Summary

  • GPULOL Inc., registered in Toronto, holds the 32-bit network number AS402324 and a direct IPv6 allocation, 2602:81D:7000::/44, from which it announces three /48 blocks — and it originates no IPv4 prefixes at all.
  • The numbering record is complete enough to audit and thin enough to say almost nothing about the business behind it; that gap, not the address count, is what matters to anyone who reads registries as a proxy for substance.

An autonomous system with zero IPv4 addresses used to be an eccentricity. It is now a plausible way to start. GPULOL Inc. is one of the companies that started that way, and its public footprint is unusual less for what it contains than for how cleanly it separates two things that used to arrive together: the right to be a network, and evidence that you run one.

What the records actually say

Start with the entity. The registry view published at bgp.tools' AS402324 page renders the organisation handle GPULO, the name GPULOL Inc., and an address of 18 King Street East, Suite 1400, Toronto, Ontario, M5C 1C4, Canada. The organisation registration date on that record is 9 March 2026, updated 20 March 2026. The autonomous system sharing that handle — an autonomous system number, or ASN, is the identifier a network uses to announce routes into the global routing table — carries the registration date 25 March 2026, with network status given as "Active, Allocated under ARIN."

RIPEstat's as-overview record for AS402324 agrees on the holder, naming it "GPULOL - GPULOL," and returns announced: true. It places the number inside block 401309-402332, described as "Assigned by ARIN" and named the IANA 32-bit Autonomous System (AS) Numbers Registry, over a query window ending 08:00 UTC on 24 October 2026.

Then the address space, which is where the picture sharpens. The same registry view reports prefixes originated as 0 IPv4 and 3 IPv6, and tags the network "IPv6 only." The three are 2602:81d:7000::/48, 2602:81d:7001::/48 and 2602:81d:7003::/48. A /48 is a block of IPv6 space large enough to number 65,536 separate subnets; it is not a token allocation. RIPEstat's announced-prefixes feed lists exactly those three and nothing else, with timelines running from 10 October 2026 to 24 October 2026 — the window over which the observation holds.

The parent block is the part that changes how the record should be read. A whois lookup on 2602:81d:7000::/48 returns the enclosing range 2602:81D:7000:: through 2602:81D:700F:FFFF:FFFF:FFFF:FFFF:FFFF, registered as CIDR 2602:81D:7000::/44, with the network type "Direct Allocation" and the organisation "GPULOL Inc. (GPULO)." A /44 holds 16 of those /48s, roughly a million /64 subnetworks. In plain terms: this is not address space carved out of a hosting provider's allocation and lent onward. GPULOL holds the block itself, from the registry, in its own name.

Reachability is supplied from elsewhere. The bgp.tools page and its accompanying summary report transit through The Constant Company, LLC, AS20473, which is how a small network with its own address space still reaches the rest of the internet: it owns its numbering and rents its path. That attribution is the page's rendering; it is not a fact I could confirm from the registry itself.

Why "no IPv4" is a decision, not an accident

An IPv4 address is a scarce, tradeable asset with a price attached. An IPv6 allocation of this size is neither scarce nor, in any practical sense, contested. For a new entrant in 2026 the arithmetic is stark: the routing layer can be entered with IPv6 alone, and the expensive half of the problem — getting IPv4 space — can be deferred, leased, or solved with translation at the edge.

That is a structural inference from the record, not a report of GPULOL's reasoning. What the record does show is the consequence: every prefix this network originates is IPv6. Traffic from an IPv4-only endpoint cannot reach it directly. Serving those users requires either address translation on the way in or a provider's IPv4 addresses in front of the service — a detail that lives in operations, not in registries, and therefore does not appear in anything I could read.

Where the evidence stops

ARIN's own machine-readable endpoints are the authoritative source for all of the above, and I could not read them. Direct requests to the RDAP autnum record for AS402324 and to the RDAP entity record for GPULO returned no readable body across repeated attempts on 24 October 2026. That is an access limit, not evidence of absence, and it changes what this article is allowed to claim: the registration dates, the handle and the Toronto address are attributed to the documents that did return text, which render registry data rather than reproduce ARIN's own field list.

The harder gap is the company. No source I obtained states what GPULOL Inc. sells, who its customers are, whether it operates its own equipment, or where that equipment sits. BTW's directory entry for GPULOL Inc. Only for Strategic Circle records the resource-holder footprint and nothing more: no recorded domains, no recorded websites, no IXP presence, no facilities. The absence of a website in a directory is not proof that none exists. It does mean that anyone who wants to know what this network is for has to ask it directly.

One number moves. That directory snapshot, taken 21 October 2026, counted eight announced IPv6 /48s. Three days later, routing data shows three. The useful conclusion is procedural rather than dramatic: the count depends on which instrument you consult and when, so anyone quoting it should name the instrument and the timestamp. I found no basis for explaining the difference, and I am not going to invent one.

One more field resists a verdict. Route origin validation — RPKI, the cryptographic proof that a network is authorised to announce a given prefix — is recorded in the directory as zero validated and eight unknown. Nothing I read confirms that these prefixes carry valid origin attestations, and nothing contradicts it either. For a network whose entire claim to existence is three announcements, that is the single most consequential unknown in the file.

What this record is good for

Everything above is available to anyone with a browser, in six fields and three prefixes, and it is all true. That is the point. Registry identity has become cheap, fast and legible: a company registered in March can hold a direct /44 and appear in three independent datasets by October. What has not become legible is the part that used to be inferred from the same records — whether there are racks, staff, customers and revenue behind the announcements. The registry used to be a rough proxy for all of it. For a network like this one, it is now a proxy for none of it.