flexible Chinese AI solutions by American enterprises. This complex interplay of competition, accusation, and technological ingenuity defines the current landscape, pushing the boundaries of what’s possible in artificial intelligence and shaping the future of global technological leadership.
