Summary

  • Address registration and network access were related but separate. A valid Internet network number established uniqueness and an administrative record; it did not provide a circuit, regional membership, NSFNET eligibility, or a route that other operators would accept.
  • NSF financed and supervised the backbone programme, Merit managed it, IBM supplied important packet-switching technology and engineering, MCI supplied transmission facilities, and later ANS operated much of the T3 infrastructure. Regional networks, universities, identifier authorities, and peer operators retained different decisions.
  • The NSFNET routing system converted institutional relationships into operational reach by validating network numbers, autonomous-system identities, and authorised regional representation. This made routing records as important as registry records without turning Merit into the allocator of Internet addresses.
  • Public investment produced substantial interoperability benefits: faster national links, shared operations, broader university access, and a route system that connected thousands of networks. Dependence on that subsidised reach nevertheless gave regional and backbone decisions consequences beyond their formal mandates.
  • The 1993–1995 transition demonstrated the separation of the two axes. Registration moved to InterNIC while regional networks procured commercial transit, routes migrated to multiple providers and exchange points, and the old NSFNET Backbone Service ended on 30 April 1995.

A layered illustration, not a documented admission case

Kent State University offers a useful view of the layers through which an address became usable, but the surviving evidence does not record a single continuous admission transaction.

The July 1990 Internet number register published as RFC 1166 listed 131.123 as KENT-STATE in the research category. That entry establishes that the number was registered to Kent State by that date. It does not reveal the original application, the date on which the number was first requested, the reasoning used to approve it, or the terms under which the university obtained external connectivity.

A separate 1991 operating account, RFC 1246, described Kent State’s position inside the Ohio Academic Resources Network, or OARnet. It recorded a DS1 circuit from Kent State to OARnet’s Akron point of presence and a 56-kilobit connection represented as 131.187.36.0. The latter number belonged to OARnet infrastructure, not to Kent State’s registered 131.123 network. The document also described a roundabout backup path through Cleveland when preferred links were unavailable.

Those are two genuine observations: a registry snapshot and a later regional-network operating snapshot. They do not prove Kent State’s original membership terms, its first successful external route advertisement, or the exact path taken by a packet on a particular day. The OARnet restart described in RFC 1246 concerned the convergence of routes to OARnet infrastructure addresses. It should not be misread as a captured end-to-end route to 131.123.

The limited example still exposes the governing mechanism. Kent State needed an identifier that would not conflict with another network. It needed an attachment to OARnet. OARnet needed functioning external connections and permission to represent the networks behind it. Backbone and peer routers then needed to accept the relevant reachability information. A failure at any one of those stages could make a destination unreachable even though the other records remained intact.

The distinction matters because the result could look unitary from the campus. A user saw only that a remote host could or could not be reached. Behind that outcome were decisions by a registry, a university, a regional network, backbone operators, circuit providers, and remote peers. Their powers interacted, but they were not interchangeable.

What NSF commissioned—and what the public record proves

The first instrument in the backbone chain was not an address-allocation policy. It was the National Science Foundation’s 1987 solicitation for management and operation of an expanded NSFNET backbone.

Merit’s later history dates Project Solicitation NSF 87-37 to 15 June 1987. The 1992 House Subcommittee on Science hearing, Management of NSFNET, reproduces important language from the solicitation. It describes the proposed system as a three-level hierarchy: a transcontinental backbone, autonomously administered second-level networks, and campus networks connected beneath them. The reproduced text also invited proposers to suggest alternative architectures or methods that might be more appropriate, economical, or effective.

That language establishes the intended separation between national backbone management and regional administration. It does not establish every term of the eventual agreement. A complete authenticated set containing the executed Cooperative Agreement NCR-8720904, all amendments, every IBM and MCI arrangement, and the later Merit–ANS operating instrument is not available through the cited public archive. Assertions about unpublished clauses would therefore exceed the evidence.

The narrower institutional chain is well supported. Merit proposed a 1.5-megabit-per-second T1 backbone with IBM and MCI in August 1987. NSF announced a five-year cooperative agreement with Merit in November. Contemporary congressional testimony and Merit’s institutional account identify Merit as the organisation responsible to NSF for management and operation of the backbone project. IBM supplied packet-switching hardware, software, and engineering. MCI supplied long-distance transmission facilities. The State of Michigan provided additional support.

