Summary

  • OpenAI OpCo, LLC holds two ARIN-registered autonomous systems — AS401518 ("OAI-01", registered 2024-11-26) and AS401864 ("OPENAI", registered 2025-07-24) — but the attributable footprint remains one 512-address IPv4 prefix, with aggregators disagreeing over whether it is even actively announced.
  • The consequential change is contractual, not routing: restructured Microsoft terms freed OpenAI to buy compute and sell on rival clouds, while a reported $300 billion Oracle purchase obligation begins in 2027.
  • Stargate's capital stack is debt-heavy: roughly 10% equity was under discussion in early 2025, project finance carried the Abilene campus ($11.6 billion secured by Crusoe and partners), and press reporting through late 2025 described a roughly $100 billion debt pile.
  • Self-operated network infrastructure remains an administrative intent signal — a PeeringDB record with an open policy and undisclosed traffic — not yet a verified operating backbone.

The public record shows OpenAI building the paperwork of network ownership faster than the network itself. What changed in its dependence on cloud landlords is not an ASN but a set of contracts: Microsoft's exclusivity loosened, Oracle became the largest compute counterparty, and the bill starts arriving in 2027.

Two autonomous systems, one small prefix

The hardest fact in this story is also the least glamorous. On 2024-11-26, the American Registry for Internet Numbers registered AS401518 under the name "OAI-01" to OpenAI OpCo, LLC, organization handle OOL-15 (https://whois.arin.net/rest/asn/AS401518). Roughly eight months later, on 2025-07-24, a second autonomous system followed: AS401864, with the more directly branded AS name "OPENAI", under the same OpenAI OpCo, LLC organization (https://whois.ipip.net/AS401864). The organization record behind both sits at 3180 18th Street, San Francisco, with technical and abuse contacts routed to an internal "Network Engineering" role at an openai.com address.

Holding an autonomous system number is an administrative act. It gives a company a routing identity in the global table — the ability to originate prefixes, announce routes and peer directly with other networks. It does not, by itself, mean the company operates meaningful infrastructure. The operational footprint attributed to AS401518 is modest: aggregators report a single IPv4 prefix, 199.47.142.0/23, totalling 512 addresses, and no IPv6, with connectivity bought as transit from Level 3 (AS3356) and AT&T (AS7018) (https://ipregistry.co/AS401518).

Even that modest footprint is disputed. IPinfo labels AS401518 "Inactive" with zero prefixes and zero addresses, while IPRegistry, IPGeolocation.io and Qrator Radar all attribute the 512-address allocation to the ASN. Qrator Radar adds that the single prefix is globally visible — above 15% of its vantage points — with zero recorded route leaks, hijacks or bogons. The contradiction matters less for what it proves than for what it prevents anyone from proving: nobody can currently demonstrate, from these sources alone, that OpenAI is routing meaningful production traffic over its own numbers.

The peering picture is equally unresolved. IPRegistry describes AS401518 as having no direct peering agreements, two transit upstreams and no downstream customer networks. Qrator Radar, by contrast, records one peering, and PeeringDB — the industry's self-reported interconnection registry — carries a record for AS401518 under organization "OpenAI" with RIR status "ok", a verified IRR AS-SET, and a general peering policy of "Open" with no ratio, contract or multi-location requirements; traffic levels and geographic scope are not disclosed, and the record was last updated 2026-06-22.

That combination — an open peering policy with no disclosed traffic, no disclosed exchange points and conflicting activity reports — is the signature of a network preparing to interconnect rather than a network already doing it at scale. It is an intent signal, administratively verified, operationally unproven.

The hiring and the stated ambition

OpenAI's own communications fill in the intent. The company posted a Stargate role for a principal network engineer to define and evolve its global network architecture, covering routing, topology and control-plane strategy across clouds, points of presence and self-built data centers, along with multicloud interconnect strategy and BGP policy including peering and route-leak mitigation. The role reports to Anuj Saharan, OpenAI's industrial compute lead, who described Stargate as requiring "the world's largest network".

