How the Internet Was Built

30.0 What this chapter gives you#

1. You will be able to tell the story of long-distance messaging from signal fires to fibre, in order, with correct names and years.
2. You will be able to explain packet switching, and say honestly who invented it, when, and why three people arrived at it separately.
3. You will be able to kill the most popular myth in computing: that the ARPANET was built to survive a nuclear war.
4. You will be able to say exactly what happened on 29 October 1969.
5. You will be able to explain what TCP/IP is, why it was split into two layers, and what happened on 1 January 1983.
6. You will be able to state the difference between the internet and the web clearly enough to correct a newspaper.
7. You will be able to explain the dot-com boom and bust with real dates, real numbers, and an honest lesson rather than a moral.
8. You will be able to describe the physical internet: submarine cables, satellites, exchange points, and where India sits in that picture.
9. You will be able to say who actually runs the internet, and why the answer is uncomfortable and interesting at the same time.
10. You will be able to read a modern network path, like the reader’s own traceroute through Mumbai and Pune, and see the whole history in it.

30.1 Before electricity: fire, flags and towers#

PLAIN30.1.1 in simple words#

1. For most of human history a message travelled at the speed of a horse.
2. There was one exception, found very early: light travels much faster than a person, and a fire on a hill can be seen from the next hill.
3. A chain of fires can carry one fact across a country in minutes.
4. But a fire says only one thing: the agreed event has happened. It cannot say what, or how many, or when.
5. So the next step was to give the signal parts, so it could spell anything.
6. In France in the 1790s, Claude Chappe built towers with two big movable wooden arms on top. Each position of the arms meant a code.
7. A watcher in the next tower copied the shape, and the shape moved down the line. This was the first working long-distance message network in the world, and it ran for over sixty years.

PLAIN30.1.2 a picture in your head#

1. Imagine a line of people on hilltops, each just able to see the next one through a telescope.
2. Each person holds two long sticks. Each stick can be held in several positions, so two arms give many different shapes.
3. Each shape is a number, and both ends of the line hold the same code book. Number 4,102 might mean “the enemy has retreated”.
4. Nobody in the middle knows what the message means. They only copy shapes. That is an idea we still use: relaying without understanding.

Where this comparison breaks: there is no second path. A single sleepy operator stops the whole country, and fog, rain and darkness stop everything. A modern network routes around a failure in milliseconds. This one could not route around anything at all.

PLAIN30.1.3 a worked example#

1. Take the Paris to Lille line, the first real Chappe line, opened in 1794.
2. It ran roughly 230 km with about 15 stations, so about 15 km between towers. Sources differ a little on the distance, so treat it as approximate.
3. One symbol took about nine minutes to reach Lille. A full message took about 32 minutes in good weather.

4. In 1794 the line told Paris that Condé-sur-l’Escaut had been retaken from the Austrians, less than an hour after it happened.

PLAIN30.1.4 what is really happening inside#

1. Strip away the wood and three ideas are already present.
2. Encoding. The message becomes symbols the channel can carry. The channel carries shapes, not sentences.
3. Relaying. No station talks to the far end. Distance is built out of many short hops between neighbours.
4. A shared code book. Both ends must agree in advance what the symbols mean, or the shapes are noise.
5. All three survive today. Encoding is now modulation, relaying is now routing, and the code book is now a protocol specification.
6. The fatal weaknesses were that it was a chain rather than a mesh, and that it carried one message at a time. Packet switching fixes exactly those two things, one hundred and seventy years later.

TECHNICAL30.1.5 the engineer’s version#

1. Chappe’s first public demonstration was on 2 March 1791 between Brûlon and Parcé, over 16 km, using synchronized pendulum clocks rather than arms.
2. The French Legislative Assembly backed a Paris to Lille relay line in 1792. The line entered service in 1794.
3. Throughput was about 2 to 3 symbols per minute, roughly 0.4 to 0.6 words per minute of plain text.
4. The Chappe code book used 92 basic symbols taken two at a time, giving 8,464 coded words and phrases. That is compression: one symbol pair stands for a whole phrase.
5. At its peak the French network had 556 stations covering about 4,800 km. It stayed in service until 1855, and one of its last messages reported the fall of Sebastopol. Claude Chappe died on 23 January 1805.

6. The modern term for the tower operator’s job is regenerative repeater: a device that decides what the symbol was and transmits a clean copy, rather than amplifying the noise along with the signal.

WORDS30.1.6 remember these#

1. Semaphore — signalling by the position of arms — visual signalling using a defined symbol alphabet on a mechanical indicator.
2. Code book — a shared list of what each signal means — a codec or protocol specification agreed by both endpoints.
3. Regenerative repeater — a station that redraws the signal cleanly — a repeater that decides the symbol and retransmits, avoiding noise build-up.