These contributions did not create a single corporate or federal actor called “NSFNET.” NSF was the funder and programme supervisor. Merit was the cooperative-agreement holder and backbone manager. IBM engineers contributed systems and routing work. MCI supplied circuits and communications expertise. Universities and regional networks remained separately administered. The IANA function and the Internet Registry handled identifiers through another institutional chain.

The distinction between a cooperative agreement and ordinary purchasing is also relevant, though it does not decide the article’s central question. NSF exercised continuing involvement in an infrastructure programme whose design and operation depended on contributions from several organisations. It could supervise the award, review performance, approve changes, and decide whether to extend support. It did not acquire ownership of all addresses whose traffic later crossed the service.

The new T1 backbone became operational across thirteen sites in the summer of 1988. Merit’s final institutional account places operation in July and reports 152 million packets per month at 1.5 megabits per second. NSF testimony to Congress in 1992 used a different baseline, describing the new backbone as carrying traffic from August 1988 and presenting growth from approximately 200 million packets per month to 11 billion by early 1992.

Those figures are not necessarily contradictory. They may reflect different reporting dates, partial versus full-month observations, or later rounding. The public documents do not define the difference closely enough to merge them into one exact “first-month” count. The defensible conclusion is that the T1 service entered operation in July–August 1988, with Merit reporting 152 million monthly packets at the July point and NSF later using an approximately 200-million-packet early-service baseline.

That qualification does not diminish the achievement. The thirteen-node system replaced an overloaded 56-kilobit arrangement with a production backbone, supported regional interconnection, and gave universities access to remote computing and information resources without requiring every campus to build a national network.

The T1 system joined several gates

The operating structure from 1988 to 1990 can be separated into distinct decisions:

Function Principal actor or instrument What it controlled
Programme funding and supervision National Science Foundation Selection and support of the backbone programme, review of performance, and conditions attached to federal support
Backbone management Merit Network under NCR-8720904 Engineering coordination, network operations, information services, regional liaison, and routing-policy administration
Technology and transmission IBM and MCI, with support from Michigan Packet-switching systems, software, engineering, and long-distance circuits
Campus-facing access Regional networks and participating institutions Membership, local circuits, equipment, fees, technical readiness, and campus attachment
Identifier administration IANA at USC’s Information Sciences Institute and Internet Registry functions at the SRI-operated DDN-NIC Unique network and autonomous-system numbers and their administrative records
Route representation Campus and regional operators Which regional network could represent a destination and with what preference
Backbone route acceptance Merit’s operations and backbone routing machinery Validation of network and autonomous-system information against policy records
Further propagation Other federal, regional, international, and emerging commercial operators Acceptance and onward advertisement beyond the immediate backbone
Traffic eligibility NSF backbone-use conditions and connected-network policies Whether particular traffic could use the federally supported path

The financial structure increased the practical importance of alignment among these gates. RFC 1192, a report from a 1990 commercialisation workshop, estimated annual backbone costs at approximately $10 million, of which NSF paid less than $3 million. It attributed much of the balance to the State of Michigan and contributed IBM and MCI services. The same report estimated that NSF supplied about 40 per cent of the costs of the mid-level networks it supported, while noting a range from zero to 75 per cent.

Those were workshop estimates, not a uniform entitlement or price schedule. They nevertheless show why an NSF-supported route could be more valuable than an address alone. Federal funds, state support, corporate contributions, regional fees, university spending, and in-kind engineering combined to create a national service whose full cost was not charged to each attached campus as commercial long-distance transit.

The benefit was collective. A university did not have to negotiate a dedicated circuit to every supercomputer centre or every other regional network. Common protocols and an operated backbone allowed one attachment to reach a growing set of destinations. The resulting network effects increased the value of every usable route.

Dependence was the other side of that benefit. Once researchers, libraries, administrators, and campus computing services relied on remote connectivity, delay at a regional link or backbone policy entry imposed a real cost on users. Yet the location of the remedy depended on the failure. A mistaken registry entry belonged with the identifier administrators. A failed leased line belonged with the campus, regional network, or carrier. An unauthorised announcement belonged with routing operations. A remote peer’s rejection could not be corrected by NSF merely declaring the destination legitimate.

