The First-Mover Dividend in Legacy Address Space

Summary

• The RFC 790 and RFC 820 cohort shows that early Class A entries could persist, transition, or return: AMPRNET's documented 2019 disposition and Stanford's early renumbering show why nominal allocation size alone is not a realised windfall.
• A defensible first-mover dividend requires chain-of-control evidence and a documented consequence such as retained capacity, avoided cost, transaction proceeds, or option value; original request and decision files remain missing for much of the cohort.

The difference between arriving early and keeping the advantage

The most revealing comparison in the early Internet address register is not between a recipient and a modern market price. It is between two rows that looked broadly similar in 1981 and had diverged by the next published register.

RFC 790, published in September 1981, listed network 36 as SU-NET, the Stanford University Ethernet. On the same printed page, it listed network 44 as AMPRNET, the amateur-radio experiment network. Both occupied Class A numeric positions. Under the address format described in the RFC, each position had a 24-bit local field, or (2^24)—16,777,216—nominal local-address combinations. That figure describes the width of the field. It is not a count of hosts, assigned addresses, routed destinations, responsive devices or economically transferable units.

The next snapshot changed the comparison. RFC 820, published in January 1983, explained that an old network number could remain temporarily in its tables with a T marker while a network moved to a new number. Its Class A table marked 36 as an old number for Stanford. The Class B table on printed page 4 listed 128.12 as the Stanford University network. AMPRNET remained at 44 without a transition marker.

January 1983 is the date of that documentary snapshot, not a certified completion date for Stanford’s migration. The two rows nevertheless establish a useful distinction. Stanford’s large numeric position was already being treated as transitional, while AMPRNET’s was still current in the register. One early entry was moving into a Class B network with (2^16), or 65,536, nominal local combinations—one 256th of the Class A field. The other continued at Class A scale.

This is the first boundary of the first-mover dividend. Early placement created the possibility of retaining a coarse and potentially durable address resource. It did not make retention automatic. The register itself records renumbering, disappearance, changed programme labels and repurposing.

The second boundary appears much later. Network 44 eventually produced a disclosed financial consequence when part of it was transferred in 2019. Stanford’s row supplies a contemporary counterexample to the idea that every early Class A entry became a permanent endowment. The comparison is stronger than a present-value exercise because it begins with observed administrative divergence rather than assuming that nominal size became wealth.

To move from timing to advantage, at least two additional links are required. The historical row must connect through a defensible chain of administration or control to a later actor. That continuity must then produce an identified consequence: usable retained capacity, work or procurement plausibly avoided, disclosed transaction income, or a conditional option over future disposition. Without both links, the evidence supports an early entry and no more.

Reconstructing the denominator from the printed register

The cohort begins with numeric positions, not with a retrospective list of famous institutions.

RFC 790’s Class A table specifically names positions 1-12 and 14-44. Position 13 is printed as unassigned. The inclusive arithmetic is straightforward: twelve positions in 1-12, plus thirty-one in 14-44, produces 43 specifically named positions.

The page contains a printing contradiction at its lower boundary. Immediately after assigning 44 to AMPRNET, RFC 790 prints the unassigned range as 44-126. Counting 44 as both assigned and unassigned would make the table internally incoherent. The narrow normalisation used here gives precedence to the specific 44/AMPRNET row and treats the following unassigned range as beginning at 45. RFC 820’s continued, specific AMPRNET entry at 44 supports that reading. This is an editorial correction of an overlapping printed range, not a claim about an unrecorded application or property right.

Those 43 units are numeric Class A positions named in the September 1981 snapshot. They are not 43 organisations. Several labels describe programme networks, experiments, testbeds or subnets, and several contacts recur. Nor are they 43 holdings, routed blocks or assets. The register records number-to-network mappings for protocol administration.

RFC 820 provides the next comparable table, but its own total has a different construction. On printed page 6, the RFC reports 31 Class A entries, 24 Class B entries and 1,042 Class C entries, or 1,097 entries in total. Its category subtotals are 26 research, four defence and one commercial Class A entries. The 31 Class A count includes ten explicitly transitional old-number rows: 1, 2, 5, 6, 9, 11, 36, 45, 47, 52. It is therefore a count of snapshot-table entries, including transition aids, rather than 31 permanent inventories.

For the RFC 790 cohort, the inclusion rule for continuity is deliberately mechanical: retain a position if RFC 790 gives it a specific name and RFC 820 also gives the same numeric position a specific name, whether or not the label changed and whether or not RFC 820 marks it T. The resulting 25-position intersection is:

1-6, 8-11, 14, 18, 21, 23, 25-28, 30, 32, 35, 36, 39, 41, 44.

The arithmetic can be checked by grouping the ranges: six positions in 1-6; four in 8-11; two singleton positions, 14 and 18; two more, 21 and 23; four in 25-28; and seven singletons, 30, 32, 35, 36, 39, 41 and 44. That gives (6+4+2+2+4+7=25).

The other 18 RFC 790 positions—7, 12, 15-17, 19, 20, 22, 24, 29, 31, 33, 34, 37, 38, 40, 42, 43—are unassigned in RFC 820. Within the 25-position intersection, seven positions—1, 2, 5, 6, 9, 11, 36—carry RFC 820’s T marker. The September 1981 denominator can therefore be reconciled at the next snapshot as 18 named positions still present without a transition marker, seven present as old-number transition rows, and 18 no longer named: (18+7+18=43).

