Why Are Big Companies Switching to Chinese AI Models?
Coinbase, Airbnb and DoorDash now run Chinese open-weight models like Kimi K2 for lower cost, data control, and no vendor lock-in.
What’s actually happening
A growing list of US companies, including Coinbase, Airbnb, DoorDash, Cursor, and the search engine Ecosia, now run Chinese open-weight AI models for parts of their stack instead of defaulting to OpenAI, Anthropic, or Google. This isn’t a handful of startups experimenting on the side. On OpenRouter, the marketplace developers use to route API calls across different model providers, Chinese models carried the majority of tokens sent by US companies as of mid-2025. The shift is driven by three practical business concerns: cost, data control, and the risk of depending entirely on one vendor.
TL;DR
- OpenRouter traffic has flipped: American models made up about 70% of OpenRouter volume roughly a year ago; that share has dropped to around 30%, with Chinese models taking all five top spots by token volume in July and carrying about 58% of tokens sent specifically by US companies.
- Download numbers tell the same story: Qwen reportedly passed 2 billion downloads on Hugging Face this year, compared to roughly 230 million for Meta’s Llama family, and over 150,000 derivative models have been fine-tuned from the Qwen base.
- Price gaps are large: Anthropic’s Claude models run roughly $2 to $10 per million input tokens and OpenAI’s flagship models can run $10 to $60 per million input tokens (more for fast modes), while Moonshot’s Kimi K2 family has been priced as low as $2 or less per million input tokens for comparable coding work.
- “Good enough” has arrived for most tasks: Most production AI workloads, like customer service bots, internal tools, or coding assistants, don’t require the single smartest model on the market, they need something fast, cheap, and reliable enough.
- Open weights mean data stays put: Running a model like Kimi K2 on your own infrastructure means your prompts, code, and internal data never have to leave your company’s servers, unlike hitting a closed API where the provider sees every query.
- Open weights also mean no one can pull the plug: Companies that self-host an open model aren’t exposed to a vendor changing pricing, deprecating a model, or shifting terms of service overnight.
- The practical move is architectural: Build the ability to swap models freely, test the most capable one first, then “downgrade” generation by generation until performance drops below what the task needs, and open-source models are increasingly where that line falls.
Auditable. Scalable.
Why is cost pushing companies toward Chinese models?
The price gap between frontier Western labs and leading Chinese model providers is large enough to change unit economics at scale. Anthropic’s Claude models are generally priced in the $2 to $10 per million input token range, and OpenAI’s top-tier models can run from $10 up to $60 per million input tokens in faster inference modes, before accounting for output tokens, which can cost far more. Moonshot AI’s Kimi K2 family, by contrast, has been offered at a fraction of that, with pricing as low as roughly $2 per million input tokens for its flagship-tier model and even lower for earlier versions optimized for coding.
For a company sending millions of requests a day, that difference compounds fast. Coinbase has reportedly cut its AI-related bill by roughly half by shifting workloads this way. The argument isn’t that Chinese models are dramatically smarter. Benchmark gaps between top Chinese models and frontier US models are often small. The argument is that most business tasks don’t need the absolute best model available, they need something fast and reliable enough to act as the execution layer underneath a broader system.
Why does data control matter so much to big companies?
Every time a company sends a prompt to a closed API like ChatGPT, Claude, or Gemini, that data passes through the provider’s servers. For a lot of use cases, that’s a non-issue. But for companies whose prompts contain proprietary processes, internal documentation, or unique business logic, it raises a real concern: the AI lab on the other end effectively sees the company’s intellectual property, and that same lab is also building its next generation of models.
Open-weight models flip this. A company can download the weights of a model like Kimi K2 or Qwen from Hugging Face, run it on its own GPU infrastructure, and fine-tune it on internal data, all without that data ever leaving the company’s own servers. It’s a point worth separating from the “Chinese AI” label: using an open-weight model developed in China doesn’t require routing any traffic through China. Once the weights are downloaded, inference can run entirely on infrastructure the company owns or rents, anywhere it chooses. That’s structurally different from using a hosted API, where usage necessarily leaves the boundaries of the company, regardless of which country the lab is based in.
Is this really about avoiding vendor lock-in?
Yes, and it may be the least talked-about reason. Companies that build critical workflows entirely on top of one closed provider are exposed if that provider changes pricing, deprecates a model, or alters terms of use. Lindy, an AI automation company mentioned as having moved its stack toward Chinese models, framed the decision bluntly: most business tasks, like drafting routine emails, don’t require the most powerful model that exists.
- Propellant load✓ Go
- Range safety✓ Go
- Weather✓ Go
- Final pollT–00:42
Open-weight models remove that dependency. Once a company has the weights, the infrastructure to run them, and the ability to fine-tune them, it isn’t beholden to any single vendor’s roadmap or pricing decisions. That matters more as AI becomes core infrastructure rather than an experimental add-on. A company running its own fine-tuned model on its own servers has something closer to a permanent asset. A company running entirely on someone else’s API has a dependency that can change under them without warning.
Is switching to a Chinese model worth it for most companies?
It depends heavily on scale and technical capacity. Self-hosting an open-weight model requires GPU infrastructure, engineering time to deploy and optimize it, and often fine-tuning work to get it performing well on a specific task. That’s a real upfront cost, which is why this trend shows up first at companies with the funds and technical teams to do it, not at small startups relying on a single API key.
For large companies running high-volume, repetitive workloads (customer support, internal tooling, code generation for well-scoped tasks) the math increasingly favors self-hosted open models. The suggested approach is iterative: start by testing the most capable model available for a given task, then step down generation by generation (both in Western and Chinese model families) until quality drops below what’s acceptable. Wherever that threshold lands, if an open-weight model clears it, that’s the one worth deploying permanently, since it comes with lower cost, full data control, and no dependency on an outside vendor.
Frequently Asked Questions
What are the most commonly used Chinese AI models in enterprise settings?
Qwen (from Alibaba) and the Kimi K2 family (from Moonshot AI) are among the most widely downloaded and deployed, with Qwen reportedly passing 2 billion downloads on Hugging Face and serving as the base for more than 150,000 fine-tuned derivative models.
Does using a Chinese open-weight model mean sending data to China?
Not necessarily. Open-weight models can be downloaded and run entirely on a company’s own servers or chosen cloud infrastructure. Data only leaves the company’s control if it’s sent to a hosted API, such as Kimi’s own chat service, rather than a self-hosted deployment.
Are Chinese models actually competitive with OpenAI and Anthropic’s best models?
On many benchmarks, top Chinese models run close to Western frontier models, though not always identical. For most production tasks, the gap matters less than cost and control, since many workloads don’t require the single most capable model available.
Why would a company choose an older or smaller model over the newest one?
Because most tasks (routine customer service replies, simple UI generation, internal automation) don’t need frontier-level intelligence. Using the smallest model that still meets a quality bar reduces cost and latency without hurting output.
What’s the business risk of relying entirely on one AI provider?
Pricing changes, model deprecations, and shifting terms of service are all outside a company’s control when it depends entirely on a closed API. Self-hosting an open-weight model removes that dependency, since the company owns the weights and the infrastructure running them.
