Electronics Guide

ARPANET in Middle Age

The Advanced Research Projects Agency funded and directed ARPANET from its inception, and operational management passed to the Defense Communications Agency in 1975. Access remained tightly controlled: participation required a defense contract or government-funded research affiliation. That model grew untenable as the 1980s progressed. Commercial firms wanted access, international partners wanted connectivity, and a research network run on best-effort terms increasingly carried traffic that institutions treated as operational.

The underlying hardware also showed its age. ARPANET's packet switches, the Interface Message Processors, had begun as ruggedized Honeywell DDP-516 minicomputers and evolved through the Honeywell 316, the multiprocessor Pluribus, and finally the BBN C/30 and C/300. Backbone trunks ran at 50 kilobits per second. That capacity, generous in 1969, had become a constraint by the mid-1980s.

The MILNET Split

In 1983 the Defense Communications Agency divided ARPANET in two. Roughly two-thirds of the nodes moved to MILNET, which carried unclassified Department of Defense operational traffic and became a component of the Defense Data Network; classified traffic ran on physically separate networks at their own security levels. The remaining nodes stayed with ARPANET and served the research community. A small number of carefully controlled mail gateways relayed messages between the two, establishing a pattern of policy-driven segmentation that reappears throughout Internet architecture. MILNET later evolved into NIPRNET.

NSFNET and the Academic Backbone

The National Science Foundation, not the Defense Department, expanded network access beyond the ARPANET community. Beginning in 1985, NSF funded NSFNET, initially linking five NSF supercomputer centers and quickly growing into a national backbone that regional networks joined. The NSFNET acceptable use policy was more permissive than ARPANET's, admitting research and education traffic broadly while prohibiting purely commercial use. Institutions that had never qualified for ARPANET connected within a few years.

The original NSFNET backbone ran at 56 kilobits per second and saturated almost immediately. A partnership of Merit Network, IBM, and MCI rebuilt it in 1988 around T1 circuits at 1.544 megabits per second, using nodal switching subsystems assembled from clusters of IBM RT PC workstations running a modified Berkeley Unix. The T3 upgrade to 44.736 megabits per second, carried out from 1990 through 1991 with IBM RS/6000-based routers, delivered nearly a thirtyfold increase over the T1 backbone. Each step demanded forwarding hardware capable of far higher packet rates without sacrificing the reliability that users had come to expect.

Decommissioning ARPANET

ARPANET was formally decommissioned on February 28, 1990. By then NSFNET had assumed the role of primary research backbone, and regional networks supplied local connectivity, so the shutdown passed with almost no disruption. That smoothness was itself an achievement: it demonstrated that the network's function had been successfully abstracted away from any particular set of wires or switches.

Backbone Privatization

Privatization completed the transition. Through the early 1990s, commercial carriers built backbones of their own. NSF's 1993 solicitation restructured the architecture around a set of Network Access Points where providers would exchange traffic, a separate research network called the very high speed Backbone Network Service, and a routing arbiter to coordinate policy. The initial Network Access Points were operated by Sprint in Pennsauken, New Jersey; Ameritech in Chicago; Pacific Bell in San Francisco; and MFS in the Washington, D.C., area. The NSFNET backbone service ceased operation on April 30, 1995, and the Internet passed into the hands of competing commercial providers.

The Flag Day of January 1, 1983

ARPANET completed its cutover to TCP/IP on January 1, 1983, a coordinated transition remembered as the protocol flag day. Sites that had not converted lost connectivity until they did. The mandate established a single internetworking protocol suite across the entire research network and gave every subsequent development a common foundation. The inclusion of TCP/IP in Berkeley's 4.2BSD Unix release in 1983 spread the protocols across the academic computing world at essentially no cost, which mattered as much as the mandate itself.

The Layered Architecture

Layering proved essential to the design's success. The Internet Protocol offered only unreliable, connectionless packet delivery across network boundaries. TCP added reliable, ordered, flow-controlled byte streams on top; the User Datagram Protocol offered a minimal alternative for applications that preferred to manage their own delivery semantics. Because IP demanded so little of the networks beneath it, Ethernet, Token Ring, X.25, satellite links, and dial-up serial lines could all carry Internet traffic without changing the applications above.

Congestion control was retrofitted after hard experience. A series of congestion collapses on the NSFNET-era Internet in 1986 dropped throughput between sites at Berkeley by three orders of magnitude. Van Jacobson's response, published in 1988, added slow start, congestion avoidance, and round-trip-time variance estimation to TCP. These algorithms let the network's endpoints regulate their own sending rates without any signal from the routers, and they remain the conceptual basis of Internet congestion control.