That accounting describes change in the register over sixteen months between publication dates. Repetition identifies register continuity rather than continuous legal control. Disappearance identifies the absence of a named entry in the later table rather than a dated return transaction. A T identifies an old number retained for transition while leaving the migration’s start, completion and administrative return dates unresolved. These source boundaries also govern the later cases: present registries establish later administrative states, routing data establish visibility, active probes establish responsiveness, and financial filings establish recognised transactions.

The denominator matters because selected later outcomes are otherwise easy to mistake for the norm. Network 44 is exceptional precisely because its later institutional and financial evidence is unusually rich. It belongs inside a cohort in which almost half of the September 1981 positions were no longer specifically named by January 1983, and in which seven surviving positions were already marked as old numbers.

What the four featured rows actually establish

The compact ledger below separates historical identity, date provenance and later consequence. The IANA comparison uses the IPv4 Address Space registry accessed on 11 July 2026 and last updated on 10 October 2025. Its dates and designations describe present top-level administrative lineage. They are not substitutes for the earlier RFC rows.

Position	Historical row and nominal scale	Entity resolution and date status	Later administrative and operational evidence	Succession and disposition finding
44	RFC 790, September 1981, Class A table, printed page 4: AMPRNET, “Amature Radio Experiment Net.” RFC 820, January 1983, Class A table, printed page 4: AMPRNET, research, no T. Class A provides (2^24), or 16,777,216, nominal local combinations.	Matching AMPRNET across the two RFCs is high-confidence register resolution. The publication dates are snapshots. ARDC’s later account assigns the request to Hank Magnuski in 1981, but the original request and decision file is unavailable; that date is institutionally reported rather than independently reconstructed from an application. AMPRNET-to-ARDC resolution is medium-high because the succession account comes from ARDC and its audited statements, not an original title ledger.	The IANA registry, last updated 10 October 2025, designates 44/8 as legacy space administered by ARIN and prints 1992-07; that is a later registry date, not the original allocation date. ARDC’s 44Net record, observed on 11 July 2026, identifies the retained network as 44.0.0.0/9 and 44.128.0.0/10 and describes its amateur-radio research and experimentation purpose. No independent BGP utilisation measurement is used for the featured-row finding.	ARDC reports informal volunteer administration followed by California nonprofit incorporation in 2011 and formal management succession. Its 2019 filings document disposition of 4,194,304 addresses to Amazon. The financial consequence is established for 44.192.0.0/10, not for the full /8 or the retained ranges.
36	RFC 790, September 1981, Class A table, printed page 4: SU-NET, “Stanford University Ethernet.” RFC 820, January 1983, Class A table, printed page 4: SU-NET, research, marked T; RFC 820’s Class B table on the same printed page lists 128.12 as the Stanford University network. The old Class A position had (2^24) nominal combinations; the Class B row had (2^16).	Resolution to Stanford University is high confidence because both RFCs spell out the institution and use the same short name. September 1981 and January 1983 are snapshot dates. The original allocation date, transition start, completion date and exact return date remain unknown.	The current IANA registry lists 36/8 as allocated to APNIC with the date 2010-10. That current top-level state belongs to a later administrative lineage, and no later Stanford routing or control state is inferred from it.	RFC 820 directly documents an old-number transition from 36 to 128.12. It supplies the principal limit on permanent-endowment assumptions and establishes renumbering in progress; completion, successful migration, formal return and monetary outcome remain outside the record.
18	RFC 790, September 1981, Class A table, printed page 3: LCSNET, “MIT LCS Network.” RFC 820, January 1983, Class A table, printed page 3: MIT, research, no T. The Class A local field contains (2^24) nominal combinations.	Matching MIT’s Laboratory for Computer Science to the RFC 820 MIT label is high confidence. Continuity from those rows to every part of the later NET-18 inventory is less secure because the RFCs are not a complete chain-of-control record. The earliest independently established date used here is presence in the September 1981 snapshot; an exact original allocation date remains unknown.	The current IANA registry designates 18/8 as legacy space administered by ARIN and prints 1994-01, a later administrative date. The MIT Student Information Processing Board later reported service configurations, renumbering work and already-transferred portions. That stakeholder account establishes observed operational effects within its stated scope.	SIPB reported in 2017 that portions had already moved to Amazon and that MIT planned to sell about half of NET-18. A complete prefix list, all buyers, every transfer date, consideration and policy status are unavailable in the reviewed evidence, leaving completed scope and proceeds unquantified.
10	RFC 790, September 1981, Class A table, printed page 3: ARPANET. RFC 820, January 1983, Class A table, printed page 3: ARPANET, research, no T. The row had (2^24) nominal local combinations.	Programme-label resolution across the two snapshots is high confidence. No corporate successor is assigned. The RFC publication dates remain snapshots rather than an original allocation date or disposition date.	RFC 1918, a Best Current Practice published in February 1996, identifies 10.0.0.0-10.255.255.255 as private-use space that enterprises may use without coordination, with uniqueness only inside the participating private network. The current IANA registry lists 10/8 as reserved for private use with the date 1995-06.	The later outcome is technical repurposing into shared private-use space rather than a recipient transfer, sale, retained holding or monetised dividend.

