China's AI Labs Disrupt Market with Cheap, Competitive Models

The whispers began in mid-July. Online observers tracking the AI community noticed signs that Moonshot AI, the Beijing-based startup behind the Kimi large language model, was preparing a major release. On July 16, the lab delivered: Kimi K3, which it described as the largest open-source model ever released.

Moonshot's timing and positioning were strategic. The company claimed K3 could perform close to the level of Anthropic's Fable 5—widely regarded as the most powerful publicly available model—at a fraction of the cost. Moonshot's official benchmarks ranked K3 among the top three AI models globally; one independent evaluation from Arena.AI went further, pegging K3 as the single best model available, surpassing even Anthropic. The founder, Yang Zhilin, a 34-year-old Tsinghua and Carnegie Mellon graduate, had named the company after Pink Floyd's album "The Dark Side of the Moon."

The market reacted sharply. Nvidia lost almost $600 billion（約96兆円） in value and briefly ceded its position as the world's most valuable company to Apple. The Philadelphia Semiconductor Index fell 1.6%. The reaction reflected a deeper anxiety: the release suggested that U.S. dominance in AI could not be taken for granted, and that Chinese firms could now compete on both capability and cost—undercutting expectations even from observers like Anthropic CEO Dario Amodei, who had not anticipated a Chinese model approaching U.S. levels for at least another six months. Tesla CEO Elon Musk had predicted first quarter of next year; Moonshot compressed that timeline by months.

The K3 launch was not an isolated event but the latest move in a rapid series of Chinese AI breakthroughs. DeepSeek, a Hangzhou-based lab attached to a Chinese hedge fund, had shocked the industry in early 2025 by releasing V3 and R1 models that matched U.S. performance. DeepSeek claimed to have achieved this with a tiny budget by engineering efficiency—using smart programming and mathematical tricks to squeeze maximum performance from second-tier hardware. In June, Z.ai released GLM-5.2, which proved particularly strong at coding and creative design; its newly listed stock surged over 1,100% through mid-July and briefly exceeded 1 trillion Hong Kong dollars ($127.6 billion（約20兆円）) in market capitalization. Meituan, best known as a food-delivery platform, entered the arena with LongCat-2.0, encompassing as much data as DeepSeek's V4 and performing at levels matching OpenAI and Anthropic releases from February. Most remarkably, Meituan claimed to have trained the system entirely on Chinese-made processors rather than U.S. chips.

The cost differential was stark. One million output tokens—roughly 750,000 words—cost $50 on Anthropic's Fable. The same volume from DeepSeek-V4-Pro cost about $0.87; Z.ai's GLM-5.2 cost $4.40; Kimi K3 cost $15. On OpenRouter, a marketplace where developers access different models through a single interface, Chinese models now dominated the rankings: at one point in mid-July, six of the top 10 models came from Chinese companies, and all of the top five. In one week in July, Chinese AI models accounted for 57% of tokens used by U.S. firms on the platform.

Adoption was moving beyond developer circles into mainstream operations. Airbnb CEO Brian Chesky disclosed last year that the company used Alibaba's Qwen for customer service. Cursor, an AI coding startup, said Moonshot AI's Kimi provided the foundation for Composer 2, its coding model. Coinbase CEO Brian Armstrong wrote in a June social media post that the crypto platform had halved its AI spending by pushing employees to use Kimi and Z.ai's GLM models. DoorDash's chief technology officer Andy Fang explained that the company delegated "lower-level work" to Kimi, achieving "better quality [at] cheaper cost."

Why could Chinese firms offer such advantages? Several factors converged. Power costs less in China, in part because the country has invested heavily in power generation and transmission infrastructure, making it easier to add data center capacity. By contrast, new U.S. data centers often face political resistance over grid strain and water consumption. Chinese companies were willing to sacrifice profit margins to capture market share and establish their models as de facto standards. U.S. export controls, paradoxically, may have accelerated this outcome by forcing Chinese labs to optimize for efficiency on constrained hardware. Labs, constrained on compute, capital, and talent, learned to deploy resources cautiously. Recent Chinese models also became compatible with cheaper, locally made processors: for the cost of one Nvidia chip, a Chinese company could buy 10 local chips from Huawei or other chipmakers. And Chinese firms embraced open-source licensing, releasing models for free under permissive terms that let users download, fine-tune, and run them locally at no licensing cost—reducing the effective price to just GPUs and energy.

U.S. policymakers began to reckon with unintended consequences. Congress probed companies like Airbnb and Cursor on their use of Chinese models, citing security concerns and competitive risk to U.S. developers. Airbnb responded that it used only "a limited number" of Chinese models, all open-source and run through "approved U.S.-based service providers." Cursor did not respond to requests for comment. Lawmakers also explored whether the U.S. should do more to support open-source models of its own.

Chinese startups, sensing opportunity, turned U.S. policy anxiety into leverage. Z.ai announced GLM-5.2 just days after U.S. officials briefly cut off access to Anthropic's Fable and Mythos models for some foreign users, writing: "Frontier intelligence should not belong to only a few people, nor be subject to withdrawal by a handful of rules at any moment." Chinese firms began courting governments seeking sovereign AI—models they could run on domestic hardware with local control over data and updates. As U.S. policy grew more protectionist and unpredictable, demand grew. Countries like Singapore and India faced growing uncertainty: one day, access to the latest U.S. model was permitted; the next, foreign citizens were blocked. Beijing capitalized on this anxiety. At an AI conference in July, President Xi Jinping pledged to "uphold openness and win-win cooperation" in AI, saying the technology "should not be a solo performance by any one country, but a symphony of global cooperation."

The release of Kimi K3 in mid-July marked a watershed moment in the global AI race. For years, U.S. policymakers had assumed export controls on advanced chips would preserve America's dominance in AI by limiting China's access to the hardware needed to train frontier models. Yet Chinese developers proved this assumption incomplete. DeepSeek's V3 and R1 models, released in early 2025, already demonstrated that efficiency engineering could compensate for hardware constraints; Moonshot's K3, Z.ai's GLM-5.2, and Meituan's LongCat-2.0 extended this lesson. These models not only matched U.S. performance but did so at a cost structure that made them attractive to U.S. companies themselves—a dynamic that rattled markets and forced a reassessment of the tech industry's competitive landscape.

The cost advantage stems from multiple sources. Power is cheaper in China, where decades of investment in generation and transmission infrastructure make it easier to add capacity. Chinese firms are also willing to sacrifice profit margins to capture market share and establish de facto standards, treating price competition as a long-term strategic investment. Open-source licensing removes the need for cloud-based inference, shifting costs to end users' own hardware and electricity. And the hardware constraint itself, ironically, may have sharpened efficiency: forced to do more with less capable chips, Chinese labs discovered techniques to extract maximum performance per unit of compute.

U.S. policymakers are now grappling with the unintended consequences of their approach. Congress has begun probing companies like Airbnb and Cursor over their use of Chinese models, citing both security and competitive concerns. Yet the restrictions that prompted this scrutiny—including recent brief cutoffs of access to Anthropic models—have inadvertently strengthened China's hand. Countries uncertain about the stability of U.S. AI policy are turning to Chinese models as a hedge, and Beijing has framed this shift as a geopolitical win, with President Xi Jinping pledging openness and cooperation in AI at a July conference.

No comments yet. Be the first to share your thoughts!