Summary

  • AS398037 is registered to NVIDIA Corporation under the name NVIDIA-NGC-NSV and announces fourteen IPv4 prefixes plus one IPv6 /48. It declares an open peering policy and holds 100 Gbps ports at AMS-IX and LINX LON1. That is a working interconnection presence, not a paperwork formality.
  • The same company's DGX Cloud documentation assigns physical compute, storage and networking to cloud service partners and NVIDIA Cloud Partners. The holder of the numbers and the operator of the building are not the same party at the layer where workloads actually run.

The registry has no interest in the product story

A number registry records who holds an address block and who announces it. It does not record what the announcements are for, and it does not care how large the market capitalisation is. That narrowness is what makes the record useful. Applied to a company whose network role is usually taken as self-evident, it splits one assumption into two separate facts.

The first fact is the autonomous system itself. AS398037 is held by NVIDIA Corporation, organisation identifier NC-1089, with the ASN-holder name NVIDIA-NGC-NSV. The association record shows registration on 14 October 2019, last updated 2 March 2021, status active and allocated under ARIN. The organisation record carries different dates — registered 26 September 2019, updated 25 November 2024 — because the two objects are maintained separately.

One attribution limit belongs beside those fields. ARIN's own RDAP endpoint returned a response type this review could not render, so the registration fields above come from a third-party reproduction of the ARIN record at bgp.tools rather than from a document read at the registry itself. That is a retrieval limitation about the source, not a doubt about the numbers.

The name on the record is the first quiet signal. NGC is NVIDIA's GPU-cloud software stack. The autonomous system is named for the software, not for a network product, and not for a region. A company that intended to sell connectivity would advertise that fact in the object that its peers look up before they agree to interconnect.

Fourteen announcements and a declared ceiling

Public routing observation shows fourteen IPv4 origin announcements: 8.44.54.0/24, 24.51.5.0/24, 24.51.6.0/23, 24.51.6.0/24, 24.51.7.0/24, then nine consecutive /24s from 198.100.160.0/24 through 198.100.165.0/24 and 198.100.171.0/24 through 198.100.173.0/24. IPv6 carries exactly one origin, 2604:1520:1000::/48.

The count needs a footnote rather than a flourish. Fourteen announcements are not fourteen distinct blocks: the 24.51.6.0/23 overlaps two of the listed /24s. Anyone who wants the address total has to do the subtraction themselves.

The peering directory completes the picture with a different kind of figure. The PeeringDB record for AS398037 lists the network as NVIDIA NGC NSV, organisation NVIDIA Corporation, also known as Nvidia NGC. It declares the network type Enterprise, geographic scope Global, and an Open peering policy with no ratio requirement, no contract requirement and no multiple-location requirement. It shows IPv4 prefixes 30, IPv6 prefixes 1, and a record timestamp of 22 April 2026.

Thirty against fourteen is not a contradiction, and the difference matters. PeeringDB's operator documentation defines those prefix fields as the maximum number of prefixes a peer should expect to see advertised. Thirty is a ceiling the network publishes for the convenience of people configuring session limits. Fourteen is what a collector actually saw. A network that tells prospective peers to expect no more than thirty prefixes is not describing a backbone.

What an open policy at 100 Gbps actually says

The interconnection detail is the part that resists the tidy corporate-connectivity reading. PeeringDB lists facilities in the United States, the Netherlands, the United Kingdom, Singapore and Japan, and names AMS-IX and LINX LON1 connections at 100 Gbps each.

An enterprise network that reaches five countries and buys 100-gigabit ports at two of Europe's densest exchange fabrics is not merely buying an office uplink. It is paying for the ability to hand traffic to many networks directly instead of pushing everything through purchased transit. Open peering with no ratio requirement is the policy of a network that wants interconnection to be easy — the opposite of a transit seller protecting a margin, and the opposite of a secretive network hiding its topology.

Those ports are still a conditional asset. A port is capacity waiting for sessions, and PeeringDB is a user-maintained directory: a declared open policy does not prove that any particular bilateral session exists. The retrieved page exposed no named exchange rows in its public-exchange section, which is a rendering limit, not evidence of absence.