Regional access was not a uniform federal rule

The three-level architecture placed regional networks between campuses and the national backbone, but those networks were not identical administrative branches of NSF or Merit.

OARnet’s 1991 operating account described a network serving Ohio higher education and permitting connections for corporations engaged in research, product development, or instruction. It used TCP/IP and DECnet, connected 29 sites directly, and operated a topology in which 13 routers functioned as autonomous-system boundary routers.

Its principal external routing relationship in the RFC 1246 account passed through a demilitarised network in Columbus connected to CICNet. Parts of OARnet generated a default route when the relevant exterior session was available rather than carrying all external EGP information through the interior. OARnet also had gateways to other systems, including the NASA Science Internet.

This arrangement gave OARnet operational choices. Its engineers determined internal routing costs, backup paths, point-of-presence design, and how external reachability became a default inside the regional system. Merit’s backbone operators did not choose the cost of OARnet’s path from Kent State to Akron, and the DDN-NIC did not configure OARnet’s OSPF routers.

The recorded Kent State links show what regional topology changed. The DS1 circuit offered a faster preferred path. The 56-kilobit connection and the longer path through Cleveland provided a less attractive alternative during a restart. As links recovered, OSPF recalculated routes. The event demonstrates regional convergence and resilience; it does not show Kent’s external prefix changing or prove that all external traffic used NSFNET.

Other regional networks used different organisational and technical arrangements. The study The Strategic Future of the Mid-Level Networks described BARRNet as distributing equipment ownership and operating responsibility among participating institutions. NYSERNet relied heavily on arrangements with telecommunications companies. PREPnet outsourced extensive functions to a carrier. NorthwestNet used Boeing Computer Services, while NEARnet used BBN.

BARRNet’s 1991 footprint included roughly 80 university, government, and commercial sites with access speeds extending from 9.6 kilobits per second to T1. It connected at Stanford to both T1 and T3 NSFNET facilities and also had links to ESnet, defence networks, and California university systems. This was not the same topology, market, or institutional environment as OARnet.

Consequently, “NSFNET access” could not be reduced to one national campus application. A university typically needed a regional organisation willing and able to connect it, a suitable leased circuit, equipment, technical staff, and a route arrangement. The regional organisation might receive federal support, but it could also rely on state appropriations, institutional fees, corporate members, carrier contracts, or in-kind contributions.

A delayed campus was not necessarily prohibited from the Internet by federal decree. It might instead lack an affordable last-mile circuit, fall outside a regional network’s membership class, or be unable to meet an equipment requirement. Whether another route existed depended on geography, provider presence, eligibility, and interconnection.

This evidentiary limit is important for Kent State. The surviving materials do not provide Kent’s original OARnet connection agreement, fee, installation date, or alternative-service quotations. They therefore cannot support a claim that a specific replacement provider was available to Kent in 1988 at a known price. OARnet’s connections to CICNet and NASA do not establish that Kent could have purchased those paths independently or used them if its OARnet attachment had been refused.

Regional admission was a real gate. It was not a nationally standardised gate, and the available Kent record does not preserve a denial, appeal, or costed alternative.

A registered number was not a service entitlement

The identifier system carried traces of earlier interconnection policy. RFC 1166 distinguished networks participating in the research and operational Internet from independent IP networks. Independent networks were marked with an asterisk and required separate permission to interconnect. Kent State’s 131.123 and OARnet’s 131.187 appeared as research networks without that mark.

That was significant administrative information in July 1990, but its meaning must remain bounded. The entry did not prove that a route was live at every moment. It did not specify which regional network was responsible for every packet. It did not order a carrier to supply a circuit or compel a foreign network to accept the destination.

RFC 1174, issued in August 1990, explains both the institutional division and the growing inadequacy of “connected status.” It identified the IANA function as being performed by USC’s Information Sciences Institute. It identified SRI International as the Internet Registry responsible for gathering and registering information about assigned network and autonomous-system identifiers.

The document described a history in which numbers had first been assigned to organisations participating in Internet research and later to government or government-sponsored networks permitted to interconnect. As TCP/IP spread into private networks, the registry assigned globally unique numbers even when an organisation did not intend to connect to the federally sponsored Internet. “Connected status” became the attempted distinction between possessing an identifier and having government sanction to interconnect.