The Domain Name System

Until the early 1980s, name-to-address mapping depended on a single HOSTS.TXT file distributed from the Stanford Research Institute Network Information Center. Paul Mockapetris specified the Domain Name System in RFCs 882 and 883 in November 1983 to replace it, and the definitive revision appeared as RFCs 1034 and 1035 in November 1987. DNS distributes the namespace hierarchically, delegating authority for each zone to servers operated by the organization that owns it, so the system grows without any central bottleneck. Caching resolvers absorb the majority of queries, keeping load on the authoritative servers manageable even as the namespace expands by orders of magnitude.

Addressing Pressure and CIDR

The original IPv4 addressing plan divided the 32-bit space into fixed classes: class A networks with roughly 16.7 million addresses each, class B with 65,536, and class C with 256. The scheme was wasteful. Organizations that needed a few thousand addresses took a class B and stranded most of it, and by 1992 projections showed the class B space exhausting within a few years while the global routing table grew faster than router memory.

Classless Inter-Domain Routing, specified in 1993, removed the class boundaries. Routes carry an explicit prefix length, so an allocation can be sized to need, and providers can aggregate many customer prefixes into a single advertisement. CIDR relieved both pressures at once and bought the time in which longer-term work proceeded: network address translation was described in 1994, and the IETF's next-generation IP effort selected the design that became IPv6, first specified in December 1995.

The RFC Process and the IETF

The Request for Comments series, begun with RFC 1 in April 1969, documented protocol specifications, design debates, and operational experience in a form anyone could read and implement. That openness allowed independent implementations from different vendors, which produced competition at the product level and interoperability at the protocol level.

The Internet Engineering Task Force, which held its first meeting in January 1986, gave protocol development an organizational home. The IETF has no formal membership and takes no votes; it works by rough consensus and running code, requiring that a proposal demonstrate independent interoperating implementations before advancing on the standards track. The approach favored specifications that could actually be built, and it moved considerably faster than treaty-style standards bodies. The Internet Society, founded in 1992, later provided the IETF with an institutional and legal home.

TCP/IP versus OSI

Through the late 1980s, the Open Systems Interconnection suite developed by ISO and CCITT appeared to be the officially sanctioned future. United States government procurement policy formalized that preference in the Government OSI Profile, and several European governments went further. OSI nonetheless lost. TCP/IP was already deployed, already free in Berkeley Unix, and simpler to implement, while OSI specifications were expensive to obtain and arrived years after the market needed them. By the early 1990s, OSI had receded to a few niches, though its seven-layer reference model survives as teaching vocabulary and its IS-IS routing protocol survives inside large carrier networks.

Electronic Mail and Usenet

Electronic mail was the Internet's dominant application for its first two decades. The Simple Mail Transfer Protocol, specified in 1982, moved messages between hosts, and the message format defined by RFC 822 in the same year still underlies email headers. Multipurpose Internet Mail Extensions, specified in 1992, added character sets beyond ASCII and attachments of arbitrary media types, making email a general-purpose file transfer mechanism almost by accident.

Usenet, established in 1980, distributed threaded discussions across a flooding network of cooperating hosts. Originally propagated by dial-up UUCP links, it moved to the Network News Transfer Protocol in 1986 and became one of the heaviest consumers of backbone bandwidth. Usenet demonstrated large-scale many-to-many discussion years before the Web, and its conventions of quoting, threading, and moderation influenced nearly every later online forum.

File Transfer, Archie, Gopher, and WAIS

Anonymous FTP servers held the Internet's public archives, but nothing indexed them. Archie, built at McGill University in 1990, solved that by periodically cataloging FTP directory listings and offering a searchable index. It is generally regarded as the Internet's first search engine.

Gopher, released by the University of Minnesota in 1991, presented Internet resources as a navigable hierarchy of menus. It was simple, fast on slow links, and briefly more popular than the Web. Wide Area Information Servers, developed by Brewster Kahle at Thinking Machines and released in 1991, offered full-text search across distributed document collections and returned ranked results, anticipating Web search engines by several years. In 1993 the University of Minnesota announced that it would charge commercial licensing fees for its Gopher server software. The announcement pushed developers and site operators toward the unencumbered Web, and Gopher traffic declined sharply thereafter.

The Morris Worm and the Birth of Coordinated Security

On November 2, 1988, a self-replicating program written by Cornell graduate student Robert Tappan Morris exploited flaws in the Unix finger daemon, sendmail, and weak passwords to spread across the Internet. A flaw in its own reinfection logic caused it to load infected machines until they became unusable. Contemporary estimates put the number of affected hosts at roughly six thousand, on the order of ten percent of the machines then connected.

The incident exposed how little capacity existed to respond to a network-wide emergency. Within weeks, DARPA funded the Computer Emergency Response Team Coordination Center at Carnegie Mellon University's Software Engineering Institute, creating the model of the coordinating incident response team that the security industry still follows. The worm also produced the first conviction under the United States Computer Fraud and Abuse Act.