The upstream picture is softer still. Observation suggests seven IPv4 upstreams — Arelion, Lumen, Telecom Italia Sparkle, GTT, Cogent, PCCW and Telstra — with only Arelion visible for IPv6. bgp.tools' own methodology page is unusually blunt about what that means: BGP does not encode business relationships, so upstream labels are inferred from observed paths through designated tier-one networks and certain route-server feeds, and a relationship can simply be missing when a collector lacks visibility.

Collector differences confirm the caution. A Hurricane Electric snapshot displayed as updated 18 August 2026 reports thirteen IPv4 origins, one IPv6 origin and ten observed peers, including AS286 and AS1000 that do not appear in the bgp.tools table. Different visibility, different day, different list. No single page is the whole truth about who this network talks to.

The vendor's own documents move the racks

Now set the routing record beside the service record, and the split becomes explicit.

At launch on 21 March 2023, NVIDIA described DGX Cloud as monthly rentals of dedicated clusters, hosted by cloud service providers and beginning with OCI. Azure hosting and expansion to Google Cloud were described then as future plans.

The current DGX Cloud product page opens differently: it presents the service as NVIDIA's internal environment for developing models, validating architectures and running production workloads across CSPs and NVIDIA Cloud Partners, while still carrying marketplace and private-offer references for AWS, Google Cloud, Microsoft Azure and OCI. The page does not say whether those commercial sections remain operative. Both readings of that page are inferences; only the text is a fact.

The managed-service documentation is unambiguous about responsibility. Run:ai on DGX Cloud, last updated 27 February 2026, states that subscriptions use GPU capacity supplied by NVIDIA CSP partners, and that NVIDIA manages cluster components, monitors CSP-hosted compute, storage and networking, administers Kubernetes and coordinates maintenance with the CSP.

The requirements document is blunter still. NVIDIA Requirements for AI Clouds, version 2.4, dated 1 September 2026, defines an NVIDIA Cloud Partner as the cloud provider responsible for delivering and operating applicable compute, networking, storage and managed-service capabilities. It also treats DGX Cloud clusters in partner environments, NVIDIA CorpIT and NVIDIA on-premises object storage as separate estates joined by private interconnect requirements — an architecture of linked but distinct operators.

Responsibility even varies by offering. The DGX Cloud Slurm cluster administration guide, last updated 17 January 2026, states that NVIDIA provisions and configures the cluster while management and operation of the cluster is the customer's responsibility, with NVIDIA managing the CSP account, the infrastructure resources and the security controls that separate tenants' compute, network and storage.

The parts are not the network

The last piece is the networking portfolio, and it is the one most often mistaken for the answer. Spectrum-X combines Ethernet switches and SuperNICs with adaptive routing that steers large flows around congestion, and telemetry-driven congestion control; Spectrum-XGS extends the platform across distributed data centres. Every one of those capabilities is a property of equipment and software a customer installs and operates. Selling the switches, the adapters and the congestion-control logic is not the same as owning the fabric they run on.

Put the layers in order and the picture is coherent. NVIDIA holds the numbers, the routing identity and a genuine global interconnection footprint. It administers a software plane — Kubernetes, tenant separation, the partner requirements document — that reaches into facilities it does not own. Its partners deliver and operate the physical compute, storage and networking. Its customers, in at least one documented configuration, operate the cluster themselves.

That is why the usual phrase is misleading. A vendor whose silicon carries AI east-west traffic is an infrastructure company in the sense that matters commercially, but the autonomous system is not a crown jewel proving it operates a backbone. It is a modest, openly peered, brand-named network: the address layer of a company whose real leverage is the design and the software plane, not the racks.

What the record does and does not settle

Four limits belong on the table. ARIN's own registration document was not readable in this review, so the AS fields rest on a third-party reproduction. Whether the retained AWS, Google Cloud, Azure and OCI text on the DGX Cloud page describes live offerings or legacy material is not established by that page. PeeringDB's open policy declaration does not prove any specific bilateral session exists. Upstream identities are inferred from routing paths, not read from contracts. Per-partner hosting share and current purchasing eligibility were not verified here.

None of those gaps touches the central finding, which needs only two documents that were read in full: a registry reproduction that names NVIDIA as the holder of AS398037, and NVIDIA's own service documentation that names partners as the operators of the compute, storage and networking. The addresses and the racks have different owners because they are different things.