By 1990 that binary field no longer described the network accurately. Regional systems served mixed memberships. Commercial networks were emerging. International networks could not sensibly be reduced to a United States sponsor’s approval. A network might carry some traffic over an NSF-supported path and other traffic over a different peer or backbone.

RFC 1174 therefore recommended that the Internet Registry remove connected status from forms and databases, gather access and transit-policy information instead, and permit any registered network to enter the Domain Name System without regard to connected status. It stated that the registry should administer number space while network administrators enforced traffic policy.

The document was an IAB recommendation, not a technical standard and not proof that every form, database, and router changed immediately. Its chronology should not be compressed into an overnight reform. What it clearly establishes is that policymakers recognised registration and interconnection as different functions and sought to remove access enforcement from the naming and identifier-registration layer.

This was not a merely theoretical distinction. An organisation could require a unique number for a private TCP/IP network while having no external transit. Conversely, a campus could have physical access to a regional network while still needing legitimate, non-conflicting address space before it could be represented safely to the wider Internet.

Routing policy created the operational join

Registration made an identifier administratively legitimate. Routing policy determined whether the backbone believed a particular network was reachable through a particular regional system.

In RFC 1092, Jacob Rekhter described a limitation of the Exterior Gateway Protocol used between the new backbone and regional networks. EGP alone could not prevent one regional network from claiming a destination belonging behind another. It also could not express a reliable hierarchy of preferred and backup paths in a meshed environment with additional “backdoor” links.

The proposed remedy was institutional as well as technical. A network would select one or more regional representatives through bilateral arrangements. Information about the chosen primary and secondary representatives would be supplied to the NSFNET Network Operations Center and entered into the Routing Policy Database. The backbone would ignore an advertisement from a regional network that was not authorised to represent that destination.

RFC 1093 described the corresponding architecture. Regional backbones were expected to use unique autonomous-system numbers. Backbone nodes checked both network numbers and the source autonomous-system number. Preferred paths were derived from information supplied by regional backbones and attached campuses. Regional networks could generate internal defaults, while the backbone maintained explicit reachability for attached and peer networks.

The route therefore depended on records from different authorities agreeing:

  1. A network number had to be unique and properly registered.
  2. A campus needed an attachment relationship with a regional network.
  3. The campus and regional network needed an agreed representation and path preference.
  4. Backbone policy data had to authorise the regional autonomous system to announce that destination.
  5. The relevant circuit and router sessions had to be operational.
  6. Other operators had to accept and propagate the route if reach beyond NSFNET was required.

These conditions were cumulative but not constitutionally unified. The registry could correct the identity associated with a number but could not repair a failed DS1 circuit. A regional operator could restore a link but could not make a duplicate number globally unique. Merit’s Network Operations Center could reject an unauthorised advertisement but could not force an independent peer to accept a route.

This is where backbone access shaped address power. The NSFNET policy database was not the address registry, yet inclusion in a widely used routing system made a registered number more useful. As the reachable network grew, correct representation through the backbone acquired greater practical value.

The same system constrained unilateral route claims. A regional network could not simply announce another organisation’s network number with a preferred metric and expect the backbone to believe it. Policy records and autonomous-system validation turned an administrative relationship into a routing permission.

The resulting authority was narrower than address ownership and broader than mechanical packet forwarding. Backbone operators controlled what their own service accepted. Because that service had exceptional reach, their operational decisions could affect many users. The scale of the consequence came from topology and adoption, not from a global mandate.

What the growth numbers count

NSFNET’s expansion is material evidence of public benefit, but its statistics describe different populations.

The July 1988 Merit figure of 152 million monthly packets and NSF’s later approximately 200-million early-service baseline concern traffic. They do not count addresses or institutions. NSF testimony reported approximately 11 billion packets per month by March 1992, a measure of rapidly increasing use rather than a census of connected organisations.

The backbone itself grew from 13 T1 sites to a 16-site T3 architecture. A backbone site was not a campus, regional network, or individual user. It was a node or attachment point within the national service.