The CERN Proposal

Berners-Lee circulated "Information Management: A Proposal" at CERN in March 1989. His supervisor, Mike Sendall, annotated it "vague, but exciting" and authorized further work. With his Belgian colleague Robert Cailliau, Berners-Lee refined the concept into a joint proposal in 1990 and began implementation on a NeXT workstation, one of the few machines then available with a development environment rich enough to build a graphical hypertext editor quickly.

HTML, HTTP, and URLs

Three inventions, all deliberately minimal, made up the Web. The Hypertext Markup Language described documents using a small tag vocabulary derived from CERN's existing SGML practice. The Hypertext Transfer Protocol defined a stateless request and response exchange so simple that its first version consisted of a single line of text. The Uniform Resource Locator gave every resource a global address combining a scheme, a host name, and a path.

The simplicity was strategic. Because HTTP was easy to implement, servers appeared for every platform within months. Because HTML tolerated errors and unknown tags, documents kept rendering as the language evolved. Because URLs could name resources reachable by FTP, Gopher, and news as well as HTTP, the Web could absorb the Internet's existing content rather than compete with it.

First Browsers and Servers

Berners-Lee's original browser, WorldWideWeb, was also an editor; it was later renamed Nexus to distinguish the program from the system. He wrote the first server software, CERN httpd, and put the first website online at info.cern.ch, describing the project itself. The Line Mode Browser, written by Nicola Pellow, provided text-only access from any terminal and made the Web usable on the dumb terminals then common at universities. Graphical browsers followed on Unix workstations: ViolaWWW from Pei-Yuan Wei at Berkeley and Erwise from a student team at the Helsinki University of Technology, both in 1992.

Release into the Public Domain

On April 30, 1993, CERN issued a document placing the Web software into the public domain, with a later release under an open license. Anyone could implement a server or browser without permission, royalty, or negotiation. The contrast with Gopher's licensing announcement the same year was stark, and the outcome decisive. To coordinate the technology's evolution without controlling it, Berners-Lee founded the World Wide Web Consortium at the MIT Laboratory for Computer Science in October 1994.

Mosaic and the Graphical Web

NCSA Mosaic, written by Marc Andreessen and Eric Bina at the National Center for Supercomputing Applications at the University of Illinois, did more than any other program to popularize the Web. The first public version for the X Window System appeared in January 1993, version 1.0 followed in April, and ports to Microsoft Windows and the Macintosh arrived later that year. Mosaic was easy to install, forgiving to use, and, critically, displayed images inline with text rather than launching an external viewer. For the first time a Web page could look designed.

Availability on all three major desktop platforms mattered as much as the interface. A researcher's Unix workstation, an office manager's Windows PC, and a designer's Macintosh could all reach the same page and see substantially the same thing. Interest in the Web grew explosively through late 1993 and 1994.

Explosive Growth

Surveys of the period counted fewer than one hundred websites at the start of 1993 and roughly ten thousand by the end of 1994; the total passed one hundred thousand during 1995. Content broadened from physics preprints and software documentation to catalogs, magazines, government records, and personal pages. No previous Internet service had grown at anything like that rate, and none had so thoroughly changed public expectations of what a computer network was for.

Origins at Xerox PARC

Robert Metcalfe and David Boggs developed Ethernet at Xerox's Palo Alto Research Center in 1973, adapting the randomized retransmission ideas of the ALOHAnet packet radio system in Hawaii to a shared coaxial cable. The experimental network ran at 2.94 megabits per second and connected Alto workstations and the first laser printers. Its central mechanism, carrier sense multiple access with collision detection, let stations transmit whenever the cable appeared idle, detect the collisions that inevitably resulted, and back off by a randomized interval that doubled with each successive collision.

DIX Ethernet and IEEE 802.3

Digital Equipment Corporation, Intel, and Xerox formed the DIX consortium in 1979 and published Ethernet Version 1.0 in 1980, specifying 10 megabits per second over thick coaxial cable. Version 2.0 followed in 1982 and remains the frame format in everyday use. The IEEE approved its own CSMA/CD standard on June 24, 1983, and published it in 1985 as IEEE Std 802.3-1985.

The two frame formats differ in one field. DIX Ethernet uses the two octets after the source address as an EtherType identifying the upper-layer protocol; IEEE 802.3 originally used the same octets as a payload length, with protocol identification moved into an 802.2 logical link control header. Because valid lengths and assigned EtherType values occupy disjoint numeric ranges, receivers can distinguish the formats unambiguously, and equipment routinely supported both. IPv4 has always been carried in Ethernet II frames.

Coaxial Ethernet: 10BASE5 and 10BASE2

The original 10BASE5 used a rigid half-inch coaxial cable, nicknamed thicknet, supporting segments up to 500 meters. Stations attached through vampire taps that pierced the jacket to contact the center conductor, connected to an external transceiver, and reached the interface card over a fifteen-pin attachment unit interface cable. Installation required care and specialized tools, which confined 10BASE5 to organizations with dedicated network staff.