OpenAI has also published a description of the networking build for its MRC supercomputer in the Abilene, Texas area — the Stargate I campus — describing the network architecture work for a frontier-scale training cluster. These are company self-descriptions of company infrastructure. They establish that OpenAI is doing genuine network engineering work at its own sites. They do not establish that the two autonomous systems carry that traffic, and nothing in the retrieved registry evidence ties the Stargate backbones to AS401518 or AS401864.

The honest reading is a two-track story. Track one: OpenAI is assembling the administrative and human apparatus of a network operator — two ASNs, a PeeringDB presence, an open peering posture, specialist hiring. Track two: the compute that actually trains and serves its models still runs on other people's infrastructure, now primarily Oracle's and Microsoft's, under contracts whose costs dwarf anything the routing estate touches.

What actually changed: the contracts

The prior state was simple to describe. OpenAI was, for most of its history, a large tenant of Microsoft's Azure cloud. The state change did not arrive with an ASN. It arrived through a sequence of contract restructurings and compute agreements.

In January 2025, Microsoft announced an evolved partnership: a restructured joint-venture framework in which Microsoft retains an increased revenue share and intellectual-property rights through an AGI verification gate, OpenAI's cloud commitment to Azure expanded, and Microsoft approved OpenAI's ability to pursue additional compute outside Azure for research and training models. That approval is the hinge. From that point, OpenAI could buy compute elsewhere without Microsoft's veto.

It did. In July 2025, OpenAI announced an agreement with Oracle to develop 4.5 gigawatts of additional Stargate data center capacity in the United States, bringing over 5 GW under development intended to run more than two million chips — while noting that Microsoft continues to provide cloud services for OpenAI, including through Stargate. By September 2025, OpenAI claimed five new US sites brought Stargate to nearly 7 GW of planned capacity and over $400 billion in investment over three years, with the July Oracle partnership characterized as "exceeds $300 billion" over five years.

The Wall Street Journal quantified the obligation differently and more sharply: OpenAI signed a contract with Oracle to purchase $300 billion of computing power over roughly five years — a commitment that, the paper noted, far outstrips OpenAI's current revenue, and the majority of Oracle's newly disclosed revenue backlog was reported to come from the OpenAI deal. The Verge added the timing detail: the contract begins in 2027. The framing difference is not trivial.

OpenAI describes a capacity partnership; the reported instrument is a purchase obligation with a deferred start date — a liability that grows silently until the payments begin.

The Microsoft relationship did not end; it was renegotiated into a different shape. Microsoft's October 2025 blog described the next chapter: Microsoft holds investment rights tied to OpenAI targeting a $250 billion Azure commitment, following the recapitalization. Press characterization of the restructuring was blunter: the exclusivity was "gutted," freeing OpenAI to sell on AWS and Google Cloud (https://venturebeat.com/technology/microsoft-and-openai-gut-their-exclusive-deal-freeing-openai-to-sell-on-aws-and-google-cloud), and The Verge described the partnership's restructured terms as a negotiated separation of Cloud commitments (https://www.theverge.com/tech/921210/microsoft-openai-partnership-divorce-notepad). By April 2026, TechCrunch reported that OpenAI had ended Microsoft's legal peril over its roughly $50 billion Amazon commitment — Microsoft consented to the AWS arrangement (https://techcrunch.com/2026/04/27/openai-ends-microsoft-legal-peril-over-its-50b-amazon-deal/).

Read as a capital-allocation story, the dependence did not shrink. It migrated. OpenAI swapped a single-landlord exclusive tenancy for a portfolio of enormous counterparty exposures: a reported $300 billion Oracle purchase stream starting 2027, a $250 billion Azure commitment Microsoft holds investment rights around, a reported ~$50 billion AWS arrangement, and NVIDIA's announced intent to invest up to $100 billion in OpenAI tied to OpenAI deploying NVIDIA systems at scale.