30.2 The electric telegraph#

PLAIN30.2.1 in simple words#

1. Electricity moves along a wire almost instantly and does not care about fog, night or hills.
2. In Britain, William Fothergill Cooke and Charles Wheatstone patented a telegraph in May 1837, using needles that swung to point at letters.
3. In the United States, Samuel Morse patented a recording telegraph in 1837. His assistant Alfred Vail built much of the practical machinery.
4. Morse and Vail also produced a code of short and long pulses, which we call Morse code.
5. On 24 May 1844 Morse sent “WHAT HATH GOD WROUGHT” from the Capitol in Washington to Baltimore, over a line paid for by the United States Congress.
6. Wires then spread fast, because railways needed them and newspapers wanted them. The hard part left was the ocean.
7. The first Atlantic cable worked for about three weeks in 1858 and died. A working cable was completed on 27 July 1866, and from that day a message crossed the Atlantic in minutes instead of ten days.

PLAIN30.2.2 a picture in your head#

1. Think of a very long garden hose full of water, with a tap at one end.
2. Open and close the tap quickly and a pattern of pulses comes out the far end. Agree what the patterns mean and you have a telegraph.
3. Now make the hose two thousand miles long and lay it on the sea floor.
4. Water pushed in at one end no longer arrives as a sharp pulse. It arrives smeared out and weak, because the hose absorbs and delays it.
5. Push the pulses closer together and they blur into each other, and the far end cannot tell them apart. That blurring sets how fast you can go.
6. And if you turn the tap on harder to fight the weakness, you can burst the hose. That is roughly what killed the 1858 cable.

Where this comparison breaks: water pressure is not voltage. A long insulated conductor sitting in seawater behaves like a chain of capacitors and resistors, which is why William Thomson, later Lord Kelvin, had to work out the mathematics before long cables could be made to work at all.

PLAIN30.2.3 a worked example#

1. Morse code is not random. Common letters got short codes.

E   .
T   -
A   . -
S   . . .
O   - - -
SOS . . .   - - -   . . .

2. A skilled operator ran at roughly 20 to 25 words per minute, so that famous four-word phrase took about 10 to 12 seconds to send.
3. Now compare that with the ocean.

4. Queen Victoria’s message to President Buchanan on 16 August 1858 was 98 words and took about 16 hours to send. Even that beat a ship by a week.

PLAIN30.2.4 what is really happening inside#

1. A telegraph key is a switch. Pressing it connects a battery to the wire.
2. At the far end an electromagnet pulls a lever whenever current flows. Morse’s design pressed that lever against moving paper, so the message was recorded.
3. On a long undersea cable the current does not switch cleanly. The cable stores charge along its whole length, like a very long capacitor.
4. A sharp pulse arrives as a slow rounded hump, and closely spaced humps overlap. That overlap is called intersymbol interference, and it is still the limiting problem in every fast link today, including fibre.
5. The 1858 cable failed because the chief electrician, Wildman Whitehouse, applied very high voltage to force a stronger signal and the insulation broke down.
6. The 1866 cable succeeded through better insulation and manufacture and, crucially, instruments sensitive enough to detect a faint signal. On a long link, improving the receiver beats shouting louder.

TECHNICAL30.2.5 the engineer’s version#

1. The Cooke and Wheatstone patent of May 1837 recommended five needles, coding 20 of the 26 letters directly.
2. A four-needle system was installed on the Euston to Camden Town section of the London and Birmingham Railway in 1837 and rejected in favour of pneumatic whistles. Their first commercial success was on the Great Western Railway in 1838, over the 21 km from Paddington to West Drayton.
3. Multiplexing arrived because copper and poles were the expensive part.

4. Émile Baudot’s 1874 printing telegraph used time-division multiplexing: a rotating distributor gave each operator a slot on the same wire. Typical working speed was 30 words per minute.
5. The Baudot code was a fixed-length 5-bit code, unlike Morse’s variable length code. Fixed length is what a machine wants; variable length suits a human ear.
6. Cable dates precisely: the 1858 cable was completed on 5 August 1858, first official message 16 August, failed in September after about three weeks. The 1865 attempt broke after 1,062 nautical miles had been paid out from the SS Great Eastern. The successful cable landed on 27 July 1866, from Foilhommerum in Ireland to Heart’s Content in Newfoundland.

WORDS30.2.6 remember these#

1. Morse code — dots and dashes for letters — a variable-length code optimized for letter frequency in English.
2. Multiplexing — sharing one wire between several messages — combining channels onto one medium by time, frequency or code.
3. Intersymbol interference — pulses smearing into each other — spreading of symbol energy into adjacent symbol periods, limiting symbol rate.

30.3 The telephone network#

PLAIN30.3.1 in simple words#

1. The telegraph sent letters. The next question was whether a wire could carry a voice.
2. Alexander Graham Bell was granted US patent 174,465 on 7 March 1876, and on 10 March 1876 made the first successful call, saying “Mr. Watson, come here, I want to see you”.
3. A voice is not dots and dashes. It is a continuously changing signal, so the wire must carry a smooth copy of the sound.
4. At first every telephone needed a direct wire to every other one. Ten phones need forty-five wires. A hundred need 4,950.
5. The fix was an exchange: one wire from each phone to a central room, where someone joins two wires together.
6. Almon Strowger, an undertaker, believed operators were sending his calls to a rival, and built a machine to replace them. His patent was granted in 1891 and the first automatic exchange opened in La Porte, Indiana on 3 November 1892, for about 75 subscribers.
7. From that grew the largest machine humans had built, working in a way that could never have become the internet.