10BASE2, or thinnet, used flexible RG-58 coaxial cable with BNC T-connectors at each station and 50-ohm terminators at both ends, supporting segments up to 185 meters. It cost far less to install, and it made Ethernet practical for small offices. Both variants shared a structural weakness: the cable was a single electrical bus, so one loose connector, missing terminator, or damaged segment took down every station on it. Fault isolation on a coaxial Ethernet was a matter of walking the cable.

10BASE-T and Structured Cabling

The IEEE ratified 10BASE-T as 802.3i in 1990, and it changed everything. Each station received a dedicated run of unshielded twisted-pair cable, up to 100 meters, terminated in an eight-position modular connector and radiating from a hub in a wiring closet. Separate transmit and receive pairs, isolation transformers at each end, and a periodic link integrity pulse gave every port an unambiguous up or down indication. A failed cable or a failed card now affected exactly one user, and a link light told an administrator which one.

The physical star topology aligned with the structured cabling practice being standardized in parallel as TIA/EIA-568, first published in 1991, which defined Category 3 and later Category 5 twisted-pair systems. Buildings could be wired once, generically, and support successive generations of networking equipment. Hubs remained shared-media repeaters, forwarding every bit to every port and preserving a single collision domain, but centralizing them in a closet made monitoring, port disabling, and troubleshooting tractable, and it set the stage for switching.

Interface Card Economics

Ethernet adapters for personal computers fell from several hundred dollars in the late 1980s to well under one hundred dollars for a 10BASE-T ISA card by the mid-1990s, driven by integration onto a single controller chip and by fierce competition among vendors. Manufacturers began building Ethernet directly onto server and business desktop motherboards. The price collapse is what turned networking from a capital project into a default assumption.

Ethernet against Token Ring, FDDI, and ATM

IBM's Token Ring, standardized as IEEE 802.5, ran at 4 megabits per second and later 16, and passed a token around a logical ring so that access was deterministic and throughput held up under heavy load, where CSMA/CD degraded. It was also markedly more expensive per port and tied to a single dominant supplier. Ethernet's cost advantage, aided by switching and then by Fast Ethernet, eroded Token Ring's position steadily; by the mid-1990s the outcome was clear.

The Fiber Distributed Data Interface, standardized by ANSI, ran at 100 megabits per second over dual counter-rotating fiber rings that could wrap around a break, and it dominated campus backbones in the early 1990s. Asynchronous Transfer Mode, promoted from the telecommunications world by the ATM Forum from 1991, promised a single cell-switched fabric spanning local and wide areas with quality-of-service guarantees. Fast Ethernet, ratified as IEEE 802.3u in 1995, undercut both in the local area: 100BASE-TX borrowed FDDI's 4B/5B encoding and ran it over ordinary Category 5 copper, and its autonegotiation scheme let 10 and 100 megabit equipment share an infrastructure during migration. Familiar, cheap, and incrementally deployable beat technically elegant.

From Gateways to Purpose-Built Routers

The first Internet routers were called gateways, and they were minicomputers running gateway software rather than dedicated hardware. BBN built early production gateways on DEC LSI-11 hardware and later on the multiprocessor Butterfly platform; universities pressed PDP-11 and Sun workstations into the same role. These machines are frequently confused with ARPANET's Interface Message Processors, but the two are distinct: an IMP was a packet switch internal to ARPANET, whereas a gateway joined ARPANET to a different network entirely.

Software forwarding on a general-purpose processor consumed a full interrupt, table lookup, and copy per packet. As link speeds climbed from 56 kilobits to T1 and then T3, the per-packet budget shrank from milliseconds to microseconds, and purpose-built forwarding hardware became unavoidable.

Cisco and the Commercial Router Market

Cisco Systems was founded in December 1984 by Leonard Bosack and Sandy Lerner, who managed computer facilities at Stanford University's computer science department and Graduate School of Business respectively. The company commercialized multiprotocol routing derived from work at Stanford, where William Yeager had written router software to join the campus's departmental networks. Cisco's first product, the Advanced Gateway Server, shipped in 1986 and established the enterprise router's defining characteristic: one box that spoke IP, DECnet, AppleTalk, IPX, and several other protocol families at once, because real campuses ran all of them.

Competitors followed quickly. Proteon, Wellfleet Communications, and 3Com built substantial router businesses, and Wellfleet merged with the hub maker SynOptics in 1994 to form Bay Networks. Cisco went public in February 1990 and used its valuation to acquire adjacent technologies, including the Ethernet switching pioneer Kalpana in 1994. Router software matured alongside the hardware, adding access control, traffic policy, and management interfaces that made the devices policy enforcement points rather than mere forwarders.

Dynamic Routing Protocols

Manually maintained routing tables did not scale past a handful of networks. The Routing Information Protocol, distributed in Berkeley Unix and later documented in RFC 1058, used distance-vector exchanges that were simple to implement but converged slowly and capped path length at fifteen hops. Open Shortest Path First, first specified in 1989, replaced this with a link-state design in which every router floods a description of its own links, builds an identical topology database, and computes shortest paths independently. OSPF converged faster, supported variable-length subnet masks, and partitioned large networks into areas.