Congressional statements in 1992 referred to approximately 5,000 networks, including about 1,500 outside the United States, connected into the wider system. Those estimates were presented in a political and institutional hearing and should not be treated as an exact routing-table snapshot.

A dated NSFNET routing update from January 1993 reported 8,997 networks configured in the T3 policy database. That count represented configured network entries and their preferred autonomous-system paths. It was not a count of unique organisations. One institution could hold multiple classful networks, and a configured entry might have primary and backup representations.

The allocation totals in RFC 1366 measured something else again. In 1992 the document reported 49 allocated Class A numbers, 7,354 Class B numbers, and 44,014 Class C numbers. Those were allocation units in the classful address system, not NSFNET customers. Some were used by private or non-NSF networks, and a Class A represented vastly more address capacity than a Class C.

The later traffic visualization preserved by CAIDA reports 18.5 trillion inbound bytes during December 1994. For that visualization, 24,435 domestic client networks were aggregated into 12,177 virtual traffic connections according to city and backbone node. Again, a client network, a virtual line on a visualization, and an institution were not equivalent.

Used carefully, the figures show several forms of expansion: more traffic, more configured routes, more address allocations, more client networks, and greater geographic reach. They do not prove that backbone funding alone caused every change. Falling equipment costs, the spread of TCP/IP software, regional investment, commercial services, campus demand, international networks, and new applications all contributed.

The causal claim can therefore remain modest but important. NSF’s investment and the Merit-led service supplied a shared high-capacity route environment that allowed much of this growth to become mutually reachable. It did not produce every allocated address, and temporal correlation between address growth and backbone growth does not establish that Merit controlled assignment.

T3 changed capacity and operations, not the identifier authority

By 1990, the T1 system was again under pressure. The T3 upgrade increased nominal backbone transmission from 1.5 to 45 megabits per second and expanded the architecture to 16 sites. It also changed the operating organisation.

Merit, IBM, and MCI formed Advanced Network & Services, Inc., or ANS, in September 1990. The 1992 congressional record describes Merit as remaining responsible under its cooperative agreement while subcontracting substantial management and operation of the upgraded backbone to the new nonprofit organisation. Merit’s final institutional history similarly presents ANS as the operating vehicle for much of the T3 work.

The available public materials establish the organisational outline but do not expose every operative clause of the 17 September 1990 Merit–ANS agreement. It is therefore safer to describe the observable division than to attribute undocumented rights. NSF remained the programme funder and supervisor. Merit remained responsible in the cooperative-agreement chain. ANS undertook extensive T3 engineering and operations. IBM and MCI continued to supply important technology, facilities, personnel, and support.

The move to T3 was not instantaneous. Node installation, initial traffic carriage, migration of regional attachments, and retirement of the T1 network were different events. Merit’s account places completion of the 16-site T3 system in 1991. T1 and T3 facilities then coexisted while attachments and routes moved.

A Merit operating notice archived in the November 1992 NANOG record scheduled the T1 backbone to shut down on Wednesday, 2 December 1992. That dated notice supplies the missing distinction between the earlier arrival of T3 production service and the later retirement of the remaining T1 service. The T3 backbone did not become fully exclusive merely because the first T3 links carried packets.

The route system also expanded. The January 1993 update reporting 8,997 configured T3 networks illustrates the amount of policy data backbone operations had to maintain. Each entry represented a network and expected autonomous-system paths, not a grant of an address. The database operationalised relationships already established elsewhere.

This phase therefore intensified the practical gate without changing its legal identity. A missing or incorrect T3 policy entry could affect reachability across a much larger service. It did not make ANS or Merit the IANA, and it did not transfer ownership of registered numbers to NSF.

Commercialisation introduced alternatives unevenly

Commercial TCP/IP services were emerging before the T3 transition was complete. AlterNet and Performance Systems International marketed connectivity. Regional networks served some industrial research organisations and sought revenue beyond federal support. The Commercial Internet Exchange offered interconnection outside the NSF-supported backbone’s traffic conditions.

ANS created a for-profit subsidiary, ANS CO+RE, in 1991 to provide commercial service. The arrangement became controversial because ANS also operated infrastructure used for the federally supported service. Participants in the 1992 House hearing disputed cost allocation, consultation, interconnection, and competitive advantage.