The date column is as important as the identity column. CAIDA’s IPv4 exhaustion analysis explains that, during a 1993 effort to update the IANA assignment file, many historical legacy records were assigned an August 1993 date because that was when precise records began. Such a normalised database date describes a cleanup boundary. It cannot replace an original application, decision or allocation date.

The same caution applies to dates printed in the current IANA registry. The registry’s 1992-07 designation for 44/8, 1994-01 for 18/8 and 2010-10 for 36/8 coexist with earlier RFC evidence. They plainly are not the first appearance dates of the historical rows. The registry is valuable for present administrative lineage, including its distinction between legacy, allocated and reserved space. It is not a frozen account of control between 1981 and 1997, and it does not resolve sub-prefix transactions or institutional succession on its own.

A cross-sector register with many broken chains

The early cohort was not a university property list. Its labels covered packet-radio systems, satellite networks, defence communications, public data networks, commercial services, contractors, research laboratories and university facilities. The uneven survival of the records is part of the result.

Position 14 illustrates a commercial label change without a proven corporate chain. RFC 790’s printed page 3 calls it TELENET. RFC 820’s printed page 3 calls the same position PDN, “Public Data Network,” and marks it commercial. The repeated position belongs in the 25-position intersection. The changed label, standing alone, leaves unresolved whether the alteration was a rename, programme reclassification or reassignment.

The defence rows are similarly heterogeneous. Position 21 is EDN, “DCEC EDN,” in both snapshots; RFC 820 classifies it as defence. Position 26 changes from AUTODIN-II in RFC 790 to MILNET in RFC 820, also under the defence marker. The register supports a changed programme label at a stable numeric position. A legal or administrative succession from one programme to another would require records beyond these tables. The current IANA designation for 26/8 points to the Defense Information Systems Agency, but a present top-level registry name cannot supply every intermediate control event.

Research entries range far beyond campus networks. Position 28 remains WIDEBAND, the Wide Band Satellite Network, between the snapshots. Position 25 remains RSRE-PPSN, associated with the Royal Signals and Radar Establishment packet-switched network. Position 41 remains BBN-LN-TEST, a BBN local-network testbed. The repeated labels establish documentary persistence through January 1983 while leaving later alienability, route visibility and financial value unresolved.

Position 41 also demonstrates why a numeric identifier should not be converted into a timeless institutional asset. The current IANA registry lists 41/8 as allocated to AFRINIC with a date of April 2005. That later top-level allocation is administratively distinct from the BBN testbed row. The documentary discontinuity separates the two uses of the same numeric position.

Even apparently stable names can mask different kinds of organisation. AMPRNET described an experimental community, not a conventional company. ARPANET was a programme network, not an asset held by a successor corporation. WIDEBAND described an operating system or programme. MIT, Stanford and BBN can be resolved to institutions, but the existence of an institutional name says nothing by itself about uninterrupted legal control over every later sub-prefix.

This makes the detailed outcome sample source-bounded. Four rows receive deeper treatment because later public records illuminate their fate. Several non-university rows are included to preserve the cohort’s composition and to show where chains break. The study does not trace all 43 positions to modern controllers. Where succession remains undocumented, the nominal 1981 position stays outside the count of later institutional dividends.

The January 1983 outcome accounting is more comprehensive than the later financial tracing. All 43 September 1981 positions enter the denominator; 25 recur, seven of those recur as transitions, and 18 disappear from the named Class A table. Later consequence evidence is necessarily narrower because surviving applications, control records, route measurements and financial statements are uneven. The cohort’s observed result is therefore a combination of persistence, transition and documentary disappearance rather than a uniform legacy endowment.

Network 44: where continuity became disclosed proceeds

AMPRNET offers the strongest demonstrated dividend because the historical record, institutional succession and financial disposition can be examined separately.

RFC 790 establishes the network-44 row by September 1981. RFC 820 retains AMPRNET in January 1983 and classifies it as research. Neither document names ARDC, which did not yet exist, or sets out a property instrument. The earliest documents establish the experimental network and its responsible contact in the assigned-number register.

ARDC’s public institutional history adds the later chain. It says Hank Magnuski requested the space for licensed amateur-radio operators; volunteers informally administered the block as AMPRNet and later 44Net; those volunteers founded ARDC as a California nonprofit in October 2011; and the organisation formally assumed management. ARDC’s audited financial statements for 2019 independently describe it as a California nonprofit established in 2011 after operating as an unincorporated association of amateur-radio operators.

This evidence supports an institutionally reported succession from a volunteer association into a nonprofit administrator. It is stronger than a bare modern registry name because it explains the organisational change and is consistent with the audited institutional account. Its source boundary remains clear: the original 1981 application, a contemporaneous title instrument and an independent ledger of every intervening management decision are unavailable.

ARDC’s 44Net record, observed on 11 July 2026, identifies the retained network as 44.0.0.0/9 and 44.128.0.0/10. Together those ranges contain 12,582,912 addresses. The total follows from the prefix sizes: a /9 contains (2^23), or 8,388,608, addresses, while a /10 contains (2^22), or 4,194,304. ARDC describes 44Net as supporting amateur-radio scientific research and experimentation with digital communications. It also reports that the community had never used more than half of the original /8. These are the administrator’s statements about mission use and capacity; no independent census of routed or occupied addresses is used here.

