Why America's startups are secretly relying on China: When their "own" AI model suddenly comes from China

80% cost savings: How Chinese AI models are silently conquering the American market

A secret revolution in Silicon Valley: Why US startups are suddenly relying on China's AI

For a long time, Silicon Valley was considered the undisputed and invincible epicenter of global AI development. But away from the big marketing stages, a remarkable shift is currently underway: More and more US startups and established tech giants are quietly adopting Chinese open-source models like Qwen, Kimi, and DeepSeek. The reason for this isn't geopolitical sympathy, but hard-nosed economics. The Chinese alternatives now offer performance in programming and automation tasks that rivals top Western models—but at a fraction of the cost. This development is putting massive pressure on industry giants like OpenAI, forcing drastic price reductions and raising fundamental questions about the future distribution of power within the industry. The following article examines the impressive figures behind this quiet revolution, uses prominent company examples to illustrate how businesses are adapting their strategies, and explains why the issue of data sovereignty plays a crucial role in this new market.

A continent silently changes sides

For two decades, Silicon Valley was considered the undisputed epicenter of software innovation, and just two years ago, the idea that Chinese technology would become the foundation for the next generation of companies there would have seemed absurd. But that is precisely what is happening now, and at a speed that is surprising even seasoned market observers. Investor Martin Casado of Andreessen Horowitz summed it up perfectly in a much-quoted interview: If founding teams pitch at a16z today and rely on open AI models, there is roughly an 80 percent probability that it is a Chinese model. It's important to note the clarification Casado later added: This 80 percent refers to the portion of startups that use open models at all, which is estimated to be 20 to 30 percent of all AI startups. Extrapolated to the entire startup landscape, this means that approximately 16 to 24 percent of all US AI companies have Chinese open models firmly embedded in their technology stack. This differentiation may at first glance lessen the drama of the statement, but it does not change the fundamental realization that a significant and rapidly growing part of the American innovation sector is turning away from domestic suppliers.

The numbers behind the claim: What OpenRouter really shows

Anyone wanting to objectively understand this market shift cannot ignore the data from the OpenRouter platform, considered the largest independent intermediary service for AI model access, allowing developers to switch between hundreds of models from different providers. The usage data collected there provides an objective picture of genuine market demand because it is not based on announcements or marketing claims, but on tokens actually processed—the basic unit in which AI language models process text. At the end of 2024, the share of Chinese models in the platform's total token volume was still around 1.2 percent, a negligible figure that received little attention. Within just a year and a half, this ratio changed more than fiftyfold. In February 2026, Chinese models first exceeded the 30 percent mark of weekly token volume among American users, and this figure has remained above that threshold every single week since. By July 2026, the share reached 46 percent in peak weeks, while American models accounted for 35.7 percent during the same period. If one considers only the ten most-used models on the platform, an even clearer picture emerges: In the week of February 24, 2026, 61 percent of the token volume among these top ten models was attributable to Chinese developers. According to OpenRouter, by the end of July 2026, Chinese models even occupied all five top spots in the global usage ranking, with the total platform volume now exceeding 20 trillion tokens per week, of which Chinese models processed more than 60 percent.

This development cannot be dismissed as mere statistical noise, but rather reveals a clear trend over the past year and a half. At the same time, methodological limitations must be taken seriously. While OpenRouter is the most important independent indicator of developer behavior, it does not represent the entire global AI market, but primarily cost-sensitive, technically savvy user groups who place great value on open weights. Consumer applications like ChatGPT or direct enterprise API contracts with Anthropic are barely reflected in these figures. This is why studies like the major annual report from OpenRouter and a16z from 2025 paint a somewhat more moderate picture, showing that proprietary North American models still accounted for the majority of tokens over the entire year, while Chinese open models averaged around 13 percent. The truth, therefore, lies somewhere between the headline and the all-clear: the trend is real, is visibly accelerating, but primarily affects specific usage segments, especially software development and automated agent work, which together now account for well over half of all tokens on the platform.

Why programming and automation in particular are driving the upheaval

A key element in understanding this shift lies in the nature of the workload that now dominates AI systems. As recently as 2024, programming accounted for just eleven percent of the total token volume on OpenRouter, whereas today more than half of all processed tokens are dedicated to programming tasks and autonomous agent workflows, where models independently execute multi-stage tasks. Chinese models have made enormous strides in this area because they are specifically optimized for coding and agent work. OpenRouter's Chief Operating Officer, Chris Clark, explained the underlying mechanism unequivocally: Chinese open models have gained a disproportionately strong foothold in precisely those agent-based workflows that are heavily used by US developers. In other words, it's not an ideological decision, but a pragmatic response to a price-performance ratio that is hard to beat for high-volume, repetitive programming tasks.