The testimony does not support converting every allegation into a finding. Critics argued that the structure favoured one path and blurred the boundaries of public support. Merit and NSF argued that the arrangement encouraged private investment while protecting research and education service. The hearing establishes the presence of a serious institutional dispute, not a proven conspiracy or ownership claim.

For address value, commercialisation mattered because it made an alternate pairing increasingly possible: a valid registered number could be routed through a commercial provider rather than through NSFNET. A customer could acquire service, arrange a circuit, and ask the provider to represent its network.

That possibility remained conditional. A provider needed geographic presence or a reachable point of presence. The customer needed a last-mile circuit, equipment, staff, a service agreement, and routing acceptance. A commercial backbone’s existence in the United States did not prove that every university could buy comparable service locally or affordably.

The surviving Kent State materials do not provide a contemporaneous AlterNet, PSI, or other commercial quotation covering Kent’s location, eligibility, installation, and complete cost. They do not show that BITNET, the NASA Science Internet, or a neighbouring regional network was available as a general IP-transit substitute. It would therefore be speculative to state that a specific alternate service was feasible for Kent in 1988 or to assign it a comparative price.

What the broader record does show is a change in the market over time. By the early 1990s, organisations had more possible upstreams and more places to exchange traffic. BARRNet’s multiple connections demonstrate that regional systems could use NSFNET alongside agency and local paths. RFC 1092 had already contemplated primary, secondary, and “backdoor” representations. Commercial growth expanded those technical possibilities into service choices, though unevenly.

The practical authority of NSFNET consequently weakened before the service formally ended. It remained highly important, but a registered network was less dependent on one subsidised national path once commercial providers and exchange relationships could deliver comparable destinations.

Registration moved on a separate timetable

Identifier administration underwent its own institutional change while the T3 backbone was operating.

NSF’s March 1992 solicitation, NSF 92-24, divided network information services into registration, directory and database, and information functions. Effective 1 January 1993, NSF’s Cooperative Agreement NCR-9218742 with Network Solutions established non-military registration services under what became the InterNIC framework.

The statement of work covered Internet domain registration, network-number assignment, and autonomous-system-number assignment in coordination with IANA and the relevant policy documents. It did not assign Network Solutions responsibility for operating NSFNET routers or selecting commercial transit providers.

RFC 1400 documented the operational transition from the DDN-NIC to the InterNIC. It set 1 April 1993 as the effective point for non-DDN registration requests to move to the new service. Military registration remained on its separate path.

This sequence matters because it occurred before the old backbone shut down. By 1993, a university could direct a number or autonomous-system request to InterNIC while its regional network continued to use the ANS-operated T3 service. The routing-policy database and the registration database were administratively distinct even when they shared identifiers and contact information.

Accurate registration still affected routing. Operators needed to know which organisation held a network and whom to contact when an announcement was disputed. But the registry’s record did not activate a backbone interface. Likewise, a valid route in NSFNET did not transfer the underlying registration function to the backbone operator.

The 1993–1995 transition redistributed access authority

NSF’s May 1993 solicitation, NSF 93-52, proposed four distinct project areas: Network Access Points, a Routing Arbiter, regional-network support, and a very-high-speed Backbone Network Service for advanced research.

The structure deliberately avoided replacing the old NSFNET service with a single commodity backbone. Commercial network service providers would carry general traffic and interconnect at Network Access Points. Regional networks would procure upstream service from those providers, with transitional NSF assistance. Routing coordination would continue through a Routing Arbiter project. The vBNS would serve advanced research requirements rather than act as the one universal successor.

The GAO decision concerning Sprint’s challenge to the vBNS award confirms that NSF 93-52 contemplated several cooperative agreements and that MCI was selected for the vBNS project in February 1994. The decision also reinforces the need to distinguish that research service from the commercial transit replacing the old backbone.