The financial event is documented more precisely. ARDC’s 2019 Form 990, Schedule N, identifies a sale of 4,194,304 IPv4 addresses to Amazon Technologies, Inc., dated 19 July 2019, with a reported amount of $109,051,904. The disposed range was 44.192.0.0/10, one quarter of a /8.

The return’s revenue schedule reports $109,051,904 in gross proceeds, a $545,260 transaction expense identified in the audited statements as a broker commission, and a net recognised gain of $108,506,644. The arithmetic reconciles: (109,051,904-545,260=108,506,644).

Dividing gross proceeds by 4,194,304 addresses yields exactly $26 per address. This is gross-proceeds arithmetic for the disclosed 2019 transaction. The denominator is the disposed /10, not the original /8 or the retained ranges. The result is not a market index. It carries the transaction’s date, prefix size, buyer, registry context, routing reputation, due diligence and negotiated terms.

The administrative records contain a small but useful date distinction. ARDC’s tax filing gives the disposition date as 19 July 2019. ARIN’s record for 44.192.0.0/10 lists the range as a direct allocation to Amazon.com, Inc., with a registration and last-updated date of 18 July 2019. The two dates describe different records: one is the registry event, the other the disposition reported in the tax filing. The one-day difference should be preserved rather than collapsed.

The ARIN record establishes the registered recipient, exact range, direct-allocation type and registration date. The particular policy provision under which the change was approved remains outside the public administrative record examined here; identifying it would require the corresponding transfer-log entry or approval record. Registration status and financial disposition are independently documented.

The transfer turned one quarter of a surviving legacy resource into a board-designated endowment. ARDC’s audited statements say the proceeds were intended for grants and other activity supporting amateur radio and digital communications, while the remaining addresses continued to be available for amateur-radio operators.

That consequence was not contained in the original row. It depended on continued administration of the resource, formalisation of the volunteer group, later registry recognition, the ability to separate a /10, a buyer prepared to acquire it and a transaction whose proceeds accrued to the nonprofit. A different break in that sequence—early renumbering, loss of control, technical reservation or absence of a buyer—would have produced a different outcome.

The transaction also sets a strict limit on valuation. Applying $26 to ARDC’s retained space would treat mission use, marketability and transaction timing as if they were identical to the sold /10. Applying it to MIT, Stanford, defence programmes or the full 43-position cohort would ignore their different chains and dispositions. ARDC’s figures measure one realised event.

ARDC’s filings describe the addresses as received without charge when no discernible market value existed. The later scarcity and transaction therefore demonstrate path-dependent value rather than original investment intent: a technical record persisted into an environment in which part of it could be exchanged for money. The affirmative finding is a documented sequence from register persistence through institutional continuity to $109,051,904 in gross proceeds and a $108,506,644 recognised gain.

Stanford: a large row already on the way out

Stanford’s 36 entry supplies the closest contemporary control case.

RFC 820 explains its T marker before presenting the table: old network numbers were retained temporarily to ease transition when assigned numbers changed. It then marks Class A position 36 as old and lists 128.12 for Stanford in the Class B table. The evidence is contemporaneous, direct and administrative.

What changed is the registered network scale. The old Class A format carried 16,777,216 nominal local combinations; the Class B format carried 65,536. The ratio is 256 to one. Those figures are capacities defined by the address formats, not measured Stanford hosts.

The transition row establishes that a recipient named in the 1981 Class A cohort could move to a much smaller network early in the period. The exact original allocation date, the date on which every Stanford system stopped using 36, migration cost, service interruption and any formal return instrument remain outside the RFC evidence.

The value of the row lies in its timing. Renumbering was not invented after address scarcity or transfer markets. The January 1983 register already had a procedure for keeping old and new numbers visible during a change. That procedure implies coordination work without quantifying it. Stanford therefore supports two bounded findings: movement to a smaller class was administratively possible, and the transition required enough accommodation for the old number to remain temporarily in the published table.

Stanford was one institution with one documented transition, so it cannot establish that every early Class A recipient could have been reassigned a Class B without consequence. Its observed discontinuity is still decisive for the cohort: by January 1983, one large university row had already become a temporary old number while a much smaller current row carried the institutional network forward.

MIT: operational consequences without a public transaction total

MIT’s record persisted through the two early snapshots. RFC 790 named LCSNET, the MIT Laboratory for Computer Science network, at 18. RFC 820 broadened the label to MIT and retained the position without a transition marker. The two documents establish an institutional association by January 1983, but they do not supply a complete later prefix inventory or disposition history.

The MIT Student Information Processing Board account provides direct evidence of what one technical stakeholder reported in 2017. It says MIT had previously treated NET-18 as its address space, planned to sell about half, and had already transferred portions to Amazon. It also describes campus systems configured on the assumption that NET-18 identified MIT.

SIPB reported several concrete categories of operational disruption. Access-control arrangements that accepted the whole range as an MIT identity boundary became unsafe once addresses outside MIT could occupy parts of NET-18. A library staff account quoted by SIPB said the entire range had been registered with hundreds of publishers of licensed electronic resources. SIPB also expected one of its services to contact about 100 users concerning Internet records for roughly 300 websites during migration.

