Cheaper Chinese AI models are reshaping enterprise AI strategy

Enterprise AI spending is pushing US companies toward cheaper Chinese AI models from DeepSeek and Z.ai, as firms like DoorDash, Airbnb, and Siemens adopt open-weight alternatives to cut costs and reduce vendor lock-in. The shift is driven by competitive performance at significantly lower prices, with OpenRouter data showing Chinese models overtaking US equivalents in usage.

What happened

OpenRouter, a platform that aggregates access to major AI models, reports that leading Chinese models from DeepSeek and Z.ai have overtaken Anthropic's Claude and OpenAI's ChatGPT in usage among US enterprises. The release of GLM-5.2 from Z.ai caused a stir in Silicon Valley, with figures hailing it as nearly as capable as US systems but far cheaper.

Cost pressure is a major factor. One organization reportedly spent $500 million in a single month on Claude usage, while the Ramp AI Index found that AI-heavy businesses spend around $7,500 per employee per month on AI. Companies like DoorDash are now using Moonshot AI for lower-level work, and the San Francisco startup Lindy completely replaced Anthropic's tools with DeepSeek's V4 models.

Beyond cost, many Chinese models are released as open-weight, meaning their parameters are fully visible to users. This allows enterprises to self-host, fine-tune with proprietary data, and maintain full control over data processing. The approach also mitigates single-vendor risk, a concern highlighted after the Trump administration suspended access to Anthropic's Mythos model overseas.

Why AI builders should care

For product teams running AI-powered services, the rise of cheaper, open-weight Chinese models changes the economics of deployment. Cost-sensitive workloads that previously required expensive US models can now be shifted to alternatives without sacrificing acceptable performance. The ability to self-host also reduces per-token costs and provides data residency control.

Builders should design their AI stacks with interoperability in mind. Supporting multiple backends including DeepSeek, Z.ai, and Moonshot AI allows dynamic routing based on cost, latency, and compliance requirements. The trend also accelerates the need for toolchains that handle model switching, governance, and monitoring across different providers.

Practical implications

Start by running pilot comparisons between US and Chinese models for your specific workload. Measure total cost of ownership including API fees, self-hosting infrastructure, and fine-tuning overhead. For tasks like classification, summarization, or customer support, Chinese models may offer 60-90% cost savings with acceptable accuracy.

Consider open-weight models for sensitive data pipelines. Self-hosting a model like GLM-5.2 or DeepSeek V4 keeps data within your infrastructure and gives you full visibility into model behavior. Evaluate readiness to fine-tune with proprietary datasets to match domain-specific requirements.

Caveats

Performance varies by model and use case. While Chinese models are closing the gap, US models still lead on certain benchmarks and complex reasoning tasks. The cost advantage may also be influenced by state subsidies, which could shift with regulatory changes.

Open-weight models introduce their own risks: you must manage model security, licensing, and potential compliance issues with data handling. Geopolitical uncertainties, such as potential export controls from China, could disrupt access. Always test thoroughly before committing production traffic.