Between administrative domains the requirements differ: an operator needs to choose routes by business policy, not just by path length. The Border Gateway Protocol addressed this. First specified in 1989 and reaching its current form as BGP-4 in 1994, BGP carries the full sequence of autonomous systems along each path, which both prevents loops and gives operators the raw material for policy decisions. BGP-4's support for classless prefixes made it the vehicle for CIDR deployment. Without it, the transition from a single government backbone to a mesh of competing commercial providers could not have happened.

Bridges, Spanning Tree, and Switching

Bridges joined Ethernet segments by learning which addresses lay on which port and forwarding selectively, which reduced traffic on each segment. Redundant bridges, however, create loops, and a bridged loop floods until the network collapses. Radia Perlman's spanning tree algorithm, devised at Digital Equipment Corporation in 1985 and standardized as IEEE 802.1D in 1990, solved this by having bridges elect a root and disable redundant links, reactivating them automatically when the active topology fails.

Switching applied bridging at every port simultaneously and in hardware. Where a hub places all stations in one collision domain and divides 10 megabits per second among them, a switch gives each port its own collision domain and forwards between ports in parallel, so aggregate capacity scales with port count. Because a switch is transparent at the frame level, existing adapters, cables, and software worked unchanged. Full-duplex operation followed: with only two stations on a link, collision detection becomes unnecessary, both directions can transmit at once, and the distance limits imposed by collision timing disappear. The IEEE standardized full duplex and its PAUSE flow control as 802.3x in 1997.

Kalpana and the First Ethernet Switch

Kalpana, founded in 1987 in Sunnyvale, California, introduced the seven-port EtherSwitch in 1990, the first commercial multiport Ethernet switch. Its distinguishing technique was cut-through forwarding: rather than buffering an entire frame before making a decision, the switch read the destination address in the first few bytes and began transmitting on the outbound port immediately, cutting latency to a fraction of a store-and-forward bridge's. The trade-off was that damaged frames were forwarded rather than discarded, which later switches addressed with adaptive modes that fell back to store-and-forward when error rates rose. Kalpana also pioneered full-duplex Ethernet and link aggregation before being acquired by Cisco in 1994.

Equipment Economics

Early enterprise routers sold for tens of thousands of dollars and made sense only for large institutions and carriers. By the mid-1990s, semiconductor integration and manufacturing volume had produced branch routers and workgroup switches at a small fraction of that price. The same curve that put an Ethernet controller on one chip put a switch fabric on a handful of them, and networking equipment followed the personal computer from the data center into the ordinary office.

The First Commercial Providers

UUNET, founded by Rick Adams in 1987 as a nonprofit offering Usenet and UUCP services, launched its commercial AlterNet Internet service in 1990. PSINet, incorporated in late 1989 from the assets of the regional network NYSERNet, began selling access in January 1990. In the same period, The World, operated by Software Tool and Die in Brookline, Massachusetts, became the first service to sell dial-up Internet access to the general public, taking its first customer in 1989. These providers connected the traffic that NSFNET's policy excluded, letting businesses exchange mail and files without an academic affiliation.

The Commercial Internet Exchange

Established in 1991 by PSINet, UUNET, and CERFnet, the Commercial Internet Exchange gave commercial providers a neutral place to interconnect. Before CIX, traffic between two commercial customers frequently had to traverse the government-funded backbone, which was both a bottleneck and a policy violation. CIX established settlement-free peering between providers as normal practice, and the Network Access Points created under NSF's 1993 restructuring generalized the model. The distributed interconnection architecture of today's Internet, in which no single operator sits at the center, dates from these arrangements.

Consumer Dial-Up and the Online Services

Institutions bought leased lines; households bought modems. Early consumer accounts were shell accounts on a provider's Unix host, reached by terminal emulation and offering mail, news, and text-mode file transfer. The arrival of graphical browsers changed the requirement: users now needed a genuine IP connection, which the Serial Line Internet Protocol and then the more capable Point-to-Point Protocol supplied over an ordinary phone call. Providers built banks of modems and terminal servers, and the busy signal became a defining consumer complaint of the era.

America Online, CompuServe, and Prodigy had built proprietary networks with their own content, forums, and mail systems, walled off from the Internet. As Internet interest grew, each added an Internet mail gateway, then Usenet, then Web access, and eventually repositioned itself as an Internet provider with extra content. The direction of travel was unmistakable: the open network absorbed the closed ones rather than the reverse.

Modem Standards and Access Speeds

Dial-up speeds advanced through a well-defined sequence of ITU-T recommendations. V.32 reached 9,600 bits per second in 1984, V.32bis raised that to 14,400 in 1991, and V.34 reached 28,800 in 1994, later extended to 33,600. The V.42bis data compression and V.42 error correction standards added effective throughput on compressible content. Each step demanded more sophisticated signal processing, and V.34 in particular approached the practical information-theoretic limit of an analog voice channel by combining trellis-coded modulation, adaptive equalization, echo cancellation, and continuous line probing that selected symbol rates and constellations to match the measured channel.

