Chinese AI Models Now Own 60% of Global API Traffic as US Providers Collapse from 70% to 30%

In June 2025, US AI providers — Google, OpenAI, and Anthropic — collectively accounted for roughly 70% of all token consumption routed through OpenRouter, the largest public API routing platform for AI models. One year later, that share has collapsed to approximately 30%.

The inverse of that decline has been captured by Chinese providers. Models from DeepSeek, MiniMax, Moonshot AI (Kimi), Zhipu AI (GLM), and Xiaomi now collectively account for more than 60% of total tokens processed on OpenRouter by mid-2026, according to platform traffic data. That is up from less than 2% in early 2025 — a structural market shift of extraordinary speed and scale.

The Numbers Behind the Shift

The scale of individual Chinese providers on the platform is striking. MiniMax’s M2.5 model processes 4.55 trillion monthly tokens through OpenRouter. Moonshot AI’s Kimi K2.6 handles 4.02 trillion monthly tokens. Xiaomi, which entered the AI model market relatively recently, is processing 4.21 trillion tokens weekly — implying a monthly volume that dwarfs both competitors.

DeepSeek V3.2 and V4 collectively hold approximately 17.6% of total platform share as of June 2026. To put that in context: DeepSeek’s token volume on OpenRouter exceeds Google’s combined share (12.5%) and OpenAI’s combined share (8.4%). Zhipu’s GLM-5.1 rounds out the top tier.

The composition of usage matters as much as the volume. Programming tasks now account for over 50% of all OpenRouter traffic, up from 11% in early 2025. Chinese models have specifically optimized for this use case, and their benchmark performance on software engineering tasks has reached parity or superiority with US closed-weight alternatives at dramatically lower price points.

Why Chinese Models Are Winning

The competitive thesis for Chinese AI providers rests on four interlocking advantages.

Pricing. The most immediate driver is cost. Chinese models — particularly the open-weight variants available for self-hosting — undercut US providers by 10 to 20 times on comparable tasks. DeepSeek V4, for example, offers context windows and coding capabilities that benchmark comparably to mid-tier US models while pricing at a fraction of the cost. For cost-sensitive developers building high-volume applications, the math is not difficult.

Open-weight availability. Several leading Chinese models are released as open weights, meaning organizations can download and run them on their own infrastructure with no per-token fees whatsoever. This makes them attractive not only to budget-constrained startups but to enterprises with data sovereignty requirements, large-context batch processing needs, or a preference for infrastructure independence. The self-hosting option effectively removes the per-query cost ceiling entirely.

Context window engineering. Chinese AI development efforts have focused heavily on extending context windows and optimizing for long-document and multi-file tasks. MiniMax M2.5 offers a context window measured in millions of tokens — a capability that is particularly powerful for repository-level code analysis, legal document review, and scientific literature synthesis.

Coding specialization. The shift of OpenRouter usage toward programming tasks played directly into the strengths that Chinese labs have prioritized. Benchmark comparisons on SWE-Bench, HumanEval, and internal AI IDE evaluations show Chinese models competing within a narrow margin of the best US coding-focused models, at a fraction of the cost.

The Regulatory and Security Dimension

The OpenRouter data tells a market story. But it lands in a geopolitical context that makes its implications more complex than a simple competitive analysis.

Directing production traffic to Chinese-controlled AI providers raises questions that go beyond cost optimization. Data processed by Chinese AI models — including potentially proprietary business logic, source code, customer data, and strategic communications — may be subject to Chinese data governance laws that require cooperation with government authorities. The legal frameworks governing Chinese technology companies’ data handling obligations differ materially from the standards applied to US or EU providers.

US legislators have noticed the shift. Multiple bills proposing restrictions on the use of Chinese AI models in government and critical infrastructure contexts have been introduced in 2026. The Trump administration postponed a planned AI oversight framework in May 2026, leaving no formal US framework governing how frontier models from any jurisdiction are evaluated before deployment in sensitive contexts. That regulatory gap has effectively allowed market forces to drive adoption decisions that have national security implications.

The defense and intelligence communities have been more direct: internal guidance at multiple federal agencies explicitly prohibits the use of Chinese AI models for any government-related tasks, and some defense contractors have begun mandating US-origin model requirements in their AI procurement policies.

For commercial enterprises outside the government sphere, the guidance is less clear. The security considerations are real but manageable with proper architecture: data can be classified, routing decisions can be task-specific, and hybrid deployments can use Chinese models exclusively for non-sensitive batch tasks while reserving US models for anything involving confidential information.

What This Means for US AI Labs

The competitive implications for OpenAI, Anthropic, and Google are real but context-dependent. OpenRouter is not the entire AI market — it is a routing layer predominantly used by developers, AI startups, and technically sophisticated enterprises building on top of foundation models. The enterprise procurement landscape, which drives a disproportionate share of AI revenue, still skews toward US providers due to compliance requirements, established vendor relationships, and risk tolerance.

But OpenRouter is a leading indicator. The developer community that drives enterprise adoption cycles builds with what is cheapest and most capable first, and then advocates for organizational adoption. A developer community that has shifted to 60% Chinese model usage is a community that, when asked to recommend AI infrastructure for their employer, will frequently cite Chinese providers.

OpenAI and Anthropic are attempting to respond on pricing — both have introduced lower-cost models designed to compete with the efficient tier where Chinese providers are strongest. But the pricing asymmetry is structural, not marginal: Chinese labs operate at lower cost structures, benefit from state support in various forms, and in the case of open-weight models have essentially decoupled their development costs from their per-token revenue requirements.

The deeper structural response is differentiation. US providers are investing in capabilities that Chinese labs are less focused on — long-horizon agentic tasks, frontier reasoning benchmarks, multimodal capabilities beyond text and code, and enterprise integration depth. The argument is that the highest-value AI workloads are better suited to frontier US models even at premium pricing. That argument is defensible. It is less convincing to the cost-sensitive majority of developers who are routing the traffic that shows up in OpenRouter’s statistics.

The View From Asia-Pacific

For readers in Taiwan and the broader Asia-Pacific region, the OpenRouter data is a particularly relevant data point. Taiwan sits at the intersection of the US-China technology divide in a uniquely complex position: it is home to TSMC, which manufactures chips for both US AI labs and, subject to export controls, Chinese AI providers. Its tech-forward developer community has access to — and is actively using — both US and Chinese AI services.

Taiwanese enterprises and developers navigating this landscape face a choice with no easy answer. Chinese AI models offer compelling cost and capability profiles. US AI models offer more predictable regulatory compliance and alignment with Taiwan’s strategic relationships. The appropriate answer, for most organizations, is a deliberate hybrid strategy rather than a default to either pole.

The OpenRouter data is a market signal about what the global developer community, unconstrained by compliance requirements, has found most economically rational. The question for organizations in Taiwan and beyond is how much of that rationality they can act on, and where the constraints make the cost premium for US models the right trade.