When Airbnb CEO Brian Chesky told Bloomberg in October that his company relies heavily on Alibaba's Qwen model for AI-powered customer service, calling it "very good, fast, and cheap," he offered a rare glimpse into a trend that's quietly reshaping enterprise AI adoption. While Silicon Valley giants battle over who can build the most powerful—and expensive—language models, a different approach is gaining traction in boardrooms worldwide.
The revelation is particularly striking given Chesky's close friendship with OpenAI CEO Sam Altman. Yet Airbnb, which uses 13 different AI models including OpenAI's, depends most heavily on Qwen. "We use OpenAI's latest models," Chesky explained, "but we typically don't use them that much in production because there are faster and cheaper models."
The Cost Equation
The economics driving this shift are compelling. DeepSeek's R1 model reportedly cost between $5.6 million and $10 million to train, according to claims made by the company—figures that remain unverified but are nonetheless significantly lower than the estimated $100 million to $300 million range for training models like GPT-4, based on industry estimates from analysts and OpenAI's own internal projections. Even accounting for research, experimentation, and infrastructure costs not included in DeepSeek's headline figure, the efficiency gap appears substantial.
Alibaba's Qwen 2.5-Max costs approximately $0.38 per million tokens, making it significantly cheaper than competing Western models while matching or exceeding their performance on several benchmarks. For enterprises processing billions of tokens monthly, this translates to real cost savings that can determine profitability.
Western Institutions Making the Switch
The adoption extends well beyond Airbnb. In 2025, several major Western institutions have begun testing or deploying Chinese AI systems:
Financial institutions: HSBC and Standard Chartered, two of the largest banks operating globally, have begun testing DeepSeek models for internal use.
Energy sector: Saudi Aramco, the world's most valuable oil company, recently installed DeepSeek in its main data center.
Cloud providers navigating both sides: Amazon Web Services, Microsoft Azure, and Google Cloud all now offer ways for their customers to deploy DeepSeek—despite White House warnings about potential security risks. The companies are effectively enabling customer choice rather than blocking access to competitive technologies.
Academic and government adoption: South Africa's University of the Witwatersrand selected DeepSeek for a research pilot project, citing its offline capability and open-source flexibility. Japan's Ministry of Economy chose Alibaba's Qwen model over U.S. alternatives for certain applications. Developer platform Latenode reports that approximately 20% of its global users now prefer DeepSeek for building AI tools.
Within China, the momentum is even more pronounced. According to Frost & Sullivan research published in 2025, more than 80% of Chinese enterprises plan to adopt open-source LLMs, with Alibaba's Tongyi (the Chinese name for Qwen) commanding a 17.7% market share.
Three Reasons Chinese Models Compete
The adoption isn't driven by marketing but by tangible advantages that address enterprise pain points:
1. Economic efficiency: Beyond headline pricing, Chinese models often use Mixture-of-Experts (MoE) architectures that reduce computational costs by up to 30% compared to traditional dense models. This allows for competitive performance with lower operational overhead.
2. Open-source flexibility: Unlike OpenAI's proprietary approach, Chinese developers have embraced transparency. Alibaba has open-sourced over 300 Qwen models under permissive licenses, leading to more than 170,000 derivative models created by developers worldwide. DeepSeek's MIT license allows companies to download, modify, and deploy without restrictive commercial terms—a significant advantage for enterprises concerned about vendor lock-in or data sovereignty.
3. Competitive performance: On key industry benchmarks, Chinese models now rival Western counterparts. Qwen 2.5-Max scores 89.4 on Arena-Hard—a rigorous comparative benchmark of model reasoning maintained by the LMSys organization—compared to Claude 3.5 Sonnet's 85.2. DeepSeek-R1 achieved 97.3% on the MATH-500 benchmark, a standard test of mathematical reasoning capability. While these benchmarks are often self-reported and may not capture all aspects of real-world performance, they suggest the quality gap has narrowed considerably.