Those figures are stakeholder-reported populations for specified service categories, not a complete host census or a monetary cost estimate. They show why address continuity could matter operationally: a prefix may be embedded in licences, allowlists, DNS, service configurations and assumptions about institutional identity. Renumbering reaches beyond changing an interface address.

The same page reports a concurrent plan to place much of the campus behind network address translation and describes building-by-building service problems during that change. It does not establish that every NAT decision was caused by every address transfer. The defensible description is that address dispositions, renumbering and service reconfiguration were concurrent changes reported by SIPB, with overlapping operational effects.

SIPB was a knowledgeable affected stakeholder rather than MIT’s complete registry, sale contract, audited financial statement or transfer ledger. The page supplies neither a final list of all transferred prefixes nor all transaction dates, buyers, policy classifications or consideration. Its phrase “about half” describes the announced plan and is not converted here into a completed-half total.

MIT’s realised proceeds therefore remain unknown. The available account establishes that some portions were reported as already transferred, that additional disposition was planned, and that address assumptions created identifiable migration work. The strongest finding is operational dependence and disposition activity rather than a quantified windfall.

MIT also clarifies the difference between retained capacity and avoided cost. Long possession of NET-18 plausibly gave the institution room for public addressing and reduced the need to obtain equivalent space elsewhere. Measuring the avoided amount would require a specified alternative: provider-assigned space, acquired IPv4, private addressing, translation, IPv6 or a smaller public prefix. No reviewed record prices that counterfactual.

What can be observed is the work that surfaced when continuity changed. Publisher registrations, access controls, DNS records, user coordination and service architecture had to be reconsidered. These categories make avoided renumbering a credible form of benefit during the period of stability. MIT’s documented discontinuity turns operational dependence from an abstraction into identifiable work across licences, identity boundaries, service records and user migration.

Network 10 left the recipient frame altogether

Network 10 follows neither the ARDC nor the Stanford path.

It appears as ARPANET in the Class A tables of RFC 790 and RFC 820. By February 1996, RFC 1918 included 10.0.0.0/8 among three blocks reserved for private internets. Enterprises could use those addresses without coordinating with IANA or a registry, provided they accepted that the addresses were unique only within the relevant private environment and were not propagated as public inter-enterprise routes.

That outcome is a form of infrastructure value, but not recipient wealth. The block became a shared technical convention used across unrelated private networks. No successor to the ARPANET row is credited with a 16.7-million-address holding. No sale or lease follows from the reservation.

Network 10 therefore matters to the cohort argument because it shows how a large early numeric position could acquire enormous later utility while ceasing to be an exclusive recipient-centred resource. Technical value and monetisable control are different consequences.

By 1992, the allocation bargain had changed

The early snapshots belong to a small, centrally administered network environment. Later policy documents show a different problem: scarcity, registry distribution, address conservation and routing scale had become explicit administrative concerns.

RFC 1366, published in October 1992 as an informational memo rather than an Internet standard, reports “Network Number Statistics” dated June 1992. The table’s units are classful network numbers. It lists 49 of 126 Class A network numbers allocated, 7,354 of 16,383 Class B numbers, and 44,014 of 2,097,151 Class C numbers. The corresponding printed percentages are 38%, 45% and 2%.

These are not address counts. One Class A network number and one Class C network number occupy radically different fractions of IPv4 space. Nor are they counts of organisations, applications, requests, active routes or later outcomes. The table is a June 1992 classful-number baseline published four months later.

RFC 1366’s assignment sections show how much more documentation had entered the process. A new Class A petitioner was expected to provide detailed technical justification about network size and structure, with assignment left to IANA’s discretion. A Class B applicant was generally expected to document more than 32 subnets and more than 4,096 hosts, subject to case-by-case treatment where Class C blocks were unsuitable. Class C quantities were tied to a 24-month projection.

The memo also exposes the conservation-versus-routing trade-off. Moving medium-sized applicants toward multiple Class C networks preserved Class A and Class B numbers, but the document warned that proliferation of Class C numbers would accelerate the growth of routing information. Granularity conserved address space while creating more registry entries and potentially more top-level routes under the then-current routing arrangements.

RFC 1519, published on the Standards Track in September 1993, responded with classless allocation and aggregation. Its central design joined two changes: distribute future allocations along routing topology and aggregate multiple destinations behind network-and-mask pairs.

Crucially for the incumbent comparison, RFC 1519 states that the plan neither required nor assumed reassignment of addresses already issued, although reassignment could have reduced routing-table size further. It encouraged renumbering in circumstances such as changing providers, but did not make wholesale renumbering of the existing Internet a prerequisite for deployment.

That was a rational transition choice. Preserving existing addresses reduced immediate disruption. It also meant that early recipients with surviving records could keep their historical address independence while later recipients increasingly entered through provider-based, aggregatable space. The policy preserved operational stability and, incidentally, preserved some early options.

By November 1996, RFC 2050 described a mature evidentiary regime. It was issued as Best Current Practice 12. The IESG note said it represented the registries’ current assignment practice, while expressly withholding endorsement or recommendation of the policy and anticipating reconsideration.

RFC 2050 distinguished allocation to an Internet service provider from assignment to an end enterprise. It described slow-start allocation for new providers, minimal initial amounts based on immediate requirements, and additional space following utilisation verification. Projected customer numbers had little influence without demonstrated requirements.