PLAIN30.3.2 a picture in your head#

1. Think of a railway where the whole route is booked for one train.
2. Every section of track from your station to your destination is reserved before you leave, and nobody else may use any part of it.
3. That is circuit switching. It is wonderful for a phone call, because once booked the quality never wavers mid-sentence.
4. It is terrible for computers, which send a burst, think for four seconds, then send another burst.

Where this comparison breaks: from the 1960s, long-distance calls were carried digitally over shared links in time slots, so there was no continuous copper path. But the reservation model stayed: a slot was booked for the whole call whether you spoke or not. The reservation, not the copper, is the point.

PLAIN30.3.3 a worked example#

1. Suppose you type at 5 characters per second to a remote machine. At 8 bits per character that is 40 bits per second of real information.
2. The circuit reserved for you is a digital voice channel, 64,000 bits per second.
3. You are using about 0.06 percent of it. The rest is reserved and wasted.
4. Now put 100 such users side by side.

5. That table is the entire economic argument for the internet.
6. It is also why the telephone companies, who were excellent engineers, resisted the idea. Their business was selling reserved time.

PLAIN30.3.4 what is really happening inside#

1. Lift a handset and you close a loop, so direct current flows from the exchange through your phone and back.
2. A rotary dial breaks that current once per unit of the digit. Dial 5 and it breaks the loop 5 times.
3. A Strowger switch counts the breaks with an electromagnet, stepping the wiper up five levels. The next digit rotates it to one of ten contacts, so two digits select one of a hundred outlets.
4. Nothing in that path stores your voice or looks at it. The network is a wire-joining machine, not a message-handling machine, and that is the deep reason it could not become the internet.

TECHNICAL30.3.5 the engineer’s version#

1. Elisha Gray filed a caveat for a liquid-transmitter telephone on 14 February 1876, the day Bell’s application was filed. Gray did not contest priority, and the examiner approved Bell’s patent on 3 March 1876.
2. The Bell Telephone Company was formed in 1877. By 1886 over 150,000 Americans had telephones.
3. Strowger conceived his switch in 1888 and received US patent 447,918 in 1891. The prototype used a collar box and pins. The two-motion selector gives 10 levels by 10 contacts, so 100 outlets.
4. Digital voice standardized on 8,000 samples per second at 8 bits, giving the 64 kbit/s DS0 channel. That sampling rate follows from the Nyquist rule applied to a 300 to 3,400 Hz voice band.

5. Signalling System No. 7 moved call setup onto a separate packet network while voice used reserved circuits. Note the detail: SS7 is itself packet based.
6. The Bell System was broken up by consent decree effective 1 January 1984, separating AT&T from the regional operating companies.
7. Stated technically, the mismatch is this: circuit switching gives constant bit rate, constant delay and zero statistical multiplexing gain. Computer traffic is bursty, often with a peak-to-mean ratio above 100 to 1, and gains enormously from statistical multiplexing.

WORDS30.3.6 remember these#

1. Exchange — the building that joins calls — a switching centre terminating subscriber loops and interconnecting them.
2. Circuit switching — booking a path for the whole call — dedicating a fixed channel end to end for a session’s duration.
3. DS0 — one digital voice channel — 64 kbit/s, 8,000 samples per second at 8 bits, the base unit of the digital telephone hierarchy.

30.4 Packet switching, invented three times#

PLAIN30.4.1 in simple words#

1. Packet switching is the idea everything on the internet rests on.
2. Instead of booking a path and sending a stream, you chop the message into small pieces.
3. Each piece carries the full destination address, like an envelope with a complete postal address on it.
4. Each piece travels on its own. Machines along the way read the address and pass it towards the destination.
5. Different pieces may take different routes, arrive out of order, or be lost. The receiving end reorders them and asks for anything missing.
6. Nobody books anything. The link is shared moment by moment between whoever has a piece ready to send.
7. Three people worked this out separately in the early 1960s, in two countries, for two completely different reasons. When an idea is genuinely ready, it tends to be found more than once.

PLAIN30.4.2 a picture in your head#

1. You want to send a long book to a friend in another city.
2. Circuit switching is hiring a private van that drives non-stop from your door to theirs with the whole book inside.
3. Packet switching is tearing the book into single pages, putting each page in its own addressed envelope with a page number, and posting them all.
4. The envelopes mix with everyone else’s post. Each sorting office moves each one closer. Some go one way, some another, and they arrive out of order.
5. Your friend sorts them by page number. If page 74 never turns up they ask for page 74 only, not the whole book.

Where this comparison breaks: the post has no congestion feedback. The internet does. When routers fill up and drop packets, senders slow down. That feedback loop is why your download speed changes second by second, and nothing in the postal picture explains it.

