Our take
Phala is a privacy-first inference host that does not train on prompts and offers a zero data retention option. Its 28-model catalogue is small but includes the cheapest tracked offer in this sample.
Use this for privacy-sensitive workloads where zero retention and no prompt training are hard requirements. Pick it for budget inference on small models — the cheapest tracked offer in this sample sits here. Skip it if you need verified SOC 2, a confirmed EU endpoint, or predictable throughput across every model you might use.
- Strong privacy defaults: no prompt training and an optional zero retention setting.
- Cheapest entry point in the sample catalogue.
- Highest-throughput model in the sample runs more than nine times faster than the lowest.
- Compliance and corporate transparency documentation is thin: headquarters country undisclosed, no URL on file, SOC 2 and EU endpoint both unverified.
- Extreme throughput inconsistency across the catalogue — five of twelve sample offers fall below 50 tokens per second.
- One model throughput is unmeasured in our data.
Point your tools here
We have not recorded what a router needs for Phala yet — no base URL and no docs link on file. That is our gap, not a sign Phala has no API; its own documentation is the place to look until we close it.
Privacy & data handling
These answers cover requests sent to Phala through OpenRouter, as OpenRouter records them, last read 4 hours ago. For Phala's own API we hold no answer: check its terms. SOC 2 and the data processing agreement below are hand-checked, not part of that sync.
Prompt retention: none.