4G mobile exits for cellular-anchored evaluation

Why 4G matters for AI — and why it usually doesn't

4G mobile is premium infrastructure. Per-gigabyte it is 5–20× the cost of datacenter. For AI data collection, most of the volume does not justify that premium. Mobile is the right tool in a narrow set of cases where IP-layer authenticity and carrier/mobile fingerprinting matter.

Cases where 4G is correct

• Cellular content policy on model APIs. Some commercial LLM providers apply different content policy to requests that originate from mobile carrier ASNs in specific regions — not dramatically, but measurably. If your eval is measuring policy boundaries, you want a sample from carrier exits.
• Mobile-first rendering paths. Some news, social, and review sites serve materially different HTML to mobile UA + mobile ASN combinations. For RAG ingestion that wants the mobile-rendered variant, only a mobile exit gets you there reliably.
• Regional carrier variation. In some regions — notably APAC and parts of the Middle East — cellular networks peer differently to the open internet than residential broadband. A model evaluation that matters for a mobile-first deployment market wants cellular-origin traffic in the eval mix.

Cases where 4G is wrong

• Training corpus collection. Bandwidth economics alone disqualify mobile for TB-scale work. Use datacenter.
• Latency-sensitive eval loops. 4G adds 50–150ms over datacenter and jitter that breaks timing-sensitive tests.
• Multi-turn agent sessions. Mobile IPs rotate aggressively on CG-NAT. Sticky sessions help but top out at 30 minutes and sometimes fail mid-session as the SIM shifts towers.

Pool shape

• Live SIMs. We operate carrier-leased SIMs in managed multi-SIM modem arrays across our 10 focus countries. Not SDK-sourced mobile proxies — those have reliability and provenance issues that make them unsuitable for eval methodology.
• Carrier targeting. Tier-1 carrier per country: AT&T, Verizon, T-Mobile in the US; EE, Vodafone, Three, O2 in the UK; Deutsche Telekom, Vodafone DE, O2 DE in Germany; etc.
• Tower distribution. We distribute SIMs across metro and non-metro tower coverage so the IP diversity looks plausible across a campaign.

Session semantics

4G CG-NAT means multiple sessions share an IP at any given moment. Per-request rotation at the gateway level gives you a new IP reliably; sticky sessions bind you to a specific SIM for 1–30 minutes. Cycling SIMs (rather than IPs) is how we rotate without breaking cellular context.