PLAIN30.4.3 a worked example#

1. Take a 10,000-byte file on a network whose maximum packet size is 1,500 bytes, the common Ethernet limit.
2. It splits into 7 packets: six of 1,500 bytes and one of 1,000.
3. Each gets a header. A minimal IPv4 header is 20 bytes and a minimal TCP header another 20, so 40 bytes of addressing and control per packet.

File 10,000 bytes
  +-> pkt 1..6  [hdr 40][data 1500]
  +-> pkt 7     [hdr 40][data 1000]
On the wire: 10,000 + 7*40 = 10,280 bytes
Overhead:    280 / 10,280 = 2.7 percent

4. Overhead is 2.7 percent. That is the price of the whole idea, and it is small.
5. Now lose packet 4. Circuit switching has no answer, because it assumed the path was reliable.
6. Packet switching resends 1,540 bytes instead of 10,280. A 15 percent cost instead of a 100 percent cost.

PLAIN30.4.4 what is really happening inside#

1. The three inventors reached the same design from different starting points, and it helps to keep them apart.
2. Paul Baran, at the RAND Corporation, worked from around 1960 on survivable communications. He proposed a distributed mesh with many alternative paths, messages cut into fixed message blocks, and store-and-forward switching with what he called hot potato routing.
3. Donald Davies, at the National Physical Laboratory in Britain, started in 1965 from a different problem: interactive computing was painful because telephone lines were slow to set up and wasteful when idle.
4. Davies coined the word packet, choosing a short word that would survive translation into other languages.
5. Leonard Kleinrock, at MIT and then UCLA, did the mathematics of queues: what happens to delay when messages wait in line at a busy node.
6. The credit question got heated later. The honest position is that historians and primary sources credit Baran and Davies with independently inventing packet switching.
7. Kleinrock’s early published work was about message switching and queueing delay, and in 2023 he acknowledged that. Donald Davies wrote that he could find no evidence Kleinrock understood the principles of packet switching from that work, and Alex McKenzie of BBN said the 1964 book contains nothing analysing packetization.
8. That is a live disagreement between people who were there. State it as a disagreement, not as a settled insult in either direction.

TECHNICAL30.4.5 the engineer’s version#

1. Baran’s work at RAND was first briefed to the US Air Force in summer 1961 as briefing B-265, published as RAND paper P-2626 in 1962 and as the report series RM-3420 in 1964. The full series, “On Distributed Communications”, ran to eleven volumes.
2. His three core ideas were a decentralized network with multiple paths between any two points, division of user messages into message blocks, and delivery by store-and-forward switching.
3. Davies conceived his design in 1965, proposed a national commercial data network publicly in 1966, and his colleague Roger Scantlebury presented the work at the Symposium on Operating System Principles in October 1967. The NPL Data Communications Network piloted in early 1969 and entered service in 1970.

4. Now the nuclear myth, stated precisely, because it needs care.
5. It is true that Baran’s funding context was US Air Force interest in communications that could survive attack, and his reports address that.
6. It is false that the ARPANET was built for that reason. ARPA director Charles Herzfeld stated plainly that the ARPANET was not started to create a command and control system that would survive a nuclear attack.
7. The ARPANET’s real motivation was resource sharing: expensive research computers were scarce, incompatible and scattered.
8. So the correct sentence is that one of the three independent inventions of packet switching had a survivability motive, and the network that first deployed it did not.
9. The model that matters in practice is Little’s Law: L equals lambda times W, relating mean number in system, arrival rate and mean time in system. It is why a link driven near 100 percent utilization has unbounded queueing delay.

WORDS30.4.6 remember these#

1. Packet — a small piece of a message with an address on it — a self-contained protocol data unit with header and payload.
2. Store and forward — receive it fully, then send it on — a node buffers the whole unit before retransmitting on the next hop.
3. Little’s Law — long queues mean long waits — L equals lambda times W.

30.5 ARPANET#

PLAIN30.5.1 in simple words#

1. ARPA was a research funding agency inside the United States Department of Defense.
2. In October 1962 J.C.R. Licklider was put in charge of its computing office. He had already published “Man-Computer Symbiosis” in 1960.
3. On 23 April 1963 he sent a memo addressed to “Members and Affiliates of the Intergalactic Computer Network”. The joke name carried a serious idea: these separate machines ought to be able to talk to each other.
4. In 1966 Bob Taylor, then running that office, started the actual project. He was irritated by having three terminals on his desk for three remote computers.
5. He hired Larry Roberts to run it, and the work went out to tender in 1968.
6. A small company in Cambridge, Massachusetts called Bolt Beranek and Newman won the contract in January 1969.
7. They built a box called an Interface Message Processor, which sat between a university’s computer and the network. On 29 October 1969 the first message was sent, and the receiving machine crashed after two letters.

PLAIN30.5.2 a picture in your head#

1. Imagine four universities, each with one expensive and completely different computer, with different operating systems and tape formats.
2. So instead you put a small identical box next to each computer.
3. Every box is the same model from the same company running the same software, and the boxes talk to each other over leased telephone lines.
4. Each box talks to its own local computer in that computer’s own language.
5. The universities never had to agree with each other. Each only had to agree with the box in its own machine room.