Business Connectivity

Organizations bought dedicated circuits. A T1 at 1.544 megabits per second became the standard mid-size business connection, with fractional T1 available for smaller needs and T3 at 44.736 megabits per second for large institutions and providers. Frame relay offered a cheaper alternative for bursty traffic by statistically multiplexing many virtual circuits over shared facilities and pricing capacity as a committed information rate rather than a fixed pipe. Integrated Services Digital Network basic rate service provided 128 kilobits per second over ordinary copper pairs and served small offices and determined home users, though North American deployment lagged the European build-out considerably.

Consolidation

The provider market consolidated almost as fast as it grew. Regional operators were bought by companies assembling national footprints, and national providers pursued international reach. Long-haul transmission, interconnection, and support carry heavy fixed costs and reward scale, so consolidation was structural rather than incidental. New entrants nonetheless kept appearing in specialized niches, a pattern that has persisted.

From Mosaic to Netscape

Mosaic's success attracted commercial attention that its university home could not accommodate. Marc Andreessen left NCSA and, with Silicon Graphics founder Jim Clark, incorporated Mosaic Communications Corporation in April 1994, recruiting several of the original Mosaic developers. The University of Illinois held the Mosaic trademark and objected, and the company was renamed Netscape Communications in November 1994. Its stated goal was a browser written from scratch that would outperform Mosaic on every axis.

Netscape Navigator

Navigator 1.0 shipped in December 1994 and reset expectations. It rendered pages progressively, showing text and laying out the document while images continued to download, instead of presenting a blank window until every element arrived. Over dial-up links that difference was transformative. Navigator 1.1, released in February 1995, added the Secure Sockets Layer; SSL version 1.0 had been designed in 1994 but was never released publicly because of flaws found in review, so SSL 2.0 in Navigator 1.1 was the first version users could actually run. Netscape also extended HTML aggressively with tags for tables, backgrounds, and font control, and introduced JavaScript, announced jointly with Sun Microsystems in December 1995 and shipped in Navigator 2.0.

Netscape's distribution model prefigured a great deal of later software business practice. Individuals and educational users downloaded the browser at no charge, while companies were expected to buy licenses, and the resulting installed base created demand for Netscape's server products. The company's initial public offering on August 9, 1995, priced five million shares at twenty-eight dollars; the stock traded as high as $74.75 during the first day and closed at $58.25, valuing a company barely sixteen months old and not yet profitable at more than two billion dollars.

Microsoft's Pivot

Microsoft initially directed its network strategy toward The Microsoft Network, a proprietary online service in the mold of CompuServe. Bill Gates's internal memorandum of May 26, 1995, "The Internet Tidal Wave," reversed that course and made Internet support a priority across every product line. Lacking a browser, Microsoft licensed Spyglass Mosaic, a commercially licensed derivative of NCSA's code, as its starting point.

Internet Explorer 1.0 shipped in August 1995 in the Microsoft Plus! pack for Windows 95, an unremarkable product. Version 2.0 followed on November 22, 1995, adding SSL, cookies, and Internet newsgroup support. Microsoft's scripting answer to JavaScript, JScript, did not arrive until Internet Explorer 3.0 in August 1996, which also introduced support for cascading style sheets and ActiveX controls. What made Microsoft formidable was not the early products but the combination of free distribution, operating system bundling, and effectively unlimited development resources.

Competing Extensions and Fragmentation

The rivalry extended well beyond browsers, into server software, scripting languages, and the standards process itself. Netscape promoted JavaScript, server-side applications on Unix, and its own HTML extensions; Microsoft promoted ActiveX, Windows server products, and extensions of its own. Competition drove genuinely rapid improvement, and it accelerated cascading style sheets, scriptable documents, and the programming model that made Web applications possible. It also produced years of incompatible markup and the era of pages that announced which browser they required, a cost that standards bodies and developers spent the following decade unwinding.

From EDI to the Open Internet

Electronic Data Interchange had automated business documents since the 1970s under standards such as ANSI X12 and the international UN/EDIFACT, but it ran over private value-added networks with per-character charges and demanded integration work that only large firms could fund. The Internet offered vastly cheaper transport and universal reach. What it did not offer was the implicit security of a closed private network, so building trustworthy transactions on an open network became the prerequisite for everything that followed.

Cryptography and Export Control

Public key cryptography, published by Whitfield Diffie and Martin Hellman in 1976 and realized as a practical algorithm by Rivest, Shamir, and Adleman in 1977, allowed two parties who had never met to establish a shared secret and to verify identities through certificates. RSA Data Security commercialized the algorithm, while symmetric ciphers such as DES and later triple-DES handled bulk encryption efficiently.

