Chinese AI Accounts for 58% of US Compute Consumption: How Open-Source LLMs Are Disrupting Silicon Valley

From 10% to 58%: A Quiet Migration in AI Consumption

According to an analysis shared by a Japan-based tech expert on Bilibili, data from the open-source model aggregation platform OpenRouter is upending assumptions about the global AI landscape. Before the first half of 2025, Chinese AI models accounted for less than 10% of overall AI consumption in the US. By July 2025, that figure had surged to 58% — Chinese AI had crossed the halfway mark in usage volume.

OpenRouter is one of the most influential model aggregation platforms in the AI developer community, providing a unified API endpoint that integrates over 400 large language models from different providers. Developers only need to connect to a single OpenRouter endpoint to freely switch between OpenAI's GPT series, Anthropic's Claude series, Google's Gemini series, and various Chinese open-source models — without registering and maintaining multiple API accounts separately. This architecture makes it an important window into real-world AI model usage preferences globally, because users are making choices in a completely free market environment with no platform lock-in effects.

The expert specifically noted that this refers to "usage volume" rather than "revenue." This distinction is crucial. In the large language model space, consumption is typically measured in "tokens." A token is the smallest unit of text that a model processes — one English word corresponds to roughly 1 to 1.5 tokens, while one Chinese character typically corresponds to 1 to 2 tokens. When we say Chinese model token usage has exceeded half, it means that on the OpenRouter platform, more than half of the actual computational load — including input prompts and model-generated responses — is flowing to Chinese models. However, since Chinese models are often priced at one-tenth or less per million tokens compared to American models of equivalent performance, their revenue share is far lower than their usage share. This "high volume, low price" dynamic demonstrates precisely how Chinese models are capturing real-world application scenarios through extreme cost-effectiveness. This tipping point occurred mainly between May and June 2025, completing the leap from minority to majority in barely two months.

Use cheap, capable Chinese models for lightweight tasks, then switch to more expensive frontier models for difficult ones — it's precisely this "switch on demand" mechanism that lets users continuously vote with their usage, and the result is that more and more people are choosing Chinese AI.

Silicon Valley's Collective Shift: DoorDash, Airbnb Lead Adoption

What's even more striking is that those adopting Chinese AI aren't fringe players — they're well-known tech companies at the heart of Silicon Valley. The expert cited multiple cases of companies that have publicly or semi-publicly adopted Chinese open-source LLMs.

According to the analysis, US food delivery giant DoorDash is rapidly rewriting code and restructuring company-wide systems, using Kimi for essentially everything except a handful of sensitive components. Kimi is a large language model developed by Chinese AI startup Moonshot AI, founded in 2023 by researchers from Tsinghua University's AI lab. Kimi initially gained fame for its ultra-long context window (supporting 2 million character inputs), then continued iterating on code generation, logical reasoning, and other capabilities. Its open-source version Kimi-K2 has performed excellently on multiple programming benchmarks, and is considered to have capabilities on par with or even superior to Claude for code refactoring tasks — at a fraction of the cost.

Meanwhile, Airbnb CEO Brian Chesky publicly stated that Airbnb has adopted Alibaba's Qwen (Tongyi Qianwen) across the board. Qwen is Alibaba Cloud's large language model series, with its open-source version Qwen2.5 covering multiple scale tiers from 500 million to 72 billion parameters, excelling in multilingual understanding, tool calling, and long-text processing. Because Alibaba chose to open-source the model weights under the Apache 2.0 license, enterprises can deploy and further develop it commercially without restrictions — highly attractive for a company like Airbnb that operates globally, needs multilingual support, and is highly sensitive to data sovereignty.

An even more extreme example comes from the well-known AI company Lindy. Lindy is a startup focused on AI automation workflows, with a core product that allows users to create AI assistants capable of autonomously executing complex tasks across customer service, sales follow-up, data analysis, and other scenarios. According to the expert, Lindy was once 100% dependent on Claude Code, then switched entirely to DeepSeek, saving hundreds of millions of yen in costs within a year and significantly boosting profits. DeepSeek was founded in 2023 by Chinese quantitative investment fund High-Flyer. Its open-source model DeepSeek-V3 and reasoning model DeepSeek-R1 attracted global attention in early 2025. DeepSeek's core innovations lie in its Mixture of Experts (MoE) architecture and Multi-head Latent Attention (MLA) mechanism, enabling it to achieve performance close to or exceeding frontier models at training costs far below competitors — reportedly only a fraction of what equivalent-scale American models cost. This demonstrates that Chinese AI is no longer a "toy" or "curiosity" — it's combat-ready and being deployed at scale in production environments at core US tech companies.

The expert also revealed an intriguing detail: a friend working at a major US tech company (a Microsoft-affiliated enterprise) told him that while the company hasn't gone public about it, they've been extensively using Kimi internally for some time. Today, more and more companies are willing to openly acknowledge their use of Chinese AI models.

Not Every Trip Requires a Ferrari: The Cost-Performance Logic of Chinese Models

Why are these Silicon Valley companies — with access to frontier labs — making the switch? The expert cited an analogy widely circulated in the industry: "Not every trip requires a Ferrari."

The logic is straightforward: to pick up groceries from a nearby supermarket, an economy sedan is perfectly sufficient — no need to drive a Ferrari. When enterprises actually categorize tasks and calculate costs, they discover that the vast majority of scenarios don't require the most premium, most expensive models. Chinese open-source models like DeepSeek and Kimi — the "economy cars" — perform far beyond expectations and are perfectly capable of handling everyday needs.