The price shock that started it all

The economic core of this development lies in a price difference so significant that it fundamentally alters every calculation made by a finance manager. One million issue tokens cost around $50 with a top-tier model like Anthropic's Fable, while the same quantity costs approximately $15 with Moonshot's Kimi K3, $4.40 with Zhipu's GLM-5.2, and a mere $0.87 with DeepSeek's V4 Pro model. These figures practically mean that a company can pay five to sixty times the price for the same task, depending on the model chosen. For a company handling millions of daily requests, this is not a minor issue, but a crucial cost factor that determines profit margins and the scalability of entire business models.

This price dynamic also explains why OpenAI reduced its prices by up to 80 percent during 2026. A company with a billion-dollar valuation and enormous market power doesn't lower its prices out of goodwill, but due to sheer competitive pressure. When customers can switch to a provider with a single click who performs the same task for a fraction of the cost, an established provider is left with only the choice between adjusting its prices or losing customers. This mechanism is a textbook example of classical competitive economics: As soon as a market is confronted with nearly homogeneous goods where the quality difference becomes marginal for the average use case, the price slides toward the marginal cost of the cheapest credible provider. This is precisely what we are currently witnessing in the language modeling market.

When corporations recalculate their accounts: Specific company examples

Abstract market data only gains real persuasive power when you see how concrete companies actually restructure their internal processes. The crypto trading platform Coinbase provides the most vivid example of this in 2026. CEO Brian Armstrong published detailed insights into his company's internal cost strategy in the summer of 2026, explaining that they had nearly halved AI spending while simultaneously increasing usage volume. The key lever was switching the default models in an internal AI gateway to Zhipu's GLM 5.2 and Moonshot's Kimi 2.7 Code – two open-source Chinese models that handle routine tasks such as code reviews, summaries, and drafting, while more expensive, high-end models like Anthropic's Opus remained reserved for complex planning tasks. It's worth noting that this switch wasn't the only cost-cutting measure. Coinbase simultaneously increased the cache hit rate of its internal AI tool from 5 to 60 percent, which in itself explains a significant portion of the savings, as repeated queries can now be answered almost free of charge from the cache instead of triggering a new, costly model query each time. This example illustrates that the shift to Chinese models is rarely an isolated decision, but rather part of a broader strategy for intelligent resource allocation, where the choice of model is just one of several levers.

Cursor, a development platform and one of the most celebrated tools for AI-powered software development, also found itself in a revealing situation in March 2026. The company presented its new model, Composer 2, as its own in-house development, boasting supposedly top-tier coding performance. However, within hours of its release, observant developers discovered an internal model identifier in the API's technical response data that unmistakably referenced Kimi K2.5, the open-source model from Moonshot AI. Cursor was subsequently forced to publicly admit that Composer 2 was indeed based on the Kimi K2.5 architecture, although the company emphasized that approximately three-quarters of the total computing power used to fine-tune the final model came from its own training, with only about a quarter based on the original Kimi foundation. Co-founder Aman Sanger self-critically acknowledged that it had been a mistake not to mention the Kimi foundation in the initial announcement. This incident illustrates a larger pattern: Many supposedly American AI innovations are now built on Chinese basic models beneath the surface, without this being communicated externally, which observers sometimes describe, somewhat smugly, as a kind of silent concealment of the actual technological origin.

Besides these prominent examples, numerous other companies have restructured their AI infrastructure to align with Chinese open models. These include the food delivery service DoorDash, which outsources simple programming tasks to Kimi while keeping critical decisions with more expensive, high-end models, as well as companies like Snowflake and the AI ​​startup Lindy, which pursue similar cost optimization strategies. According to publicly available reports, the software company Siemens relies on training Qwen with its own industrial data to develop industry-specific assistance systems, rather than building entirely new base models from scratch—a logical economic decision given the immense costs of training new language models.

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

Four names to remember

To understand the current market shift, one must know the key players behind this development. First and foremost is Qwen, developed by the Chinese technology company Alibaba, which is now considered the world's most downloaded open AI model, boasting an impressive reach with around 700 million downloads. Qwen is regarded as a versatile all-rounder with particular strengths in natural language processing and document analysis, making it well-suited for text production, reporting, and customer service. The second model deserving special attention is Kimi K3 from Moonshot AI, which was introduced in July 2026 as the largest open model ever released and performs in programming benchmarks approaching leading proprietary models. It is particularly well-suited for software development and processing long documents, as it operates with a context window of up to 256,000 tokens, corresponding to the amount of text the model can process and retain simultaneously. As a third model, GLM, from the Chinese company Zhipu AI (now operating under the Z.ai brand), has established itself as a specialist in the reliable control of tools and systems, making it particularly valuable for automation tasks and agent-based workflows. Finally, DeepSeek positions itself as the undisputed price-performance champion of the market, whose models deliver logical reasoning and programming tasks at a tiny fraction of the cost of established providers, making them ideal for all tasks that occur at very high scale.