United States export regulations classified strong cryptography as a munition and limited exported software to short keys, which is why browsers of the era shipped in a domestic edition with 128-bit symmetric keys and an international edition restricted to 40 bits. The restriction was demonstrably inadequate: in August 1995 a French researcher, Damien Doligez, recovered the key from a 40-bit SSL session by brute force using idle workstations. Weeks later, Ian Goldberg and David Wagner showed that Netscape seeded its random number generator predictably, allowing session keys to be reconstructed regardless of length. Both findings were fixed, and both illustrated that transaction security depends on implementation quality as much as on algorithms.

The Secure Sockets Layer

Netscape's Secure Sockets Layer, designed in 1994 and first shipped as SSL 2.0 in February 1995, gave the Web a practical security layer. SSL sits between the application protocol and TCP, so HTTP required no modification: an ordinary request runs over an encrypted, integrity-protected channel to a server whose identity has been attested by a certificate authority. Early versions had genuine weaknesses, corrected in SSL 3.0 in 1996 and in the IETF's successor, Transport Layer Security, in 1999. The architecture, however, has proven durable, and the padlock indicator it introduced remains the public's primary cue for a secure connection.

Payment and Early Merchants

Consumer payment meant credit cards, which raised well-founded fears about transmitting account numbers over a public network. Card associations and technology vendors pursued dedicated protocols, and Visa and Mastercard announced a joint Secure Electronic Transaction specification in 1996; SET required certificates on every cardholder's machine, proved too cumbersome to deploy, and was abandoned, leaving SSL-protected card entry as the mechanism that actually prevailed. Alternative approaches from DigiCash and First Virtual likewise failed to reach critical mass.

Early merchants nonetheless proved the concept. Pizza Hut's PizzaNet, launched in August 1994 in Santa Cruz, California, accepted orders through a Web form and relayed them to the local restaurant. Amazon.com, incorporated by Jeff Bezos in July 1994 and opened to the public in July 1995, ran initially out of the garage of a rented house in Bellevue, Washington. Its early advantage lay in offering a searchable catalog far larger than any physical store could stock, together with customer reviews that gave shoppers a reason to return, and it demonstrated that Internet retail could take real share from established chains.

Business-to-Business Commerce

Less visible than consumer shopping, business-to-business trade represented the larger opportunity. Firms began carrying supply chain messages over Internet protocols rather than leased value-added networks, cutting transport costs by orders of magnitude and bringing smaller suppliers within reach of automated ordering. These early integrations laid the groundwork for the enterprise resource planning and supply chain automation projects that reshaped corporate operations in the following decade.

Venture Capital and the Netscape Signal

Silicon Valley venture firms, experienced with semiconductors and personal computers, recognized a familiar pattern of platform creation and moved early into Internet companies. Netscape's August 1995 offering converted that private conviction into a public phenomenon. A company sixteen months old with modest revenue and no profits closed its first trading day worth more than two billion dollars, establishing that Internet firms could be valued on growth narratives rather than earnings. Entrepreneurs and bankers drew the obvious conclusion, and the pipeline of Internet offerings filled rapidly.

Capital, Talent, and Coverage

Public markets receptive to unprofitable technology companies changed corporate behavior. Firms went public earlier, funding expansion with public capital and creating liquid markets for employee stock options, which in turn let startups outbid established employers for scarce engineering talent with equity rather than salary. Business and general media amplified the cycle, with coverage of instant fortunes drawing retail investors into technology stocks and expanding the capital available to the next cohort of ventures.

Infrastructure Investment

Not all of the money went to websites. Carriers committed billions to long-haul fiber, equipment vendors expanded manufacturing for routers, switches, and servers, and developers built the first generation of purpose-designed data centers. Much of this capacity was overbuilt relative to near-term demand, and the resulting glut destroyed enormous shareholder value after 2000. The physical plant, however, remained, and cheap abundant fiber underwrote the broadband and streaming era that followed.

Policy Environment

Public policy was broadly permissive. NSF's decision to privatize the backbone removed the government as gatekeeper of who could connect and for what purpose. Administration of names and numbers moved toward civilian institutions, with the InterNIC established in 1993 and, in September 1995, NSF authorizing Network Solutions to charge an annual fee for domain registration, which made the domain name a commercial asset. The Telecommunications Act of 1996, though aimed principally at telephone and cable markets, set competitive terms that shaped later broadband deployment. Regulators largely chose not to intervene in the Internet itself, and by 1995 every element of the boom was in place: a genuinely transformative technology, eager private capital, receptive public markets, and a policy stance that encouraged experiment.

Network Interface Controllers

Early Ethernet adapters were multi-chip assemblies: a media access controller, a separate encoder and decoder handling Manchester coding, bus interface logic, buffer memory, and discrete magnetics. Successive generations integrated these functions, and by the mid-1990s a single controller chip implemented the MAC, the physical layer interface, and a bus master DMA engine, leaving only the isolation transformer and connector outside. Integration cut board area, power, and cost simultaneously while removing the interconnections most likely to fail, which is what made an Ethernet port a standard motherboard feature rather than an option.

