Western Companies Turn to Chinese AI Models to Cut Costs
Claire Weston
DoorDash, Siemens, and Airbnb are migrating workloads to Chinese AI models like DeepSeek, Z.ai, and Kimi K2.6 — driven by cost savings of 10× to 60× and a growing need to reduce dependence on U.S. frontier labs.
Which major companies have already switched?
DoorDash co-founder Andy Fang said last week the company now routes "low-level work" to Kimi K2.6, keeping Anthropic's Fable only for "the hardest work." The new mix "massively outperforms" the previous all-American setup at lower cost.
San Francisco startup Lindy has fully switched from Anthropic to DeepSeek V4. Founder Flo Crivello called the move "transformational," saving the company millions of dollars while improving performance across multiple core use cases.
Siemens runs DeepSeek, Z.ai, U.S. frontier labs, Nvidia, and France's Mistral simultaneously, citing a need for "flexibility." Airbnb uses "a small number of Chinese-origin models," routed only through "U.S.-approved providers" to protect data security.
How much cheaper — and why?
Together AI CEO Vipul Ved Prakash says the best open-weight models — models whose parameters are publicly released, letting firms deploy and fine-tune them on their own servers — are 10× to 60× cheaper than comparable proprietary models.
This means → companies aren't just saving on subscription fees — they're reclaiming data control, running models on their own infrastructure instead of handing data to a third party.
At the same time, Anthropic and OpenAI have shifted some enterprise services from flat subscriptions to usage-based billing, pushing costs higher. In plain terms = U.S. models are getting more expensive while Chinese models are getting cheaper — the gap is widening fast.
Can Chinese models keep up on capability?
OpenRouter data shows that token consumption for DeepSeek, Z.ai, and other Chinese AI models has rapidly overtaken U.S. rivals this year. Featherless AI CEO Eugene Cheah calls Chinese models "the elephant in the room."
Z.ai's GLM-5.2, released in June, drew wide praise from Silicon Valley. Andreessen Horowitz co-founder Marc Andreessen wrote on X that GLM-5.2 is "the first Chinese AI model to match or exceed the open models from major U.S. labs without compromise."
This reflects a tipping point: companies are realizing they "don't need the best model — they can use a faster, cheaper one." For the bulk of everyday workloads, Chinese models are already good enough.
Why are European firms even more aggressive?
The Trump administration imposed export controls on Anthropic's Mythos and Fable models last month. The ban was later reversed, but Newton chief AI officer Ben Grinnell says the episode "permanently changed market perception" — "the genie is out of the bottle, and it can't be put back."
This means → for European companies, relying on U.S. models has itself become a policy risk. RTP Global U.S. vice president Tom Sheridan says his advice to European startups has changed: "Self-hosted Chinese models are a safer bet than U.S. models."
German HR startup Timebutler began migrating some tasks from Anthropic's Claude to Alibaba's Qwen model about six months ago, aiming to reduce dependence on U.S. frontier labs.
What does this mean for U.S. AI vendors?
Georgetown CSET researcher Sam Bresnick frames it bluntly: "Why pay a premium for Anthropic or OpenAI when Chinese models are good enough for a large share of workloads?"
In plain terms = the challenge facing U.S. frontier labs isn't a technology gap — it's pricing-power erosion. When a "good enough" alternative costs tens of times less, premium pricing becomes very hard to sustain.
If this trend continues, it will put sustained pressure on U.S. frontier model vendors' enterprise revenue.
Content is for reference only, not financial advice.