Who pays: the Stargate capital stack

Stargate was announced on 2025-01-21 as a new company intending to invest $500 billion over four years in US AI infrastructure for OpenAI, beginning with $100 billion deployed immediately; the initial equity funders are SoftBank, OpenAI, Oracle and MGX, with SoftBank holding financial responsibility, OpenAI operational responsibility and Masayoshi Son as chairman. Even at launch, OpenAI said it was still evaluating sites and finalizing definitive agreements — the baseline condition against which everything since must be measured.

The capital stack was never mostly equity. Axios reported that a portion of the initial $100 billion was expected to be funded through third-party debt rather than equity. Bloomberg reporting in February 2025 described SoftBank exploring structures in which partners contributed roughly 10% of overall cost in equity and tapped debt markets for most of the rest; one modeled breakdown was roughly 10% common equity, 20% preferred equity and mezzanine debt, and 70% senior debt — with no decisions made and viability questioned because cash flows from AI services were, at that point, hypothetical.

The project-finance pattern shows up concretely at Abilene. Crusoe, Blue Owl Capital and Primary Digital Infrastructure entered a joint venture to develop the Stargate I campus, and Crusoe secured $11.6 billion in debt and equity for the campus. Kirkland & Ellis, advising Bank of America-led funds, confirmed the third-party joint-venture and financing structure. Note whose balance sheet that debt sits on: the developer's, not OpenAI's. OpenAI is the tenant.

Its own borrowing arrived separately — Reuters reported in September 2025 that OpenAI, under pressure to meet demand, widened Stargate's scope and was eyeing debt financing for chips, and by late November 2025 The Irish Times reported OpenAI and its partners amassing roughly $100 billion in debt to fund their ambitions.

This is the mechanism that matters for OpenAI's independence claims. Self-operated infrastructure changes bargaining power only if it is funded by capital OpenAI controls. When the data centers are built on developer-side project debt and the compute is bought through a $300 billion purchase obligation, the counterparty relationship is preserved — arguably deepened — even as the logo on the facility changes from Microsoft to Oracle.

Where the mechanism can still break

Three failure paths deserve monitoring. First, the revenue condition. Project finance of this shape services itself from contracted cash flows. The February 2025 Bloomberg reporting was explicit that AI-services cash flows were hypothetical; the $300 billion Oracle obligation begins in 2027. If OpenAI's revenue growth disappoints relative to the commitments, the debt stack and purchase obligations converge on the same income statement.

Second, the deployment condition. NVIDIA's up-to-$100 billion investment is tied to OpenAI deploying NVIDIA systems at scale — the capital follows the deployment, not the reverse. Stargate's own milestones are company claims: "nearly 7 GW planned" is a plan, not energized capacity, and the claim of a clear path to the full $500 billion by end of 2025 was never independently confirmed in the sources retrieved here.

Third, the network condition. The verifiable test of self-operated infrastructure is observable in public data: larger, multi-region prefixes announced by AS401518 or AS401864, populated PeeringDB facility and exchange-point records, and settlement-free peering sessions visible in global routing tables. Today the record shows one 512-address prefix with disputed activity and an open policy with undisclosed traffic. Until those observables move, "OpenAI runs its own network" remains an administrative statement, not an operating fact.

What this means

The state difference worth tracking is not that OpenAI acquired routing numbers. It is that the company converted a single-supplier dependency into a portfolio of leveraged, dated obligations — Oracle from 2027, Azure around a $250 billion target, AWS at a reported $50 billion — while the balance sheets carrying the infrastructure debt belong mostly to partners. OpenAI OpCo, LLC's two autonomous systems are the cleanest evidence of administrative intent to own network identity. The $300 billion purchase contract is the cleanest evidence that, for the compute that matters, OpenAI has chosen to rent at historic scale rather than own.

The next observable condition is unglamorous and decisive: watch the routing tables, the PeeringDB facility records, and the first Oracle invoices in 2027.