Where this comparison breaks: the box was not merely a translator. It did the routing, buffering, error checking and retransmission. It was the first router, and it did the hard part so the host computers could stay ignorant of it.

PLAIN30.5.3 a worked example#

1. On 29 October 1969, at about 10:30 pm Pacific time, Charley Kline at UCLA tried to log in to the SRI machine over the new network.
2. To log in he had to type LOGIN, and the remote system echoed each character.

UCLA (SDS Sigma 7)          SRI (SDS 940)
    |-- 'L' ------------------>  |
    |  <---------------- echo 'L'|
    |-- 'O' ------------------>  |
    |  <---------------- echo 'O'|
    |-- 'G' ------------------>  |
    |                        [crash]

3. So the first message ever sent between two computers on the ARPANET was “LO”. The crash was in SRI’s host software, not in the network, and a full login worked about an hour later.
4. The four original nodes, installed through the last months of 1969, were these.

PLAIN30.5.4 what is really happening inside#

1. The Interface Message Processor, always called the IMP, was a ruggedized Honeywell DDP-516 minicomputer with 24 kilobytes of magnetic core memory.
2. Each IMP connected to its local host on one side and to two or more other IMPs on the other, over leased telephone lines.
3. Each IMP kept a table saying which neighbour was the best next hop for each destination, updating it as conditions changed. That is exactly what a router does today; the word did not exist yet.
4. Above the IMP layer, host computers spoke the Network Control Program, or NCP, which assumed the network underneath was reliable.
5. That assumption is the key difference from TCP/IP and is what eventually had to be thrown away. Note the split, though: identical switches doing the networking, wildly different hosts doing the applications. That is still the shape of the internet.

TECHNICAL30.5.5 the engineer’s version#

1. The honest version of the dates: Licklider arrived at the Information Processing Techniques Office in October 1962, and the famous Intergalactic Computer Network memo is dated 23 April 1963. Use 1962 for his arrival and 1963 for the document.
2. The request for quotation went to 140 potential bidders in 1968 and 12 bid. BBN won, and the contract was awarded in January 1969.
3. The host-to-IMP interface was specified in BBN Report 1822, which is why engineers still call it 1822. Lines between IMPs ran at 50 kbit/s.
4. The public demonstration was at the International Conference on Computer Communications in Washington in October 1972, organized largely by Robert Kahn, with about 40 machines connected live in a hotel.
5. That demonstration matters more than it sounds. Until then, telephone engineers widely considered packet switching doubtful. Afterwards it was considered obvious.

6. In 1973 the ARPANET gained links to Norway, via NORSAR, and to University College London, making it international.
7. The tools that echo this era are still on your machine. ping measures round-trip time and traceroute shows the hop sequence, because a hop-by-hop store-and-forward network needs hop-level diagnostics.

WORDS30.5.6 remember these#

1. ARPANET — the first packet switched network of this kind — the ARPA-funded research network, operational 1969 to 1990.
2. IMP — the box that did the networking — Interface Message Processor, a Honeywell DDP-516 acting as the first packet switch, specified in BBN Report 1822.
3. NCP — the early host-to-host rules — Network Control Program, the ARPANET host protocol before TCP/IP, which assumed a reliable subnet.

30.6 Email and the applications that came first#

PLAIN30.6.1 in simple words#

1. The ARPANET was funded to let people share expensive computers. What people actually did with it, almost immediately, was send each other messages.
2. Machines already had local mail: you could leave a note for another user of the same computer.
3. In 1971 Ray Tomlinson at BBN joined a local mail program to a file transfer program, so a note could be left on a different computer.
4. He needed a way to write “this user, on that machine”, and picked the @ sign because it was not used in user names and read naturally as “at”.
5. Email spread on its own, with nobody planning or funding it, and within a couple of years it was the bulk of the traffic.
6. That is the pattern to remember: the important application is usually not the one in the funding proposal.

PLAIN30.6.2 a picture in your head#

1. Think of a research building where every office has a pigeonhole for internal notes. That works, but only inside the building.
2. Now someone realizes the courier who carries parcels between buildings could also carry notes.
3. So they write a note, put it in a parcel, and mark it for the pigeonhole of Dr Rao, in the Chennai building.

Where this comparison breaks: unlike a parcel, an email is copied rather than moved, so the sender still has it. That single difference is why email became a permanent archive, and why mailing lists were possible at all.

PLAIN30.6.3 a worked example#

1. An email address has exactly two parts, split by the @ sign. In tomlinson@bbn-tenexa, the local part is tomlinson and the host part is bbn-tenexa.
2. Here is a modern message as it actually travels.

S: 220 mail.example.net ESMTP ready
C: EHLO client.example.org
S: 250 STARTTLS
C: MAIL FROM:<student@example.org>
S: 250 2.1.0 Ok
C: RCPT TO:<teacher@example.net>
S: 250 2.1.5 Ok
C: DATA
S: 354 End data with <CR><LF>.<CR><LF>
C: Subject: chapter 30
C: (blank line, then the body)
C: .
S: 250 2.0.0 Ok: queued as 4A2F1C
C: QUIT

3. Notice the envelope and the content are separate. MAIL FROM and RCPT TO are the envelope; the Subject line inside DATA is content.
4. Mail is delivered to the envelope address, not the To: header. That is why blind copies work at all.
5. That design is from RFC 821 in 1982, and the modern version is RFC 5321 from 2008. The shape has not changed.

