September 4, 2026
“The AI Is Slow” Is Usually a Routing Ticket
A model that answers in 800 ms from a phone and 8 seconds from the office is rarely a different brain. It is a different path to the same prefix: a resolver in another country, a proxy that exited on another continent, a peer that prefers a cheap detour. Treat “AI is slow” like “the website is slow.” Name the prefix. Time the path. Then argue about tokens.
Unbundle the feeling
Users mash four clocks into one complaint: DNS time, TCP/TLS setup, time to first token, and tokens per second. The last two can be the vendor. The first two are yours until proven otherwise. Resolve the API name. Look up the A and AAAA on IP Lookup. If the origin ASN is a CDN or the vendor’s announced infrastructure, you are on the public front door. If it is a random VPS ASN, you are not talking to who you think.
Anycast will gaslight you
Many AI fronts use anycast or geo-DNS. A global ping that is 12 ms from one city and 180 ms from another is a map of PoPs, not a sick rack. Traceroute from the angry office should show the detour. If hop 4 is a peer you do not use from home, you found the ticket.
The office extras that murder TTFT
- SSL inspection that does not like HTTP/2 or long-lived streams.
- Split-tunnel VPN that sends AI to the internet and DNS to HQ.
- A “secure DNS” app on the laptop pointing at a resolver 4,000 km away.
- IPv6 enabled on the LAN, broken on the firewall, happy AAAA on the vendor.
Probe both families. A broken v6 path with a published AAAA is the 2026 version of “it works in incognito.” Incognito does not disable AAAA.
A workflow that ends arguments
- Capture the resolved address at the moment of slowness.
- Prefix, origin, RPKI on BGP Lookup and the validator.
- Ping and traceroute from the angry network and from a clean network.
- Looking glass on that prefix from two regions.
If those four agree and first-token time is still awful, then you may email the vendor. If they disagree, you have a path story. Resizing the prompt will not fix a peer that prefers cheap latency.
Resolve the API, then traceroute it from the slow site. Screenshot the origin ASN next to the hops.
FAQ
Can the model be the problem?
Yes. After the path is clean. Measure time-to-first-byte on a tiny prompt from two networks before you blame weights.
Does a VPN always help?
Only if the concentrator sits closer to a good peer of the vendor. A VPN in the wrong city is a self-inflicted hairpin.