For assignments, its common criteria were 25% immediate utilisation and 50% within one year. Section 3.6 defines the utilisation rate’s numerator as hosts connected to the network and its denominator as the total hosts possible on that network. The percentages were guidelines, not immutable thresholds. Topological conditions could justify exceptions, one-year utilisation did not always have to fall exactly at the stated level, and applicants were expected to document a high-confidence projection.

The document also allowed registry-specific refinements. It called the guidelines a common operational base while acknowledging additional regional requirements. It would therefore be inaccurate to treat the 25% and 50% figures as identical implementation by every registry, or as criteria retroactively imposed on AMPRNET, Stanford, MIT or ARPANET.

The late entrant faced a different administrative contract. A service provider in 1996 was expected to document immediate need, use classless technology, accept incremental allocation and supply utilisation evidence. A research network visible in 1981 entered when fixed address classes and central assignment shaped the available choices. The comparison identifies a historical asymmetry, not a controlled experiment: the applicants, technology, routing environment, institutional structure and demand were all different.

Later concentration is context, not a verdict on the 1981 rows

Address concentration can be measured in later registries, but the result depends heavily on the population and entity method.

CAIDA’s IPv4 Address Space Concentration analysis uses a full ARIN WHOIS dump dated 31 August 2005. It defines an allocation event as all blocks assigned to one OrgID on the same date. An “old player” is an organisation with multiple allocation events by the observation point; a “new player” has one. This is not a classification by age in years and does not identify the RFC 790 cohort.

The analysis covers allocated address space, not routed or occupied space. It explicitly distinguishes the three: allocated space appears in registry records; routed space appears through BGP; occupied space has been assigned to hosts, routers or other devices and is only partly observable through measurement.

CAIDA pruned IANA-direct /8 assignments to end sites, including MIT’s 18/8, as well as DoDNIC-managed and JPNIC-managed allocations. The resulting ARIN-derived population covered the equivalent of 42.6 /8s after overlapping records were resolved. This exclusion alone prevents the concentration percentages from being used as a measurement of the early Class A cohort.

Dataset 1 retained fine-grained customer records. It contained approximately 1.07 million OrgID values and about one million overlapping allocated blocks. At the August 2005 endpoint, old players held 56.4% of the allocated-address-space denominator while constituting fewer than 2% of the recorded organisations.

CAIDA then removed 883,000 specially identifiable customer records to create Dataset 2. That population contained approximately 188,000 OrgID values and roughly 250,000 overlapping allocated blocks, covering essentially the same address space apart from a reported 0.026% discrepancy. Under this treatment, old players held 63.4% of the allocated space and represented 11.2% of organisations.

The two space percentages are methodological bounds. Including customer records attributes suballocated space to many customer organisations and produces the 56.4% lower estimate of concentration. Removing those records treats customers as part of the provider organisation and produces the 63.4% upper estimate. Neither percentage measures motive, favour, ownership, route visibility or original need. Both depend on OrgID resolution and overlapping-prefix attribution.

The finding still matters. By the 2005 snapshot, repeated allocation recipients held a majority of the studied allocated space under either customer treatment. The difference between 56.4% and 63.4% also shows why concentration cannot be discussed without specifying what counts as an organisation.

For the legacy first-mover question, the study supplies aggregate context rather than row-level proof. Its population excludes direct legacy /8 end sites such as MIT and therefore cannot reconstruct a particular RFC 790 request or its later chain. Its affirmative contribution is narrower and still material: under both customer-record treatments, repeated recipients controlled a majority of the allocated-space denominator in the studied 2005 ARIN-derived population.

What the transfer evidence measures

The later transfer market provides another aggregate lens, with different units and exclusions.

The CAIDA-associated paper On IPv4 transfer markets: Analyzing reported transfers and inferring transfers in the wild, published in 2017, analysed transfer lists published by ARIN, APNIC and RIPE through September 2015. LACNIC and AFRINIC were not part of the reported-transfer population used for the main analysis. The authors excluded 26 ARIN transactions involving space reserved for Internet exchange points and 111 APNIC transactions that moved resources within the same organisation.

Within that defined population, legacy space accounted for 63.82% of total transferred address-space volume. The denominator is the quantity of addresses represented in the retained reported transfers. It is not the number of transactions, blocks, sellers or recipients; it is not every transfer that may have occurred outside public reporting; and it is not the full legacy or RFC 790 cohort.

The paper assessed post-transfer routing visibility using BGP data spanning 2004 through 2015. It defined six classes. A-I was never routed during the study; A-II stopped being advertised at least two years before the reported transfer; and B appeared only before the transfer. C-I appeared only after transfer, C-II had been absent for at least two years before transfer and reappeared afterwards, and D appeared before and after.

On an address-space-volume basis, 94% of the transferred space fell into the post-transfer-visible classes C-I, C-II and D. The paper’s 85% result covers C-I and C-II: space routed for the first time after the transfer or routed again after an absence of at least two years. These percentages describe BGP visibility classes in the reported-transfer population at the study’s observation horizon.

A BGP advertisement reveals that a route to a prefix was visible through the measurement system. Ownership, physical occupation, legal transfer, service quality and the number of active addresses require different evidence. Origin changes can arise from traffic engineering, service-provider changes, internal reorganisations or routing incidents. The paper itself found that BGP-based transfer inference produced substantial false positives until supplemented with registry, organisational and DNS evidence.

