The divergence between Sino-US AI is no longer a passive outcome but an active choice.

—Lead

01

Two Reports

After the fierce AI-to-Consumer (AI2C) battle during the Spring Festival, two less flashy but highly substantive reports revealed a critical yet underappreciated trend in Sino-US AI competition.

The first type of report goes beyond mere statistics—it stems from one-on-one enterprise-level market research and industry tracking. For instance, in February 2026, Frost & Sullivan, in collaboration with HeadAI, released the . This report systematically disclosed structural shifts in China's enterprise-level large model adoption, including query volumes, vendor market shares, deployment models, and open-source/closed-source preferences.

Key data point: In H2 2025, daily average queries for enterprise-level large models in China reached 37.0 trillion tokens, up 263% from H1.

Notably, the shift from closed-source to open-source dominance accelerated significantly post-H2 2025, doubling from 25.9% in H1 2025 to 48.5% in H2.

The second type of report comes from "the footprints left by machines." Studies like OpenRouter and a16z's (released in late 2025), based on metadata analysis of over 100 trillion tokens processed on their platforms in the past year, address a fundamental yet hard-hitting question: Which models are developers and enterprises actually using, how, and at what scale?

When "survey choices" align with "behavioral traffic patterns," a striking conclusion emerges: Sino-US AI is clearly diverging.

The U.S. remains focused on flagship model breakthroughs, chasing the "ceiling" of capability; China, meanwhile, appears to be executing a massive engineering project—disaggregating, orchestrating, cost-reducing, and packaging model capabilities for integration into complex industrial settings.

02

The 37-Trillion-Token Counterattack

Among the two reports, I first encountered OpenRouter's findings, but many data points only gained context through Frost & Sullivan's analysis.

My core assessment: Sino-US AI is diverging, but this isn't a simple "who's stronger" contest. Instead, it reflects systemic bifurcation in technical approaches, industrial organization, and commercial logic.

The pivotal factor here is open-source.

Globally, China's dominance in open-source AI is intensifying.

OpenRouter's report shows open-source models' market share steadily rising over the past year, reaching nearly 30% by late 2025. Chinese models contributed significantly to this growth—in some weeks of H2 2025, they accounted for 30% of global traffic, up from just ~2% in early 2024.

Moreover, this isn't driven by a few Chinese tech giants alone. Frost & Sullivan noted that in H2 2025, Chinese vendors accounted for 90.2% of all newly released open-source large models globally, far surpassing overseas contributions.

Clearly, China has evolved from a key AI participant to the core engine and primary exporter of global open-source large model innovation.

Open-source is no longer just a technical choice but a strategic advantage in China's global AI competition.

To some extent, this reflects Chinese enterprises' success in dismantling "computational hegemony" and "cost barriers." While the U.S. attempts to build a "computational moat" through restrictions on top-tier chips (e.g., H100/B200), Chinese firms have achieved "efficiency-driven scaling" via widespread open-sourcing, architectural innovations (e.g., Mixture-of-Experts models), and algorithmic optimizations.

Evidence shows that with constrained computational resources, China can achieve results comparable to leading U.S. models at 1/10th or even 1/50th the cost through hyper-efficient training. This grants Chinese AI a dimensional advantage in commercial deployment and inference costs.

This mirrors a classic "rural surround urban" strategy in the global developer ecosystem. By late 2025, Chinese open-source models accounted for over 17% of downloads on Hugging Face, even surpassing U.S. models in some months.

While precise data is unavailable, if 50% of global startups and developers fine-tune and deploy AI using open-source models like QianWen, Doubao, DeepSeek, Wenxin, or Kimi, China effectively sets the de facto global AI implementation standard. Such ecosystem-driven penetration is more resilient than mere API subscriptions.

Open-source models can run on local enterprise servers, eliminating the need to upload data to U.S. clouds. This provides an irreplaceable "Plan B" for privacy-sensitive clients worldwide, particularly in Southeast Asia, the Middle East, and Europe.

03

Will U.S. AI Business Models Collapse?

To date, U.S. AI giants like OpenAI and Anthropic still rely on exorbitant computational investments and high API pricing to sustain valuations, with their customer bases continuing to expand.

However, China's high-performance, low-cost (or even free) open-source models are directly eroding these firms' gross margins.

Simply put, if a $0.10 open-source model can solve 95% of the problems a $20 closed-source model addresses, the U.S. giants' "premium myth" collapses, triggering capital market doubts about their business model sustainability.

More critically, open-source enables bidirectional transparency in technical pathways. Talent and innovation spillovers become inevitable. For example, when Chinese firms open-source their R1 reinforcement learning frameworks, they effectively "light the path" for global innovation. Against such forces, the "moats" built by U.S. firms through painstaking R&D become exceedingly difficult to defend against collective open-source community wisdom.

This explains why U.S. AI progress remains tied to a few giants' innovations, while China nurtures a cohort of lighter, faster AI firms (e.g., Kimi). Rather than pursuing Google-style vertical integration, these firms partner with leading AI cloud providers like Alibaba, Huawei, Baidu, and Volcano Engine, focusing intensively on model R&D and monetizing rapidly through open-source distribution.

