America is Building the Best AI. China is Building the Default.

The United States produces the world’s most capable AI models. It attracts the most capital, operates the largest hyperscalers, and controls much of the advanced computing infrastructure on which modern AI depends. China is pursuing a different strategy: make models capable enough, inexpensive enough, and open enough to become the default choice for developers. Early evidence suggests that strategy is working.

Follow the Tokens

OpenRouter gives developers access to hundreds of AI models through a common interface. Its traffic does not represent the entire AI industry. It favors developers who compare models, monitor inference costs, and switch providers easily, and it underrepresents enterprises locked into long-term Microsoft, Google, or Amazon contracts. That said, it is my canary in the coal mine.

Over the past year, the share of OpenRouter traffic going to U.S. models has fallen sharply as Chinese-model usage has surged. DeepSeek nearly doubled its share of platform tokens from 9 percent in January to 18 percent in June, reached nearly 20 percent in early June, and has held OpenRouter’s top model position since mid-May.

OpenRouter represents only one segment of global AI usage, so these figures do not describe the whole market. They do show that developers in one of the world’s most price-sensitive, model-agnostic marketplaces are increasingly choosing Chinese models.

Revenue and Usage Tell Different Stories

The United States remains dominant by revenue. Fortune 500 companies do not choose models based solely on token price. They need security reviews, indemnification, regulatory assurances, data-residency controls, support agreements, and an accountable vendor.

Those requirements favor American hyperscalers and the closed models they distribute. Usage measures something else: which technologies developers learn, which models they incorporate into products, and which defaults become embedded in tomorrow’s software. The United States can lead in revenue and still lose share among the developers building the next generation of AI applications.

The Next Big Customers Aren’t Human

A growing share of AI usage is moving from conversational interactions to coding, agents, and machine-driven workflows. OpenRouter’s analysis of more than 100 trillion tokens documented rapid growth in programming workloads, while OpenAI’s Codex research found that intensive users run longer, more complex tasks and coordinate multiple agents concurrently.

A person may prefer a model’s writing style, personality, or conversational tone. An agent focuses on task completion, speed, reliability, and cost. It calls a model, evaluates the output, invokes a tool, reads the result, revises its plan, and repeats the process. A single workflow may generate thousands of model calls, so small price differences quickly become material opex.

Great Is the Enemy of Good

Stanford’s 2026 AI Index found that the capability gap between leading U.S. and Chinese models has nearly closed on several major evaluations. Anthropic’s leading model held an advantage of just 2.7 percent in the report’s aggregate comparison as of March. A U.S. government evaluation by NIST’s Center for AI Standards and Innovation found that DeepSeek V4 trailed leading U.S. models by approximately eight months across the domains it evaluated.

At the frontier, an eight-month lead can be strategically significant for national security, warfighting, weather forecasting, healthcare research, and other advanced applications. It matters far less to an agent reconciling invoices, rewriting product descriptions, testing software, conducting routine research, or sorting customer-service requests. In those use cases, the least expensive model that reliably clears the quality threshold wins.

Open Models Create Their Own Gravity

The largest U.S. AI labs generally keep their most capable models closed and accessible through APIs. Chinese companies including DeepSeek, Alibaba, Z.ai, Moonshot AI, MiniMax, and Xiaomi have emphasized open-weight distribution.

Open-weight models can be downloaded, modified, fine-tuned, and deployed on infrastructure chosen by the customer. They can run in private clouds, sovereign data centers, corporate facilities, or specialized devices. Organizations can optimize them for specific tasks without routing every interaction through the model developer’s servers.

Local models already run on laptops, tablets, and smartphones, and their capabilities are improving quickly. As more useful models operate entirely on personal devices, open weights, privacy, security, and national origin will move to the center of the AI debate.

Adopting a Model-Agnostic Strategy

Security, provenance, licensing, and political risk remain serious concerns. Regulated, confidential, proprietary, and national-security workloads require controls that often favor American vendors. For lower-risk workloads, developers choose the model that completes the task reliably at the lowest cost.

A research loop may use one model to gather information, another to classify it, a third to draft, and a frontier model to review the result. Orchestration layers select the appropriate model from the model garden based on the task, cost, latency, and required capability. The AI tech stacks we are helping our clients design are all model-agnostic.

The United States produces the world’s most capable models. China produces models the world can afford to deploy at scale. Capability creates strategic leverage and enterprise value. Distribution creates developer habits, technical standards, ecosystem dependencies, and market power. America leads at the frontier. China is positioning itself as the default. In an agentic economy, the default may prove more valuable than the benchmark.

Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.