Chinese AI labs have shipped models that compete with the best on quality while undercutting them dramatically on price. That combination is hard to ignore, and it is also where a lot of teams get nervous. The honest position is that these models are worth understanding on the merits, and that the important decision is less about the model than about where your data goes when you use it.
They are genuinely capable, and cheap
Models from labs such as DeepSeek and Qwen are competitive on many everyday tasks, frequently at a fraction of the cost of premium Western APIs. Several are released as open weights, which means you are not limited to the maker's own hosted service to run them. On raw capability per dollar, they are among the strongest options available.
The real question is where your data goes
Using a Chinese provider's hosted API sends your prompts to that provider, under its terms and its jurisdiction. For regulated, confidential, or customer data, many organizations treat that as a non-starter, and that is a reasonable policy call. But the open-weight versions of these models can be run in your own environment or through a Western inference host, which changes the calculus entirely. The model and the route it travels are two different things.
Hosted API versus open weights
This distinction matters more than the brand on the model. A capable open-weight model reached over a US-based inference host is processed on that host's infrastructure, not the maker's, and inherits that host's data commitments. Governance should therefore be decided per route, not per country of origin. Judging a model purely by where it was trained misses the question that actually affects your risk: who processes the request.
Compliance and policy considerations
- Data residency and jurisdiction: which laws apply to data sent to a given route.
- Procurement and legal policy: some organizations have explicit rules on vendor origin.
- Sector requirements: regulated industries may constrain where data can be processed.
- Internal risk posture: many teams allow open-weight models self-hosted while blocking the hosted APIs.
A pragmatic approach
Evaluate these models on capability and cost like any other, then decide the hosting and routing that meets your data rules, and make it an explicit allow or block decision rather than something that happens by accident when an employee finds a cheap API. The goal is to capture the value where it is safe to and to close the door clearly where it is not.
How Switchboard helps
Switchboard surfaces each model's data residency, retention, training practices, and the route it is served over, and lets administrators allow or block models per policy. So you can take advantage of strong, low-cost models where your governance permits, keep sensitive workloads on approved routes, and make the whole thing an enforced, visible decision instead of a surprise on someone's expense report.