The question of data sovereignty and regulatory grey areas

One aspect often underestimated in public debate concerns the fundamental difference between using a Chinese model via the original provider's cloud API and running the open model weights on one's own or European infrastructure. All four models presented here—Qwen, Kimi, GLM, and DeepSeek—are released under open licenses. This means that companies can download the underlying mathematical parameters of the model and run them on their own servers or with a European hosting provider. In this scenario, all data remains physically and legally with the operating company, which, from the perspective of the European General Data Protection Regulation (GDPR), can be a significantly cleaner setup than using an American cloud infrastructure, where data may be subject to the American CLOUD Act. This distinction is crucial because it demonstrates that the political origin of a model and the actual risk to data protection and national security are two distinct issues that are often inappropriately conflated in public discourse.

At the same time, legitimate questions remain that preclude the uncritical adoption of Chinese models. In 2026, the US Congress launched investigations into companies like Airbnb and Cursor related to their use of Chinese AI models. These investigations focus not primarily on the open model weights themselves, but rather on issues of security auditing, potential influence, and strategic dependence on technology from a geopolitical rival. Independent security assessments have also shown that some of the most widely used Chinese models receive comparatively weak scores in standardized security tests, which could be particularly relevant for companies with stringent compliance requirements. However, these concerns largely relate to the use of models via hosted cloud services provided by the Chinese vendors themselves, not to the independent operation of the open weights on European or US infrastructure, where companies retain full control over data flows, logging, and security configuration.

The strategic logic behind the model mix

The observed company examples reveal a clear strategic pattern that has become established in professional discourse as a kind of barbell strategy – a deliberate distribution of tasks between two opposing poles. Simple, high-volume, and repetitive tasks such as text summaries, standard code reviews, or translation tasks are increasingly delegated to inexpensive, often Chinese, open-source models, while strategically critical, complex, or particularly sensitive tasks remain with more expensive, high-end models from Western providers. This division is highly rational from an economic perspective, as it optimizes the cost-benefit ratio of each individual task instead of using a single, usually overpriced, high-end model for all use cases. The food delivery service DoorDash is already implementing this principle in practice by systematically forwarding simple programming tasks to Kimi, while business-critical decisions remain with the more expensive, established high-end models. For companies that want to professionalize their AI cost structure, this ultimately means moving away from the question of which provider is fundamentally the best, towards a differentiated question of which specific task best suits which model with which cost profile.

What this means for the Western AI industry

The shift towards Chinese open-source models presents established American vendors with a structural dilemma that extends beyond short-term price adjustments. Companies like OpenAI and Anthropic have built their business models for years on the premise that technological superiority justifies a sustained price premium. However, once performance differences become so small for most practical use cases that they no longer represent an economically relevant difference for companies, competition inevitably shifts toward price. In this race, Chinese vendors have a structural advantage due to lower operating costs, government subsidies, and a more aggressive open-source strategy. OpenAI's response of cutting its own prices by up to 80 percent is less a sign of strength than an admission that its previous price premium was no longer sustainable without a corresponding competitive response. For the entire Western AI industry, this signifies a transition from a phase in which innovation alone guaranteed sufficient pricing power to a phase in which cost efficiency, infrastructure control, and specialized additional services become the decisive differentiating factor.

The upcoming postponements

Current developments suggest that this trend will not reverse in the foreseeable future, but rather become even more diversified. Instead of a single dominant model, an increasingly fragmented landscape is emerging, in which no single open model holds more than a quarter of the total market share among open models, indicating a healthy, competitive market with several equally viable alternatives. For companies in Europe, and especially for medium-sized businesses with limited IT budgets, this presents a rare opportunity to benefit simultaneously from falling costs and increasing technological maturity, provided they are willing to address the legal and security frameworks for operating open models. Those who base their AI strategy solely on a single, usually American, provider today risk not only excessive costs but also strategic dependence on a market that is currently changing faster than any other technology sector in recent memory. The most pragmatic answer to this uncertainty lies in a conscious diversification of one's own model portfolio, combined with clear internal governance that assigns which task class to which model and hosting location.

From visibility to trust: Your scalable path with Xpert.Digital - Image: Xpert.Digital

In industrial B2B, sustainable business relationships rarely emerge overnight. They develop step by step – through visibility, professional relevance, recurring touchpoints, and growing trust. Xpert.Digital's 4-stage model addresses precisely this: It offers a structured path that begins with a manageable entry point and can evolve into deeper collaboration in business development if needed.

Instead of relying on loud marketing promises, this model puts the relationship at the forefront. Companies start with clearly defined, easily calculable measures and then decide, based on their own experience, how far they want to expand the collaboration. A key factor for this undisturbed trust-building process: The platform completely avoids annoying advertising ads, so the editorial focus remains solely on the companies' expertise.

More information here:

• From visibility to trust: Your scalable path with Xpert.Digital

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors