Chinese open-weight models opened up new possibilities. Before using them for real workloads, I wanted to understand exactly who would handle the data—and under what terms.
Explore the reportTrust follows
the data.
A model’s origin and the location of the service running it are different questions. An open-weight model can be hosted by another company; a seemingly local service can also forward requests to an upstream API.
I wanted to evaluate the actual inference path: the provider, its infrastructure, its contractual commitments, and where a prompt might end up.
I used an AI-assisted research workflow with four lanes: corporate identity and ownership, product and infrastructure, privacy and data flow, and reputation. Each provider received a profile built from public sources.
The brief combined Canadian privacy considerations with a specific restriction on China-bound data flows. The resulting assessments distinguished between a clear fit, a fit with conditions, and a provider that did not meet the brief.
The July review classified 35 providers as GO, 27 as conditional GO, 8 as conditional NO-GO, and 6 as NO-GO. Conditions included contractual terms, retention settings, region selection, and unresolved service availability.
The useful output was the reasoning behind those labels: what was known, what remained uncertain, and what would have to change before a workload could be considered.
Read the July 2026 deliverable