This reflects an important technical reality in AI applications: roughly 70% to 80% of enterprise AI calls are "routine tasks" — text summarization, format conversion, basic code generation, simple Q&A, data cleaning, and so on. These tasks don't require the strongest reasoning capabilities or the largest parameter scales. A medium-performance model at extremely low prices can complete them with nearly identical accuracy. Only the remaining 20% to 30% involving complex multi-step reasoning, frontier scientific problems, or highly creative writing truly require flagship models like GPT-4o or Claude Opus. Once enterprises recognize this, routing large volumes of routine traffic to low-cost models becomes the rational choice — and Chinese open-source models perfectly fill this need.

The expert extended this into a broader point: give the market free choice, and people will naturally select the optimal solution. This also indirectly suggests that heavy-handed regulatory policies on AI may not work — the market is voting with its feet, breaking what was assumed to be America's AI monopoly.

The expert also mentioned that Cursor, the well-known AI programming startup acquired by SpaceX for a massive sum, has been found to use Chinese models behind its new features and services. Cursor is a deeply modified AI-native code editor based on VS Code, developed by Anysphere. It deeply integrates large language models into the programming workflow, supporting cross-file code understanding, auto-completion, intelligent refactoring, and natural language instruction programming. Cursor rapidly became one of the most popular AI programming tools in the developer community in 2024, with its valuation exceeding $10 billion at one point. In 2025, reports indicated Cursor was brought into Elon Musk's xAI acquisition portfolio (rather than being directly acquired by SpaceX), a transaction that itself reflects the strategic value of the AI programming tools space. When a company whose core product is AI itself starts adopting Chinese open-source models on the backend to reduce inference costs, the signal is especially powerful — even the most cutting-edge AI companies are acknowledging the practical value of Chinese models.

When even "AI-native companies" are running Chinese models behind the scenes, a consensus that "Chinese AI works just fine" is spreading across Silicon Valley.

The Self-Sovereignty Advantage of Open Weights

Beyond cost and performance, another key factor in the rise of Chinese models is "Open Weights." It's worth distinguishing between "open weights" and "fully open source" — two concepts that are easily confused. Full open source means publishing all training code, training datasets, data processing pipelines, model weights, and evaluation methods. Open weights primarily means publishing the model's parameter weight files, allowing users to download and run the model locally or on private clouds, without necessarily revealing training data or the complete training process. Currently, mainstream Chinese LLMs (such as DeepSeek, Qwen, Kimi-K2, etc.) mostly adopt the open weights model with commercially friendly licenses.

Open weights mean enterprises can self-host and self-manage models, which provides decisive advantages in data sovereignty, privacy compliance, and long-term operational costs. Specifically, enterprises can deploy models on their own servers or designated regional cloud infrastructure, ensuring business data never leaves their jurisdiction and meeting strict privacy regulations like GDPR (EU General Data Protection Regulation). Self-hosting also eliminates dependency risks on third-party APIs — business continuity won't be affected if a model provider adjusts pricing, changes terms, or faces geopolitical sanctions. Enterprises can also fine-tune models to more precisely fit specific business scenarios, which is typically restricted under closed-source API models. Companies don't need to entrust their core business to a closed-source API — they can truly maintain control.

From Kimi to DeepSeek to Qwen, Chinese open-source LLMs are penetrating global developers' daily workflows precisely through this combination punch of "low price + high performance + self-deployable."

A Victory for Pragmatism and the Open-Source Spirit

The expert concluded by noting that the abundance of training courses on the market teaching people to "make money with Claude Code" are essentially teaching people to "drive a Ferrari to the supermarket" — seemingly premium, but with questionable cost-effectiveness. Claude Code is a command-line AI programming tool from Anthropic that can understand entire codebases directly in the terminal environment, autonomously execute edits, run tests, and perform Git operations. Because it's based on flagship models like Claude Sonnet/Opus, API costs for a single complex programming session can run several to tens of dollars. For individual developers or small teams whose daily programming tasks don't involve complex architecture design or cutting-edge algorithm implementation, using open-source alternatives of equivalent capability but an order of magnitude cheaper (such as programming tools based on DeepSeek or Kimi) is often the more rational economic choice. The truly pragmatic approach is matching the right model to the task requirements.

Of course, the data and cases in this article come primarily from a single source (a Japanese expert's analysis reposted on Bilibili), and OpenRouter's statistics only reflect usage behavior on that platform — they cannot be entirely equated with real market share in the global AI market. It should be noted that OpenRouter's user base consists mainly of independent developers, small-to-medium startups, and tech enthusiasts. Large enterprises' AI consumption typically occurs through directly negotiated enterprise API contracts or self-built inference infrastructure, and wouldn't appear in OpenRouter's statistics. Therefore, the 58% figure more accurately reflects the usage preferences of the developer community and SMEs, not the full picture of the entire AI market. Even so, this trend remains profoundly significant — because today's developer community preferences often foreshadow the direction of the enterprise market one to two years out.

Regardless of the exact numbers, the trend of Chinese open-source AI rapidly penetrating global markets with extreme cost-effectiveness — even entering the core systems of American tech companies — is a reality that can no longer be ignored.

The essence of this disruption is not a victory for nationalist narratives, but a victory for pragmatism and the open-source spirit.