Summary

  • Puzl publishes a cloud-computing service, resource prices and a dedicated NVIDIA A100 offer. Its company number connects those products to the business recorded in network registries as GPU Computing OU.
  • Usage-based cloud resources and long-term GPU reservations make different bargains with the customer. Network observations identify part of the operating footprint; they do not measure available computing capacity.

The lowest hourly price on Puzl's GPU page has a three-year condition attached. At the time of this review, the company advertised one NVIDIA A100 with 40GB of memory, 24 virtual CPUs and 128GB of system memory for $0.82 an hour with a three-year commitment. The same displayed configuration cost $1.12 an hour with a six-month commitment. The attractive number is therefore also a question about how long the customer expects to need the machine.

That GPU offer sits beside a rather different pitch elsewhere on the website: stop paying for computers when they are idle. Puzl describes cloud instances whose CPU and memory charges follow actual resource use, rather than simply the amount reserved. One offer sells a longer reservation; the other sells a different meter. Conflating them would obscure what a customer is buying.

The build finishes; what happens to the bill?

Puzl's main service page puts software jobs near the front of its business. RunJob is advertised for GitHub Actions and GitLab CI, with management through a dashboard, an API or Terraform. Those are familiar places for a development team to spend money without thinking of itself as a buyer of cloud infrastructure: compiling software, running tests and waiting for another job to begin.

The company's resource-pricing page describes Spike Instances as KVM-based virtual machines that scale CPU and memory for changing workloads. CPU use is measured in CPU-seconds; memory use is measured separately in GB-seconds. Persistent storage has its own rate. Puzl's proposition is that a customer need not keep paying the same compute charge through the quiet periods between bursts of work.

This is a description of the product's billing model, not a measured saving. A useful comparison would require the same job, its completion time, resource consumption and the resulting bill. The published page also mixes euro symbols on individual rates with a general USD/EUR explanation. Those details need resolving in a quote rather than being turned into a precise cost comparison here.

The dedicated GPU page addresses a different need: reserving an accelerator configuration over a stated period. It advertises EU infrastructure, on-premises or private-cloud options and Kubernetes-ready AI clusters. Additional dedicated NVMe storage carries an extra charge. These are concrete published offers, but they are not an independently checked inventory of machines or a guarantee of stock on a particular date. The usage-based promise for Spike Instances cannot simply be carried across to a committed GPU reservation.

The company behind the service

Puzl's paperwork supplies a more useful connection than the word “GPU” in a network holder's name. The Estonian business register lists Puzl Cloud OÜ under number 14476114 and records GPU Computing OÜ as a former name. The website footer uses Puzl Cloud OÜ and that same number. The customer agreement still defines Puzl Cloud using GPU Computing OÜ, again with number 14476114.

The network records retain the older name too. RIPEstat identifies AS57063 as PUZL-AS GPU Computing OU. The matching registration number connects the company and the service without requiring a story about an acquisition or an undisclosed parent. Neither such a transaction nor a parent relationship has been established here.

There is an important operational detail in the customer agreement. It allows a customer to choose a location for customer data. Where that choice is unavailable or not made, the terms allow storage and processing wherever Puzl maintains or uses facilities. Account information is treated separately in that clause. An EU location on a GPU product page is consequently not a complete description of where every category of a customer's information will go; the selected service and its terms matter.

A network footprint, with a corrected neighbour

The RIPEstat announced-prefix result captured for 10–24 October 2026 showed two prefixes: 213.5.72.0/24 and 2001:67c:f74::/48. This is a dated observation from the routing system, not a count of servers. The endpoint's documentation also explains that its default visibility threshold excludes announcements seen by fewer than ten RIS peers. “Two in this view” is more precise than “only two anywhere”.

The routing-consistency view, with a query time of 24 October at 08:00 UTC, observed AS15830 as a neighbouring network. The holder record for AS15830 identifies Equinix (EMEA) Acquisition Enterprises B.V., not RIPE NCC. Earlier coverage on this page got that identification wrong.

The same consistency view distinguished policy-listed AS50794 and AS200675 from the neighbours it observed in BGP. That difference can prompt a network operator to inspect routes and declarations. It does not establish a hosting contract, the location of GPUs or a failed interconnection. Nor does a small prefix count classify a business as an internal corporate network rather than a cloud supplier. An address table has no column for accelerators delivered to customers.

Puzl is therefore not merely a suggestive registered name. It has published products and commercial terms that can be examined on their own merits. The useful next evidence would connect those terms to a workload: which configuration was supplied, when it became available, how the job ran and what was charged. Those answers would say considerably more about the service than another attempt to count computers through its IP addresses.

Correction and update, 24 October 2026: An earlier version incorrectly identified AS15830 as RIPE NCC's network and overlooked Puzl's published service, pricing and contract pages. This report corrects those points and removes unsupported inferences about computing capacity and hosting from prefix counts. The original article URL and publication date are retained.