The authors therefore ran a separate active-measurement analysis. Their IP census data spanned November 2009 through December 2015 and probed allocated IPv4 addresses with ICMP echo requests every two to four months. For each transferred prefix and census snapshot, the measured response fraction was the number of addresses replying divided by the number of addresses in the prefix.

They selected each prefix’s maximum response fraction in windows of three, six, nine and twelve months before and after its reported transfer, then calculated regional medians. Comparing the twelve-month windows, the paper reported an increase of at least 50% in the median response fraction across each studied RIR.

This is evidence of a change in ICMP responsiveness under the paper’s method. Hosts configured not to reply, intermittent addresses, internal use, services behind translation, dark infrastructure and non-ICMP activity remain outside that observation. The measurement neither supplies title nor transaction value. Routing visibility and ICMP response are complementary observations rather than a combined measure called “routed and used.”

The aggregate evidence challenges a static-hoarding story. Much of the address-space volume in reported transfers was visible in BGP after transfer, and the active-measurement response fraction generally increased. The measured discontinuity is affirmative: within the defined three-RIR population and observation windows, transferred prefixes became more externally observable through both routing and active-response measures.

Turning an early row into an economic consequence

The governing states have to remain distinct. An address range can be allocated by a top-level authority, delegated down a hierarchy, registered to a named organisation, routed through BGP, visible at particular collectors, responsive to a probe, occupied by network devices, operationally used for services, controlled by an administrator, treated as owned under a legal claim, deemed transferable under registry policy, leased, sold, and finally monetised through recognised proceeds. Evidence at one state does not automatically move the range into the next.

The cases establish different points along that chain. The early RFCs supply registered mappings in publication snapshots, and RFC 820 records Stanford’s documentary transition from an old Class A number to a Class B row. The current IANA table supplies present top-level administrative designations. ARIN’s network record identifies Amazon as the registered recipient of the disposed 44/10. CAIDA’s BGP and ICMP analyses observe visibility and responsiveness in aggregate populations. ARDC’s filings carry the evidence further by recording a transaction and its recognised financial result.

In the early classful setting, the immediate benefit of a Class A row was architectural room. A programme or experiment received one network identifier with a large local field in which to design subnets. That arrangement could simplify network planning even when the nominal field was never fully occupied. Missing applications and host forecasts leave the recipient-specific degree of necessity unresolved, but the assigned-number format itself establishes the scale and administrative convenience made available.

For AMPRNET, that initial technical position persisted into retained operational capacity. ARDC continues to administer two named ranges for amateur-radio networking and experimentation. The later nonprofit did not merely inherit a historical label; it managed a separable resource whose mission use continued after one quarter of the original /8 was disposed of. The operational and financial consequences coexist in the same case without becoming interchangeable.

MIT shows how retained capacity could become embedded in institutional systems. NET-18 appeared in access rules, publisher registrations, DNS records, service configurations and assumptions about institutional identity. During the period of stability, those dependencies made continued addressing an operational benefit. When portions changed hands and services were reconfigured, the same dependencies generated migration work. The evidence identifies the affected categories but supplies no monetary total for capacity acquired elsewhere or labour avoided during earlier years.

ARDC’s 2019 disposition crosses the further threshold into realised income. The disclosed /10, buyer, registration date, disposition date, gross proceeds, broker commission and recognised gain make the financial consequence auditable. MIT’s reported transfers remain on a different evidentiary footing because the reviewed materials do not disclose complete scope or consideration. No featured case supports a claim of audited lease income.

The ability to retain, subdivide or dispose of space also created a conditional option. Its value depended on registry recognition, operational encumbrance, prefix reputation, fragmentation, buyer demand and timing. ARDC exercised part of that option through one documented sale while retaining mission capacity. Stanford’s early transition and network 10’s later private-use reservation show that the same nominal Class A scale could instead lose its recipient-centred optionality through renumbering or technical repurposing.

The chronology matters more than any tidy taxonomy. A row that eased experimentation in 1981 could support operations decades later, impose migration work when continuity changed, or generate proceeds only after transfer institutions emerged. The documented cases therefore connect early timing to heterogeneous outcomes: Stanford’s administrative transition, network 10’s shared technical utility, MIT’s operational dependence and ARDC’s realised financial consequence.

A more granular beginning would have shifted costs, not erased them

One period-feasible alternative would have assigned smaller initial networks and expanded or renumbered recipients as demonstrated needs grew.

RFC 820 shows that such movement was administratively possible. Stanford moved from the old 36 position toward 128.12. Packet-radio systems at positions 1, 2, 5, 6 and 9 also carried transition markers, with several replacement Class B rows present in the same register. The evidence establishes a transition mechanism while leaving its price and success rate unmeasured.

Starting smaller would have conserved scarce numeric capacity. It also would have created more registry transactions and, under classful routing, potentially more independently visible networks. RFC 1366 explicitly describes the tension: substituting multiple Class C numbers for scarce Class B numbers conserved the larger class while accelerating the growth of routing information.