PLAIN30.6.4 what is really happening inside#

1. Three applications defined the first decade.
2. Telnet gave you a terminal session on a remote machine. FTP, the File Transfer Protocol, moved files; its first specification is RFC 114 from 1971. Mail began as a bolt-on to file transfer and then got its own rules.
3. MsgGroup started in 1975. Another early list, SF-Lovers, was about science fiction, and administrators argued about whether a defence-funded network should carry it. That argument about acceptable use returns in the 1990s and matters enormously.
4. In 1979 and 1980, Tom Truscott and Jim Ellis at Duke University, with Steve Bellovin, built something different: Usenet.
5. Usenet was not on the ARPANET. It ran over ordinary dial-up calls between Unix machines using UUCP, Unix-to-Unix Copy. Each machine phoned its neighbours and swapped new articles, so over days an article reached every participating machine in the world.
6. That is store-and-forward at the application level, with no central server, no live connection and no permission needed from anybody. It is where the words FAQ, flame war and spam entered common use.

TECHNICAL30.6.5 the engineer’s version#

1. Tomlinson’s 1971 work combined SNDMSG, a local message program, with CPYNET, an experimental file copy program, on TENEX systems at BBN. The exact date is not recorded, only the year.

2. Usenet was publicly established in 1980 at Duke University and the University of North Carolina at Chapel Hill. The first group was net.general.
3. The Usenet transport protocol over TCP is NNTP, RFC 977 from 1986, updated by RFC 3977 in 2006.
4. Email’s structural weakness was baked in early. SMTP has no authentication of the sender, and everything added since, SPF (RFC 7208), DKIM (RFC 6376) and DMARC (RFC 7489), is a patch over that original openness.
5. You can still watch the protocol directly with openssl s_client -starttls smtp -connect mail.example.net:587.

WORDS30.6.6 remember these#

1. Email — a note left in a mailbox on another computer — a store-and-forward messaging system using SMTP for transport and RFC 5322 for format.
2. Envelope — the delivery instructions, separate from the letter — the SMTP MAIL FROM and RCPT TO commands, distinct from message headers.
3. Usenet — a worldwide notice board copied machine to machine — a distributed discussion system using flood-fill propagation over UUCP or NNTP.

30.7 TCP/IP and the flag day#

PLAIN30.7.1 in simple words#

1. By the early 1970s there was not one network but several, and they were all different.
2. The ARPANET ran over leased telephone lines, PRNET was a radio network in the San Francisco area, and SATNET went over satellite to Europe.
3. The question was how to join networks that disagree about everything.
4. Bob Kahn posed the problem and Vint Cerf worked on it with him from spring 1973. In May 1974 they published “A Protocol for Packet Network Intercommunication” in IEEE Transactions on Communications.
5. Their answer was not to make the networks agree. It was to put a common protocol on top of all of them, with boxes called gateways in between.
6. That is what the word internet means: internetworking, networks of networks. In 1978 the design was split into TCP and IP, and on 1 January 1983 the ARPANET switched to them in a single day.

PLAIN30.7.2 a picture in your head#

1. Imagine shipping goods across a continent where each country has different trucks, roads and load limits.
2. The good solution is the shipping container. Agree only on the box.
3. IP is the shipping container: a standard box with an address, able to travel over any underlying network.
4. TCP is the paperwork saying how many boxes there are, in what order, and what to do if one is lost. Gateways are the border crossings.

Where this comparison breaks: a real container is never opened at the border, but an IP packet may be cut into smaller pieces if the next network cannot carry one that big. That is fragmentation, and it is one of the messier parts of the design. Also, no shipping firm would tolerate quietly losing containers. IP does, on purpose, leaving recovery to TCP.

PLAIN30.7.3 a worked example#

1. Look at what happened on the reader’s own machine when it tried to reach github.com, which DNS resolved to 20.207.73.82.

 1  192.168.0.1          home router
 2  172.31.0.17          ISP private core
 3  137.97.29.249        ISP public edge
 7  ...del01...msn.net   Microsoft, Delhi
 8  ...bom01...msn.net   Microsoft, Mumbai
10  ...pnq21...msn.net   Microsoft, Pune
13+ * * *                no reply

2. Hops 1 to 6 are inside the reader’s ISP. Hops 7 onwards are inside Microsoft’s own backbone. Two completely separate organizations.
3. Nothing there required them to run the same equipment, software or internal addressing. They only had to agree on the IP packet format and on how to exchange routes.
4. Then look at the failure. The connection to port 443 timed out after 15 seconds with no reply of any kind: no reset, no error, silence.
5. IP did its job by promising nothing. TCP did its job by retrying and then giving up. Correct layering does not mean success. It means each layer fails in a well-defined way.