This creates a stark contrast: U.S. AI leadership is increasingly concentrated among a few players—efficient but path-limited—while China's more numerous, agile model firms, though unlikely to outcompete giants in closed-source markets (the "Hundred Models War" there has ended), thrive in open-source domains. With focused capabilities and maturing, self-sustaining business models, China's open-source output capacity continues to surge without apparent bottlenecks.

04

The "Half-Step" Dilemma: Following or "Lane-Changing"?

Admittedly, China still trails the U.S. in top-tier model development, fueling concerns that Chinese AI will "always lag by half a step." This is the dominant worry today, but two perspectives offer clarity:

1. Why does it always seem like "half a step"?

Partly because the U.S. currently controls the definitions of "leading edge" (e.g., o1 defined reasoning, Sora defined video generation). As long as Chinese firms chase U.S.-defined "SOTA (State of the Art)" benchmarks, they'll appear as perpetual followers.

This perception isn't mere illusion. The U.S. holds advantages in computational power and corpus scale, with leading positions in top-tier compute clusters and global English-language datasets.

Though Chinese has become the second-largest language in open-source communities, its actual market share is just 5%—a significant gap versus the leader.

I must also note: When "$0.10 solves 95% of problems," premium pricing inevitably contracts. However, U.S. giants can reprice using "system capabilities" (toolchains, inference stability, security, enterprise services) rather than model performance alone.

But does open-source mean permanent passivity? I think not.

This is precisely what Chinese firms are doing—eliminating the "half-step" gap through "lane-changing":

In short: From "stacking compute" to "stacking logic." While the U.S. pursues AGI's absolute upper bounds, China optimizes for "intelligence per unit of compute." In hardcore domains like reasoning, mathematics, and coding, Chinese open-source models (e.g., DeepSeek R1) matched or slightly outperformed U.S. leaders in some 2025 benchmarks.

A massive advantage lies in AI's productivity land (implementation): Victory hinges not on benchmark scores but on solving real-world problems in industry, education, and healthcare. China boasts the world's densest application scenarios, creating "application feedback → model iteration" cycles that could enable vertical-domain surpassing.

According to Frost & Sullivan, as large models integrate into core processes like customer service, risk control, marketing, and R&D, most queries are now system-triggered and workflow-embedded, driving steady, continuous token consumption growth. The query surge in H2 2025 stemmed primarily from task-level automation expansions in business processes, with intelligent agents becoming the main driver of new queries.

Chinese firms now rank second globally in agent development after the U.S. An MIT-Stanford report identified 30 leading global agents, with the U.S. first and China second, while others are scattered across nations.

China not only has the world's largest demand scenarios (top global manufacturer) but is also accelerating in agent R&D. For most agent firms, success hinges on access to robust open-source large models as innovation foundations—and as noted earlier, China's dominant global open-source models and superior Chinese-language support are key enablers for massive high-quality agent development.

In short: If AI is electricity, the U.S. is developing the most powerful "generators," while China is making "electricity costs" negligible via open-source and building a global "power grid." Chinese firms need not rush to define generators; by controlling global grids and appliances, they can still win the competition.

05

An Active Choice

The divergence between Sino-US AI is no longer a passive outcome but an active choice.

Some things have changed permanently—like our confidence.

This means that in China's and even the global developer ecosystem, Chinese open-source forces are no longer just "options" but have become de facto "industrial foundations."

To some extent, Chinese AI firms, through widespread open-sourcing, have shattered the U.S. closed-source models' "black box barriers." They demonstrate to global enterprises that AI need not be costly fixed assets but can become controllable, even privatizable, "digital assets."

In late 2024, Chinese models held just 1.2% of global token usage. Back then, Chinese developers lamented GPT-4's insurmountable barrier (barriers).

By mid-2025, a watershed emerged. According to OpenRouter, from mid- to late 2025, Chinese open-source models' token contributions neared 30% of global traffic (including both open- and closed-source, hence a larger base) within a short period.

This surge saw growth among both tech giants and "light-and-fast" large model firms. While consolidation trends are evident, niche players remain vibrant.

In short, the future Chinese AI market will likely feature:

—In infrastructure, full-stack enterprises like Alibaba, Baidu, and Huawei—with open-source capabilities spanning hardware to applications and sustained cloud investments—will serve as the bedrock for AI adoption across industries.

—In ecological diversity, China's model firms won't rapidly oligopolize (and perhaps shouldn't). Instead, leveraging infrastructure from giants, a growing array of differentiated large model firms will coexist. This decentralized approach offers a unique advantage against U.S.-style centralized AI.

—In agent development, open-source ecosystem friendliness will be decisive, with Chinese firms accelerating their catch-up.

Thus, China's productivity-focused large model market has embarked, during this brief transition, on the path of deep industrial empowerment it should have taken—but hadn't pursued far enough—until now.

Epilogue

China's AI Deep Dive

Are Sino-US AI paths diverging? The answer is clear.

The U.S. still seeks radical breakthroughs, pushing toward AGI's ceiling; China focuses on infiltrating AI into factories, banks, hospitals, and millions of entrepreneurs' codebases.

Over the next decade, as Sino-US AI differentiation sharpens, such industrial-level penetration will become the hardest foundation in China's tech rivalry.

Divergence doesn't imply confrontation but rather two paths exploring humanity's intelligence boundaries—one reaching skyward, the other grounding deep. Ultimately, they may converge in the vast expanse of the cosmos.