RFC 1519 later made finer-grained allocation more workable by joining variable-length prefixes to route aggregation. It also identified limitations. Multi-homed sites could require explicit routes, and customers that changed providers without renumbering could punch holes in aggregates. Smaller assignments carried bookkeeping and routing effects shaped by topology, protocol deployment and willingness to renumber.

Host and service work would also have moved forward in time. RFC 820’s old-number rows show the need to preserve transition information. RFC 1918 later identifies changes to host addresses, DNS and configuration references as consequences of moving between address regimes. MIT’s stakeholder account gives concrete later examples involving licences, access controls, service records and user coordination.

Together, these sources establish plausible mechanisms: additional registry labour, more route state under some architectures, host renumbering and service breakage. They do not measure a cohort-wide cost, so no defensible monetary estimate can be assigned to the 43 positions.

The alternative might still have produced a more conservative distribution. Stanford proves that at least one early university network could be represented in a Class B row by January 1983. The wider cohort contains satellite, packet-radio and defence networks whose corresponding transition requirements remain unmeasured. The observed result is bounded but affirmative: smaller reassignment was feasible in at least one documented early institutional case, and the register provided an explicit transition mechanism to support it.

Review and return would have required an institution capable of enforcing it

Another period-feasible design would have allowed a large initial assignment but required periodic documentation of use and partial return where capacity no longer served the programme.

This would have preserved the planning simplicity of a contiguous block while creating an administrative checkpoint. It would also have required answers to questions the early RFCs leave open: what counted as use for an experimental network, how uncertain research growth should be treated, which evidence the administrator could demand, how a block could be divided, and how affected systems would be renumbered.

The later policy documents show that such governance was conceptually available by the 1990s. RFC 1366 required technical justification for exceptional Class A requests and used projected need for Class C assignment. RFC 2050 made utilisation documentation, prior assignment history, network engineering plans and audit part of the assignment framework. It also contemplated invalidating assignments when the requirement ceased to exist, while calling for reasonable efforts to notify the organisation.

Those later rules describe an institutional response that evolved once scarcity and routing scale became central. They cannot be projected backward as obligations governing the RFC 790 rows.

A review-and-return system would have conserved capacity only if records were reliable and return decisions enforceable. It would have imposed recurring reporting and verification work on registries and recipients. Experimental programmes could face legitimate uncertainty about future topology. Separating an unused portion might affect addressing plans or routes elsewhere in the block. These are plausible policy costs rather than observed totals.

ARDC’s eventual disposition shows that a large block could be divided: one /10 was separated while a /9 and another /10 remained. That 2019 fact belongs to an environment transformed by CIDR, registry procedures and transfer institutions, so it cannot establish that identical division, policy recognition and operational separation were readily available in 1981.

MIT supplies a different warning. Portions of a large institutional range can become embedded in access rules and external registrations far beyond a network team’s direct inventory. A return or transfer process focused only on the registry row would miss those dependencies. Periodic review might have reduced excess capacity, but a credible process would also need transition procedures and a realistic account of service breakage.

The governance choice was therefore not between costless conservation and irresponsible abundance. More granular issuance shifted costs toward routing, registry work and renumbering. Large issuance with review shifted costs toward measurement, documentation, enforcement and partial-return transitions. The observed early system minimised some immediate administrative constraints and left a subset of recipients with unusually durable capacity.

The measured finding

The full cohort does not support a universal windfall story.

The September 1981 denominator contains 43 specifically named Class A numeric positions after resolving the printed overlap at 44. Only 25 positions remain specifically named in January 1983. Seven of those are marked as old numbers for transition. Eighteen positions from the original denominator are unassigned in the later Class A table.

The surviving positions cover research, defence, programme and commercial labels. Some names persist; others change. Several later institutional chains remain unknown. The present registry shows that numeric positions can move into entirely different administrative uses, as with 36/8 under APNIC and 41/8 under AFRINIC. Network 10 became shared private-use space. Stanford’s position was already transitional. These are substantive limits on recipient-centred valuation.

A realised first-mover dividend is clearest at network 44. The early AMPRNET row persisted; an institutionally reported volunteer chain became a nonprofit administrator; a specific /10 was separated; ARIN registered the range to Amazon; and public financial records disclose the buyer, date, address count, gross proceeds, commission and net recognised gain. The evidence measures one transaction rather than a nominal fortune.

MIT demonstrates a different consequence. Its stakeholder account records operational dependencies and migration work, while public evidence of the complete transfer scope and consideration remains incomplete. The case supports retained capacity and costly discontinuity more strongly than it supports a quantified financial gain.

Later aggregate evidence supplies context rather than a shortcut. CAIDA’s concentration analysis shows that repeated recipients held a majority of the allocated space in its 2005 ARIN-derived population under two customer treatments, while excluding direct legacy /8 end sites such as MIT. The transfer study shows that legacy space formed 63.82% of transferred address-space volume in its reported three-RIR population through September 2015 and that much transferred space became visible in routing and more responsive to ICMP measurement. Neither study reconstructs original need or continuous title for the 1981 cohort.

The distributional asymmetry is nevertheless real. Some early records survived from a coarse classful regime into a world of address scarcity, incremental allocation and transfer. Later entrants faced documented utilisation, provider-based aggregation, smaller increments and renumbering expectations. For the early recipients whose chains held, administrative continuity could become operational independence, work avoided, transaction proceeds or a strategic option.