Static ISP exits for long-session evals and authenticated RAG

Why ISP exists between residential and datacenter

ISP proxies are datacenter-hosted IPs that have been announced under a real residential ASN. You get the persistence and latency of datacenter with the ASN reputation of residential. The trade-off: no diversity — you have a specific, stable set of IPs assigned to your account.

For AI workloads, ISP is specifically useful when:

• A multi-turn agent evaluation depends on the same IP across a long conversation, because mid-session IP rotation breaks session cookies, CSRF tokens, and some chat-history features on commercial APIs
• A RAG ingestion pipeline logs into a knowledge source (with permission) and the source's rate limiter is per-IP
• You need sub-50ms latency to a US or EU eval target but the target ASN-classifies AWS subnets and downgrades you

Where ISP is wrong

• SquadProxy diversity. A training-corpus workload sampling broadly from the open web does not benefit from ISP's persistence. Use residential or datacenter instead.
• Mass parallel evaluation. Running 50k concurrent eval requests from 100 ISP IPs means 500 requests per IP — which is not residential-shaped behaviour even though the IP has a residential ASN. Mix with rotating residential above scale.

Regions and ASN mapping

• US — Comcast, Spectrum, AT&T, Verizon.
• UK — BT, Sky, Virgin Media, TalkTalk.
• Germany — Deutsche Telekom, Vodafone DE, 1&1.
• France — Orange, Free, SFR, Bouygues.
• Japan — NTT, KDDI, SoftBank.
• Singapore — Singtel, StarHub, M1.

Korea, Canada, Netherlands, and Australia ISP allocations available on request but are smaller pools. Ask before quoting.

Provenance

ISP announcements are done through upstream partnerships with carriers that lease residential IP space. Every allocation has a RIR-visible ownership chain ending at the carrier. For customers publishing model cards or data cards that have to disclose source infrastructure, this is documentable.