The access chain changed accordingly:

Function in the transition Principal actors Operational consequence
Transitional funding NSF Supported migration without choosing one permanent commodity backbone
Old backbone continuity Merit and ANS Kept the existing service available while regional networks moved
Replacement transit ANSNet, internetMCI, SprintLink, PSINet, and other providers Sold connectivity under separate service arrangements
Regional migration Regional networks and their members Selected providers, installed circuits, tested routes, and bore local transition risk
Identifier administration IANA, InterNIC, and emerging delegated registries Continued number and contact administration independently of provider choice
Interconnection Network Access Point operators and participating providers Supplied locations for exchange among multiple backbones
Routing coordination Routing Arbiter participants, providers, and customer operators Maintained route information and diagnosed inconsistent reachability
Advanced research backbone NSF and MCI through the vBNS project Supplied a distinct high-performance service rather than general commercial replacement transit

A Merit transition report dated 30 September 1994 shows why registration alone could not complete the migration. It tracked five operational dependencies: working Network Access Points, NSFNET attachment to those points, new provider attachments, Routing Arbiter services, and each regional network’s connection to its chosen provider. A failure in any of those areas could leave an institution with valid identifiers but incomplete reachability.

The final transition was staged rather than ceremonial. Merit’s notice of 14 April 1995 reported that only seven organisations had completely severed their old NSFNET relationship. Many were using new providers while retaining NSFNET as a backup.

Remaining sessions were scheduled for a test shutdown on 21 April to reveal unreachable networks. Temporary restoration remained possible while problems were corrected. The notice then called for permanent termination of the remaining sessions on 28 April, followed by termination of the backbone service on 30 April.

An NSF release dated 15 May confirmed that the NSFNET Backbone Service had been decommissioned at midnight on 30 April 1995.

The shutdown strongly illustrates the separation between address registration and backbone service, but it does not prove that every individual prefix migrated without interruption. Demonstrating continuity for Kent State’s 131.123, for example, would require matched before-and-after routing observations in addition to the registry entry. The available transition notices show a system designed to preserve reachability while services changed; they do not supply a prefix-specific Kent trace.

What can be stated with confidence is that the old backbone’s termination did not abolish the identifier system. InterNIC and IANA functions continued. Regional networks bought replacement transit. Providers exchanged routes at new interconnection points. The operational value of addresses persisted only because those new actors carried and accepted the routes.

Testing the two axes without inventing alternatives

The distinction can be tested through several bounded scenarios.

A valid identifier without usable external transit

Suppose a university possessed a globally unique registered network number but lacked a functioning regional attachment or acceptable upstream.

It could use the number internally without colliding with another registered network. It could operate local TCP/IP services and exchange traffic over any bilateral path that agreed to carry it. Registration would remain meaningful.

What it would lack was general external reach. The university would need a regional connection, a commercial provider, an eligible agency path, or a dedicated peer. Each option required its own agreement and physical facilities. Registration would not compel any of them to provide service.

For Kent State, the historical record does not establish which substitute, if any, met all those conditions in 1988. OARnet’s later multiple gateways prove topology, not an independent service entitlement for Kent. The existence of BITNET would not by itself provide general IP transit. Commercial providers became more plausible after 1990, but no complete Kent-specific price or availability record has been recovered.

The justified conclusion is therefore limited: a registered address could survive without NSFNET, but its external usefulness depended on obtaining another carrier and accepted route. Whether that was feasible or affordable for Kent at a particular date remains unknown.

Physical access without a valid public identifier

Reverse the conditions. A regional network might have an available circuit and be willing to attach a campus, but the campus might lack a valid network number for general Internet use.

The circuit could carry traffic under a local addressing plan or another technically coordinated arrangement. Generic engineering possibilities included using valid provider-controlled address space or delaying public advertisement until registration was resolved. The surviving OARnet materials do not establish which remedy it would have offered Kent, so none should be presented as OARnet policy.

What the campus could not safely do was select another organisation’s public number and expect global routing to work. Duplicate numbering could misdirect packets or cause filters to reject the announcement. NSFNET’s policy machinery was specifically designed to compare network and autonomous-system information against expected representation.

The remedy would begin on the identifier side: obtain or correct a legitimate assignment and ensure that the responsible contacts and representation were accurate. Opening a backbone port could not make a duplicate identifier unique.

A legitimate route that one peer declined to carry

A destination could be properly registered, attached to a regional network, and accepted by NSFNET while remaining unreachable through another provider.

Each peer or backbone controlled its own routing policy. NSFNET’s scale made its information influential, but it could not command every agency, commercial, or international operator to propagate a network. A routed address was therefore evidence of acceptance along some path, not proof of universal approval.

