Summary

  • The APNIC aut-num object for AS17494 declares import/export relationships only with AS3491, AS7374, AS5400, AS702 and AS6762, and was last modified on 18 January 2021 — a policy document frozen for more than five years.
  • Observed routing in 2026 tells a different story: paths via AS2914 (NTT), AS6453 (Tata Communications America) and AS9498 (Bharti Airtel); upstream inventories that diverge by observer (from 6 to roughly 68–70); a customer as-set enumerating hundreds of ASNs — including the reserved AS23456 — against 58 to 63 networks observed downstream; and self-reported capacity figures that measurement contradicts.

A policy frozen since January 2021

The first routing document any peering engineer consults is the APNIC aut-num object. For AS17494, it carries the name BTTB-AS-AP and declares import/export relationships only with AS3491 (PCCW Global), AS7374, AS5400, AS702 and AS6762 (Telecom Italia Sparkle), with a last-modified date of 18 January 2021 https://bgp.he.net/AS17494. This field is self-filed: it describes what the operator wrote, not what the network does. The distinction matters because IRR-based import filters are built precisely on this document.

Recall the identity frame established by our earlier coverage: the entity operating AS17494 is named by a free-text description line in the registry, the organisational object points to the Bangladesh Telegraph & Telephone Board, and PeeringDB attributes the network to BTCL https://btw.media/en/as17494-two-ledgers-registration-versus-operation. This article does not re-argue that question; it measures a different divergence, between declared policy and observed use.

What the BGP table shows in 2026

Observed paths for prefixes originated by AS17494 traverse AS2914 (NTT), AS6453 (Tata Communications America) and AS9498 (Bharti Airtel) — for example "2914 17494" and "7018 6453 17494" — with 81,664 addresses originated and 148,992 addresses in transit, according to CIDR Report https://www.cidr-report.org/cgi-bin/as-report?as=AS17494&view=6447. None of those three operators appears in the declared import/export list: AS2914, AS6453 and AS9498 are all absent from the 2021 document.

The upstream inventory then depends on each observer's methodology, and it should be attributed source by source rather than reconciled into a single number. bgp.tools lists 7 upstreams — AS9498, AS6453, AS174 (Cogent), AS6762, AS2914, AS3491 (PCCW) and AS1299 (Arelion) — with 354 peers, 58 downstreams and a customer cone of 278 https://bgp.tools/as/17494. IPinfo retains 6, including AS6461 (Zayo), absent from the other lists, and 50 peers https://ipinfo.io/AS17494. IP.CC counts 8, adding AS6939 (Hurricane Electric) https://ip.cc/topic/asn/AS17494/. records.ping.pe, by contrast, reports roughly 68 to 70 upstreams, 116 peers and 62 to 63 downstreams, with AS9498 the most frequently observed upstream https://records.ping.pe/peers/AS17494. The spread between 7 and 70 upstreams is irreconcilable without knowing the sampling methods; that is exactly the point. BGPView, another window on the same network, shows 100 Gbps ports at AMS-IX and Equinix Singapore https://bgpview.io/asn/17494.

An oversized customer as-set

The APNIC IRR object "AS17494:AS-CUSTOMERS", described as "BTCL's AS-SET for its Customers" with country BD, enumerates hundreds of member ASNs — beginning with AS17494, AS17469, AS153392, AS137934, AS63139, AS137453 — including the reserved AS23456, an AS_TRANS placeholder rather than a real network https://bgp.he.net/irr/as-set/AS17494:AS-CUSTOMERS. PeeringDB marks this as-set as verified https://www.peeringdb.com/asn/17494. Yet trackers observe 58 downstreams according to bgp.tools and 62 to 63 according to ping.pe. The declared list exceeds observed use by an order of magnitude: whether it reflects historic customers, sub-allocations or data noise, it cannot be verified from the outside, and a counterparty that builds an acceptance filter from this as-set absorbs a routing-hygiene risk the declaration does not price.

Self-reported capacity against measurement

PeeringDB self-reports traffic of 500–1000 Gbps, 500,000 IPv4 and 50,000 IPv6 prefixes, a "Heavy Inbound" ratio, and four exchange points — AMS-IX, DE-CIX Mumbai, Equinix Singapore, LINX LON1 — with the network record last updated on 26 January 2023 and public peering information on 24 April 2026 https://www.peeringdb.com/net/13897. Measurement tells another story: bgp.he.net counts 49 originated prefixes (36 IPv4, 13 IPv6) against 729 announced prefixes (716 IPv4, 13 IPv6), 388 observed BGP peers, and an RPKI status of 37 valid against 12 invalid, plus an "announces bogons" flag https://bgp.he.net/AS17494. Port speeds diverge too: IP.CC concurs on 100 Gbps at Equinix Singapore but differs elsewhere, and it adds interface addresses consistent across trackers (80.249.211.151, 103.27.171.154, 27.111.228.145, 195.66.226.78) https://ip.cc/topic/asn/AS17494/, while PeeringDB shows 10G or a blank field https://www.peeringdb.com/asn/17494.

A freshness anchor: WorldIP provides a per-prefix snapshot last updated on 25 September 2026 at 02:09 UTC, observed via RIPE RIS collectors, including 203.112.216.0/24, 103.110.215.0/24, 114.130.128.0/18, 123.49.0.0/18, 180.211.128.0/17 and 209.58.24.0/24 — the last being an ARIN-range block for an APNIC-registered network, an anomaly the sources do not resolve https://worldip.io/asn/17494.

What all of this proves is bounded: the declaration layer and the operational layer of AS17494 describe different networks. None of these observations establishes misconduct — documentation staleness is not evidence of wrongdoing. But for a peer, a transit customer or a due-diligence team, each gap has a cost: mis-calibrated IRR filters, an unusable peer count, unverifiable capacity.