The Chinese Enterprise Wave
Following DeepSeek's January 2025 breakthrough, Chinese enterprises moved with remarkable speed. CNBC reported in February that eight automakers including BYD, at least nine financial securities companies, three state-owned telecommunications operators, and smartphone brand Honor all rushed to integrate DeepSeek within a single week.
"This is quite unprecedented," Wei Sun, principal analyst at Counterpoint Research, told CNBC, pointing to the rate of adoption, scale of business integration, and breadth of industries covered.
Cloud computing giants Alibaba, Huawei, Tencent, and Baidu now all offer streamlined access to DeepSeek's models. China's national supercomputing network announced subsidized access for eligible companies—a state-backed infrastructure push that some analysts compare to the planned U.S. Stargate network for frontier AI research.
The Ecosystem Advantage
What makes this shift potentially sustainable is the ecosystem developing around Chinese models. According to Alibaba, over 90,000 enterprises across sectors from e-commerce to insurance have adopted Qwen. FAW Group, one of China's largest automakers, built its internal AI agent OpenMind using Qwen. By January 2025, more than 290,000 customers across robotics, healthcare, education, finance, and automotive sectors had adopted Qwen via Model Studio, Alibaba's development platform.
This creates a reinforcing cycle: more users generate more feedback, leading to better models, attracting more developers—the same dynamic that helped establish AWS's early cloud dominance.
Addressing Security Concerns
The adoption of Chinese AI by Western enterprises raises legitimate security questions. DeepSeek's privacy policy explicitly states it stores user information on servers located in China and will "comply with legal obligations" and perform tasks in the "public interest." China's National Intelligence Law requires all organizations to "support, assist, and cooperate with national intelligence efforts."
A comprehensive security analysis published in March 2025 by AppSOC Research Labs found that both DeepSeek-R1 and Qwen-2.5 exhibited significant security vulnerabilities in testing, including high failure rates on jailbreaking tests and concerning rates of generating potentially harmful code when prompted adversarially.
However, this requires a closer look. Many enterprises are actually managing these risks by self-hosting open-source versions rather than using cloud APIs. Running models on internal infrastructure mitigates some—though not all—data sovereignty concerns. Model weights could potentially contain hidden behaviors, though the open-source nature allows for community scrutiny.
Microsoft President Brad Smith warned a U.S. Senate hearing that "the No. 1 factor that will define whether the U.S. or China wins this race is whose technology is most broadly adopted in the rest of the world."
The American Response
U.S. companies are responding to the competitive pressure. OpenAI has cut prices, and both OpenAI and Meta have accelerated their own open-source releases. But Western AI companies face a structural tension: their business models depend on premium pricing for hosted services, while Chinese competitors can give away the software and monetize through cloud infrastructure and ecosystem services.
Legislation is also emerging. U.S. lawmakers have floated proposals to block federal agencies from using Chinese-developed AI. Italy blocked access to DeepSeek and its mobile app in early 2025. However, such regulatory approaches may prove difficult to enforce with open-source software that can be downloaded, modified, and deployed anywhere.
What This Signals
The rapid adoption of Chinese LLMs represents more than a pricing disruption—it reflects a fundamental divergence in how AI technology is being commercialized. The open-source approach, combined with aggressive cost optimization and permissive licensing, has created a viable alternative to the Silicon Valley model, at least in the enterprise segment.
For enterprises, this means more options, competitive pricing, and greater deployment flexibility. For Western AI companies, it suggests that premium pricing for increasingly commoditized capabilities faces pressure. For policymakers, it raises complex questions about maintaining technological leadership while not cutting off access to competitive tools.
Brian Chesky's candid assessment of Airbnb's AI stack wasn't just corporate transparency—it offered a preview of the shifting landscape. When even close friends of Sam Altman choose Chinese alternatives for production workloads based on pragmatic considerations of cost and performance, the AI race has entered a new phase. It's no longer just about who builds the most capable model, but about whose technology achieves the broadest adoption at the most sustainable economics. And in that competition, Chinese developers are proving to be formidable contenders.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.