The appropriate remedy was route diagnosis and inter-operator coordination. The registry could help identify contacts but could not compel carriage. Merit could correct its own policy database but could not configure every remote network.

Eligible connectivity with traffic that required another path

A campus could also possess a valid identifier and working route while some traffic was ineligible for the federally supported backbone. That was a traffic-policy issue, not cancellation of the address or denial of physical membership.

The institution might need to route the relevant traffic through another provider or demonstrate that it served the permitted research and education purpose. The important point here is institutional: acceptable-use compliance was another condition in the path, not the authority that created the number.

These scenarios avoid false symmetry. Their consequences were not identical, and the available remedies differed. Registration failure threatened uniqueness and stable representation. Regional failure threatened local attachment. Backbone failure threatened interregional carriage. Peer rejection threatened reach beyond the immediate provider. Traffic-policy conflict affected which uses could traverse a particular supported path.

To a campus user, all could produce the same symptom: a remote destination did not respond. Governance analysis has to reconstruct the layer at which the failure occurred.

Where the authority and consequence sat

NSF directly governed the programme it funded. It selected and supervised the cooperative-agreement structure, supported regional connectivity, reviewed performance, approved changes, and later designed the transition towards Network Access Points, commercial service providers, a Routing Arbiter, and the vBNS.

Merit managed and coordinated the backbone under its agreement with NSF. Its responsibilities included network operations, information services, regional liaison, and the policy data needed to keep a growing routed system coherent.

IBM and MCI supplied distinct technical and communications contributions. They were neither university-membership authorities nor Internet number registries. During the T3 phase, ANS assumed extensive engineering and operating work while Merit remained in the NSF agreement chain.

Regional networks controlled the campus-facing layer. They decided whom they could serve, how circuits and equipment would be arranged, what fees and local conditions applied, and how internal routing worked. Their structures were heterogeneous, and the consequences of geography or limited carrier choice varied substantially.

The IANA function, the DDN-NIC’s Internet Registry work, and later InterNIC registration services controlled identifiers and records through a separate sequence. Those records mattered because duplicate or misattributed numbers could not support dependable global routing. The identifier authorities did not operate Kent State’s DS1 circuit or select its upstream path.

Campus, regional, backbone, and peer operators converted those relationships into reachability. They configured routers, exchanged routing information, validated expected representatives, restored failed sessions, and decided which routes to propagate. Their work determined whether a registered number was usable at a given moment.

The proposition that NSFNET amplified address power therefore holds only in a distributed sense. NSF-funded connectivity made the Internet more useful. The more destinations the backbone and its peers connected, the more valuable it became to possess an identifier that was correctly represented through that environment.

That effect also increased the consequences of regional admission and routing-policy decisions. A record in the number register was necessary but insufficient. A regional circuit without a legitimate identifier was also insufficient. Neither axis alone produced general reachability.

Public investment should not be treated as incidental to this result. NSFNET created national interoperability at a speed and scale that individual universities were unlikely to reproduce separately. It shared fixed costs, developed operations, supported regional growth, and gave researchers access to distant resources. Its success is part of the explanation for its power: dependency followed utility.

Nor should dependency be confused with a formal monopoly over addresses. Commercial services, agency networks, direct peers, and regional backdoors existed or emerged at different times. Their availability was uneven, and their existence did not guarantee a feasible alternative for every institution. But they demonstrate that route value could migrate without requiring the identifier to be recreated.

The 1995 transition made that separability visible at system scale. The old backbone ended. Registration continued. Regional networks changed providers. Routes moved across commercial networks and exchange points. Routing coordination persisted in new organisations.

This was not proof that every migration was seamless. It was evidence that Internet identity and Internet carriage could survive institutional replacement because they had never been the same function.

NSF did not own the addresses carried over NSFNET. Merit did not decide all global Internet admission. ANS did not become IANA by operating the T3 service. The DDN-NIC and InterNIC did not provide campus circuits. Regional networks did not control every peer beyond their borders.

Yet their decisions aligned tightly enough that users could experience them as one gate. The governance significance of NSFNET lies in that alignment. Backbone access shaped address power not by absorbing identifier authority, but by determining whether a legitimate identifier could participate in one of the most valuable routing environments of its time.

Sources