ASICs and Packet Processing

Forwarding decisions on a general-purpose processor cost hundreds of instructions per packet. At 10 megabits per second and 64-byte frames, a link can deliver more than fourteen thousand frames per second per port; a switch with a dozen ports leaves a processor no headroom at all. Application-specific integrated circuits designed around the forwarding path solved this by executing address lookup, filtering, and queuing in dedicated logic, in parallel, at line rate. The ability to specify and fabricate such devices became a defining competency for networking companies and a durable barrier to entry.

Content-Addressable Memory

Address lookup was the critical inner loop. Conventional memory returns data for a supplied address, so finding whether a MAC address is known requires a search or a hash computation, both of which cost variable time. Content-addressable memory inverts the operation: it compares a supplied value against every stored entry simultaneously and returns the matching location in a fixed number of cycles. Deterministic single-cycle lookup made wire-speed forwarding predictable, and later ternary CAMs, which allow individual bits to be masked as wildcards, extended the technique to longest-prefix matching for IP routing. The cost is silicon area and power, and managing that trade-off has occupied switch designers ever since.

Optical Transmission

Long-haul carriers had used optical fiber since the early 1980s, but terminal equipment costs restricted it to high-capacity trunks. The SONET and SDH standards, defined in the late 1980s, gave carriers a synchronous optical hierarchy with disciplined multiplexing and rapid protection switching, and it became the transport under most Internet backbones of the period. In the enterprise, cheaper short-reach transceivers using light-emitting diodes and multimode fiber brought optics to campus backbones, where FDDI at 100 megabits per second offered both the bandwidth and the electrical isolation that copper between buildings could not. Erbium-doped fiber amplifiers, demonstrated in the late 1980s and deployed commercially during the first half of the 1990s, removed the need to convert signals back to electrical form for regeneration every few tens of kilometers and set up the capacity expansion of the following decade.

Modem Signal Processing

Consumer modems were among the most sophisticated signal processing devices of the era. Extracting 28,800 bits per second from a channel roughly 3 kHz wide requires packing nine bits or more into each transmitted symbol, which leaves almost no margin for distortion. V.34 modems ran dedicated DSP chips executing adaptive equalization to invert the channel's frequency response, echo cancellation to separate the local transmitter from the far-end signal, trellis coding to recover margin lost to noise, and a line probing sequence at connection time that measured the channel and chose the symbol rate and constellation to match it. The result operated close to the Shannon capacity of a voice-grade line, which is why further gains required abandoning the analog channel altogether.

Power, Thermal, and Manufacturing Realities

Complexity brought heat. High-end routers filled chassis with processor and interface boards whose combined dissipation required engineered airflow, redundant fans, and, in carrier installations, thermal planning at the rack and room level. Successive CMOS process shrinks offset part of the increase, but power density trended upward throughout the period and became a first-order design constraint rather than an afterthought. Meanwhile production followed the broader electronics industry offshore: design work generally stayed in the United States and Europe while assembly moved to contract manufacturers in Asia, lowering costs and introducing the supply chain management problems that networking vendors have contended with ever since.

Complementary Standards Bodies

The IETF produced the protocols above the wire, moving from proposal to deployed standard in a fraction of the time treaty organizations required, and enforcing the discipline that a specification must be implementable and implemented. The IEEE 802 committee produced the standards for the wire itself, and its 802.3 working group defined each successive Ethernet generation from the original 10 megabits per second through Fast Ethernet in 1995 and Gigabit Ethernet in 1998. The division of labor was clean, and the two processes rarely collided.

Interoperability Testing

Specifications contain ambiguities, and implementations contain bugs. The University of New Hampshire InterOperability Laboratory, founded in 1988, gave vendors a neutral venue to test products against competitors' equipment before shipping. The Interop trade show, which grew out of the TCP/IP Interoperability Conference of the mid-1980s, built a live multi-vendor show network that functioned as a public integration test. Both exposed defects and specification ambiguities early, improving products and the standards themselves.

Management and Cabling

The Simple Network Management Protocol, first specified in 1988, let a single management station poll devices from any vendor for status and statistics using a common data model, the Management Information Base. Implementations varied in quality and coverage, and the protocol's write capability was often disabled for security reasons, but SNMP made multi-vendor networks manageable, which removed one of the strongest arguments for buying everything from one supplier.

Physical standards mattered just as much. Convergence on eight-position modular connectors and TIA/EIA-568 structured cabling, with Category 5 as the workhorse specification, meant a building could be wired once to a published standard and accept equipment from anyone. Generic cabling turned a bespoke installation project into a commodity trade, and it let successive Ethernet generations reuse infrastructure that was already in the walls.