PLAIN30.7.4 what is really happening inside#

1. The original 1974 design was one protocol doing both jobs, addressing and reliability.
2. That was wrong for one important reason: not every application wants reliability. Digitized voice would rather lose a packet than wait for it.
3. So in 1978 the design was split. IP does addressing and forwarding and promises nothing. TCP sits on top and adds order, retransmission and flow control.
4. IP is deliberately best effort: it will try, it may fail, and it will not tell you.
5. This is the hourglass or thin waist model: many technologies below, many applications above, one narrow protocol in the middle. That narrow waist is why the internet survived every change underneath it, from 50 kbit/s leased lines to 400 Gbit/s optics, without rewriting the applications.
6. The companion idea is the end-to-end principle: put function in the endpoints, not the network, unless there is a strong reason. That is why routers are simple and your laptop is complicated. The telephone network did the opposite, and could not change.

TECHNICAL30.7.5 the engineer’s version#

1. Cerf and Kahn, “A Protocol for Packet Network Intercommunication”, IEEE Transactions on Communications, volume 22 issue 5, pages 637 to 648, May 1974.
2. RFC 675, December 1974, by Cerf, Dalal and Sunshine, was the first detailed specification and is where “internet” appears as a contraction of internetwork.
3. Version 4, in 1978, split the Transmission Control Program into IP, connectionless, and TCP, connection-oriented. Jonathan Postel drove the split. That version number is why we have IPv4 and never deployed an IPv1, 2 or 3.
4. The core specifications settled as RFC 791 (IP) and RFC 793 (TCP), both September 1981, plus RFC 768 (UDP, 1980) and RFC 792 (ICMP, 1981).
5. A flag day is a cutover with no compatibility period: at a set instant everyone stops the old thing and starts the new one, with no fallback.
6. The NCP to TCP/IP flag day was 1 January 1983. It was announced years ahead, tested in stages during 1982, and enforced by turning NCP off. Sites that were not ready lost service.
7. It worked because the network had a few hundred hosts, one funding agency could compel every site, and every administrator was reachable by name.

8. IPv6 proves the point. It was specified in RFC 1883 in 1995, updated as RFC 2460 in 1998 and RFC 8200 in 2017, and adoption is still partial in 2026, with Google’s measurement of IPv6-capable users in the range of roughly 40 to 50 percent in recent years.
9. The reader’s machine reported IPv6: (none), an ordinary result on many Indian home connections, which illustrates the point exactly.
10. Instead of flag days we use dual stack, running both protocols at once, with transition mechanisms such as NAT64 and 464XLAT. The cost is running two internets in parallel for thirty years.

WORDS30.7.6 remember these#

1. Internetworking — joining different networks into one — providing a uniform service across heterogeneous networks via a common protocol.
2. IP — the addressed envelope — the connectionless best-effort internet layer protocol, RFC 791 for version 4 and RFC 8200 for version 6.
3. TCP — the part that makes sure everything arrives in order — a connection-oriented reliable byte stream with congestion control.
4. Flag day — everyone switches at the same instant — a cutover with no backward compatibility period.
5. End-to-end principle — keep the clever parts at the edges — place function in end hosts unless it cannot correctly be implemented there.

30.8 Other networks and the standards war#

PLAIN30.8.1 in simple words#

1. It is easy to think the internet was the only game in town. It was not.
2. Through the 1980s there were many networks, most of them bigger and better funded.
3. Telephone companies offered X.25, a public data network standard first published in 1976.
4. France gave away terminals for a service called Minitel and had millions of users while most of the world had none.
5. Universities that could not join the ARPANET built BITNET in 1981, and hobbyists built FidoNet from 1984, linking home bulletin boards by nightly phone calls.
6. CompuServe and America Online sold access to their own private worlds, with their own content and their own rules.
7. And an enormous international standards effort, OSI, was supposed to produce the one true networking suite. Governments mandated it, textbooks taught it, and TCP/IP beat it completely.

PLAIN30.8.2 a picture in your head#

1. Imagine two teams designing a bicycle for the whole world.
2. Team B builds a rough bicycle, rides it, fixes the problems, and gives away both the plans and a working example for free.
3. Ten years later everyone rides Team B’s bicycle, and Team A’s document sits on a shelf being technically excellent.

Where this comparison breaks: OSI was not a failure of intelligence. Its vocabulary is still how we teach networking, and the seven-layer model is genuinely useful even though the seven-layer protocols lost. The failure was of process and timing, not of ability.

PLAIN30.8.3 a worked example#

1. Compare the rivals of the 1980s side by side.

2. Minitel is the most instructive. France Telecom gave terminals away free from 1984, and by early 1986 there were 1.4 million connected.
3. By the end of 1999 there were close to 9 million terminals used by around 25 million people, with revenue of 620 million francs in 1985 and about 832 million euros in 1998.
4. The reason matters: Minitel was closed, with one operator deciding what services could exist. The web let anybody publish without permission.
5. FidoNet is the opposite lesson. With no money at all it peaked at around 39,000 systems in the mid 1990s and roughly 4 million users by 1993, using hierarchical addresses like 1:250/250.10.

