Chinese AI Labs Gain Ground Through Efficiency, Not Just Chips

Efficiency Over Hardware

U.S. intelligence agencies this week accused six Chinese AI companies—including DeepSeek and Moonshot—of using bulk subscriptions to American AI services to train their own models, allegedly extracting billions in value while understating development costs. China's foreign ministry rejected the claims as groundless.

But the allegations miss a more fundamental story about how Chinese labs closed the performance gap: they engineered their way around hardware limitations.

When U.S. export controls cut off access to Nvidia's most advanced chips, Chinese developers couldn't simply scale up compute power the way American labs did. Instead, they optimized the core attention mechanism that powers all large language models—the system that determines which parts of text matter most for understanding context.

Brendan Burke, semiconductors analyst at Futurum Group, explained that Chinese labs developed algorithms reducing the computational complexity of attention calculations by an order of magnitude. "They're able to summarize the most relevant tokens" more efficiently than U.S. models, which he described as "token hogs" designed for exploratory research rather than operational efficiency.

The approach delivered results. Stanford research earlier this year found Anthropic's leading model ahead of DeepSeek's by just 2.7%.

The Cost Advantage

The efficiency gains translate directly to enterprise economics. Ameya Kanitkar, cofounder of AI measurement platform Larridin, reported that Chinese models like GLM 5.2 and Kimi handle roughly 75% of engineering tasks at one-fifth the cost of U.S. alternatives.

"Frontier U.S. models still have an advantage on the most complex tasks, but Chinese open-weight models are becoming more than capable enough for the majority of everyday enterprise engineering work," Kanitkar noted.

That cost differential matters as AI spending grows. McKinsey found 20% of business leaders cite token costs as a constraint on AI adoption.

Enterprise Adoption Accelerates

U.S. companies are responding. DoorDash CEO Andy Fang called Moonshot's Kimi "cheaper" and "better quality" for code generation. AI coding startup Cursor incorporated Kimi into its Composer 2 agent. Airbnb and Siemens are testing Alibaba and DeepSeek models, with Airbnb CEO Brian Chesky praising Qwen as "fast and cheap."

Thomson Reuters built its Thomson-1 document-review model by adapting Alibaba's open-source Qwen, replacing work previously handled by Anthropic's Claude.

Hugging Face reported Chinese open-source models accounted for 41% of platform downloads last year—more than U.S. models. Ramp's AI spending index showed businesses paying for platforms with Chinese-developed models rose from 4.5% in January to 6.1% by July.

Open-source availability helps explain the adoption. DeepSeek's R1 reasoning model can be downloaded and run through U.S. cloud providers like AWS, letting companies fine-tune models without sending data to China-based servers.

Why it matters

The shift reveals that AI leadership isn't solely about raw computing power or the largest training budgets. Chinese labs demonstrated that algorithmic efficiency can compensate for hardware disadvantages—a lesson with implications as compute costs and energy consumption become limiting factors for the entire industry. For enterprises, the emergence of cost-effective alternatives creates pricing pressure on U.S. providers while expanding deployment options for budget-constrained AI initiatives.

Mike Finley, CTO of enterprise AI firm AnswerRocket, noted that U.S. companies still provide the "existence proof" Chinese labs innovate from: "The work they do would simply not be possible without the frontier labs blazing the trail."

These details were first reported by Fortune.