Summary

  • CoreWeave's public number-resource record resolves to a single autonomous system, AS33425, at the same time as the company discloses revenue, backlog and capital spending in the billions of dollars.
  • No public document retrieved for this article states CoreWeave's total IPv4 or IPv6 holdings, and no readable per-prefix registry record was obtained. That gap between financial disclosure and registry disclosure is the story.

A company that has signed $104.2 billion of backlog can afford to be precise. CoreWeave is precise about money: revenue of $2.575 billion in the second quarter of 2026, a 112.3% increase; backlog of $104.2 billion at the end of that quarter; 2026 capital expenditure guidance raised to a range of $35 billion to $39 billion; interest expense climbing to $640 million from $267 million. Those figures come from a 27 October 2026 market analysis citing the reported results (247wallst, 27 October 2026). A companion account of the autumn AI-sector selloff adds 2025 annual revenue of $5.1 billion, a net loss of $1.2 billion, and roughly $25 billion of new commitments booked early in the third quarter of 2026 on top of that backlog (Economic Times).

The same account records that Nvidia agreed to buy unsold CoreWeave compute capacity through April 2032 (Economic Times). Read those two items together and you have the shape of the business: enormous contracted demand, enormous capital intensity, and a financier's dependency on a single silicon supplier's willingness to keep absorbing capacity.

Now ask the internet's own records what CoreWeave is.

One number, four seconds

An autonomous system is the unit the internet routes with. It is a network, or a group of networks under one operator, that presents a single routing policy to everyone else and is identified by an autonomous system number, an ASN. When a carrier or a data centre hands traffic toward CoreWeave, that is the handle on the label.

Public peering records give CoreWeave one: AS33425, listed as CoreWeave's network in the PeeringDB entry for that number (PeeringDB net/33425) and in Hurricane Electric's BGP toolkit record for the same ASN (bgp.he.net/AS33425). CoreWeave is a United States company, which puts its American number resources under ARIN, the regional internet registry for the United States, Canada and a swathe of the Caribbean (ARIN).

That is the documented floor. One ASN, one registry of authority, one jurisdiction.

It is a real and meaningful fact: a company operating GPU clusters in the hundreds of megawatts cannot route customer traffic without an autonomous system and transit relationships behind it.

It is also a startlingly thin public profile for a company of this scale, and the thinness is not an accident of my searching method.

Financial disclosure is a legal obligation enforced by securities regulators. Registry disclosure is a side effect of how addresses are allocated: the registry records who holds a block of addresses and who is responsible for it, not what the holder earns or how much it has promised to spend.

What the peering record says, and what it does not

Peering databases answer a narrow question well: which network is this, and with whom does it exchange traffic. They are not asset registers. Neither PeeringDB nor a BGP toolkit states how many IPv4 addresses CoreWeave controls, how many IPv6 prefixes it originates, or when those resources were first registered. Those answers live in per-prefix registry records: WHOIS, the long-standing query service that returns the registered holder of an address block, and RDAP, its structured successor, which returns the same registration data as machine-readable JSON.

I did not obtain those per-prefix records as readable text.

What I can say without qualification is what the public network records show and what an authoritative registry says about jurisdiction.

What I cannot say is the number everyone actually wants: how many addresses CoreWeave holds.

There is one further registry-shaped fact worth recording, because it is the kind of thing that gets assumed rather than checked. CoreWeave operates in Europe, and a company with European infrastructure might be expected to hold a membership in RIPE NCC, the registry that allocates address space in Europe, the Middle East and Central Asia. CoreWeave UK Limited does not appear in RIPE NCC's public membership indices (RIPE NCC membership indices). That is an observation about a membership list, not proof about where any packet terminates. CoreWeave's European capacity can perfectly well sit inside another operator's address space, leased or contracted, with the announcing network belonging to a partner.

Which is precisely why the address layer matters more than it looks.

The address layer is the portability layer

A GPU cluster is welded to a building. Power, cooling, fibre routes and land are all immovable, which is why the AI build-out has become a story about substations and planning committees. Addresses are the one input in the stack that moves freely.

An organisation that holds its own registered address space can take it from one facility to another and keep the same numbers reachable, because the addresses belong to the organisation rather than to the landlord. An organisation that uses its hosting provider's space gets reachability for as long as the contract lasts, and renumbering when it moves. The first model is capital and optionality. The second is operating expense and dependency.

The distinction does not show up in a capex line. It shows up in how quickly a company can walk away from a site, a supplier or a country, and in how much of its own continuity it actually controls. It also shows up in the transfer market, where address blocks change hands under registry policy and published transfer records, and where the price of a scarce resource is set by everyone who needs one.

For an AI cloud that is opening capacity faster than any company in its category, those are not abstractions. Every new site needs addresses before it can serve a customer, and routing announcements before those addresses mean anything. The company that plans this well buys itself the ability to move. The company that does not discovers that its infrastructure is only as portable as its least portable contract.

How a $104 billion backlog meets a routing table

The financial record and the routing record describe the same company from two directions, and the mismatch is instructive.

Backlog is a promise: $104.2 billion of contracted future revenue, per the 27 October analysis (247wallst). Delivering on a promise of that size requires an extraordinary amount of physical capacity, and every increment of capacity is a new place on the internet that must be addressed and announced. Capital expenditure guidance of $35 billion to $39 billion for 2026 is the price tag on that expansion (247wallst). Interest expense of $640 million, up from $267 million, is the cost of financing it (247wallst). The 2025 net loss of $1.2 billion, on revenue of $5.1 billion, is what the growth phase looks like on the income statement (Economic Times).

Yet the public registry artefact remains a single ASN.

Companies at this scale usually accumulate a visible trail: multiple autonomous systems for regional operations, documented transfers, prefixes registered to named subsidiaries. CoreWeave's trail, as far as public records I could retrieve go, is one number and a jurisdiction.

Two readings are available and only one of them is supported. The first is that the company holds a modest, tightly managed address estate and keeps its network simple — a legitimate engineering choice that many operators would defend. The second is that the register is richer than the public view suggests and simply is not indexed in the places I could read. Without per-prefix records, neither reading can be demonstrated, and a journalist who picks one has stopped reporting and started guessing.

What can be said is that the circulating numbers about CoreWeave deserve the same scepticism as the register. CoreWeave's own news feed does not carry the 3.1 GW contracted-capacity figure or the Q3 2025 revenue-backlog figure that travel widely in commentary about the company (coreweave.com). A figure that no primary publisher carries is not thereby false. It is unattributed, and unattributed numbers in a capital-intensive sector acquire a life of their own precisely because everyone repeats them.

The evidence boundary

Three documents would settle most of this piece, and I did not obtain any of them as readable text: the per-prefix ARIN WHOIS and RDAP records for CoreWeave's announced prefixes; any public document stating CoreWeave's total IPv4 or IPv6 holdings; and CoreWeave's own SEC filings and quarterly shareholder materials.

The financial figures in this article therefore travel with named attribution and stated periods rather than as independently verified line items, and no prefix list, allocation size or registration date is asserted here.

That is the boundary of this report, and it is a boundary drawn around what I could read, not a claim about what exists.

What would close the gap

Three public events would each make this measurable. First, a published address total from the company itself — the sort of number that appears in a technical talk, a peering policy or a procurement notice long before it appears in a marketing page. Second, growth in the per-prefix registry record: more prefixes registered to CoreWeave entities, or a documented transfer that brings address space under its control. Third, any RIPE NCC membership or registry presence for a CoreWeave entity, which would tell you whether European expansion is being carried on its own numbers or somebody else's.

None of those is a scandal waiting to happen. They are the ordinary paperwork of an extraordinary build-out, and the fact that the financial paperwork is public while the network paperwork is not says something about which layer the market has learned to watch. The address layer is where an AI cloud's options are stored, and right now almost nobody is reading it.