PLAIN30.8.4 what is really happening inside#

1. To understand why TCP/IP won, separate four advantages that people usually collapse into one.
2. Running code. TCP/IP existed and worked while OSI was still being written. In particular the 4.2BSD release of Berkeley Unix in 1983 shipped a usable TCP/IP stack with source, on the machines universities were already buying.
3. No licence fees and no owner. RFCs were free to read and free to implement. ISO standards had to be purchased, and still do.
4. Government backing of the useful kind. The US Department of Defense required TCP/IP on its networks, which created a large market immediately. OSI mandates promised a market later.
5. Simplicity. IP has one job. OSI’s network layer had modes, options and variants, so two implementations could both be correct and still fail to interoperate.
6. The honest counter-argument, still made by experts: OSI had better answers for network management, addressing scale and formal description, and some of the pain of the 1990s came from TCP/IP’s weaker versions of those. The better process beat the better committee, and we inherited both the benefits and the gaps.

TECHNICAL30.8.5 the engineer’s version#

1. X.25 was finalized by CCITT in 1976 as the Orange Book, revised in 1980, 1984, 1988, 1993 and 1996. It provided virtual circuits, typically at 2,400 bit/s to 64 kbit/s, rarely above.
2. BITNET was founded in 1981 by Ira Fuchs of CUNY and Greydon Freeman of Yale, first linking those two. It used IBM’s RSCS and NJE protocols, peaked around 1991 at almost 500 organizations and about 3,000 nodes, and lost support in 1996. LISTSERV came out of that world.
3. FidoNet was created by Tom Jennings in San Francisco in early 1984, first released in June 1984. Zone Mail Hour was a reserved period, commonly around 04:00, when nodes exchanged mail and human callers were locked out.
4. OSI: ISO began the work in 1977; the reference model was published as ISO 7498 in 1984, after a merged CCITT and ISO text was agreed in May 1983.

5. The US Government OSI Profile, GOSIP, became a federal procurement requirement in 1990 and was effectively abandoned in 1994 and 1995 when agencies were allowed to buy TCP/IP instead.
6. ASN.1, an OSI-era notation, is everywhere in security: the X.509 certificates your browser checks on every HTTPS connection are defined in ASN.1 and encoded in DER.
7. So the accurate summary is not that OSI failed. It is that the OSI protocol stack lost and several OSI ideas won quietly.

WORDS30.8.6 remember these#

1. X.25 — the phone companies’ data network — a CCITT virtual-circuit packet network standard from 1976.
2. Walled garden — a service where one company decides what exists inside — a closed platform with controlled content and gated third-party access.
3. OSI — the official international networking suite — the ISO and CCITT protocol family and its seven-layer reference model, ISO 7498.

30.9 From research network to public internet#

PLAIN30.9.1 in simple words#

1. Until the late 1980s you could not simply buy internet access. There was nothing to buy.
2. You got on the internet by being at a university or research lab with a connection paid for by a government agency.
3. In 1986 the US National Science Foundation built NSFNET to link supercomputer centres and then regional university networks.
4. It came with an acceptable use policy saying the network was for research and education. Commercial traffic was not allowed.
5. So a national network existed and businesses were forbidden to use it. They built their own connections and their own exchange points instead.
6. Between 1991 and 1993 the restrictions were relaxed and then removed, and commercial providers appeared quickly.
7. On 30 April 1995 the NSFNET backbone was switched off. That date is a good candidate for the day the internet became a public utility rather than a research project.

PLAIN30.9.2 a picture in your head#

1. Imagine a country with excellent motorways built only for university buses. Lorries are banned.
2. Eventually the private roads are better than the public one, the government closes its motorway, and everyone drives on the private roads.
3. The interchanges remain, and they are now the most important places on the map.

Where this comparison breaks: unlike a road interchange, two networks meeting at an exchange point may simply refuse to exchange traffic with each other. Peering is voluntary, and refusals happen, sometimes publicly and expensively.

PLAIN30.9.3 a worked example#

1. Follow the NSFNET backbone as demand grew.

2. Put a modern number beside it. DE-CIX Frankfurt, one exchange point, recorded a peak of 18.73 Tbit/s in December 2025.
3. That is roughly 416,000 times the entire 1991 national backbone, in one building, in one city.
4. That is why the acceptable use policy felt reasonable at the time. The capacity genuinely could not have absorbed commercial traffic.

PLAIN30.9.4 what is really happening inside#

1. The acceptable use policy was not a law. It was a condition attached to money, which is often stronger.
2. Commercial networks that wanted to exchange traffic without touching NSFNET formed the Commercial Internet eXchange, CIX, in 1991.
3. That is the first clear example of settlement-free peering, still how most large networks interconnect.
4. In 1991 the NSF removed access restrictions, and on 23 October 1992 US legislation authorized the NSF to support networks used substantially for purposes beyond research and education.
5. For the handover, NSF funded four Network Access Points, in New York, Washington DC, Chicago and San Jose. They were transitional and were replaced by hundreds of independent exchange points worldwide.
6. The structural change is the important part. Before 1995 there was a backbone. After 1995 there is no backbone, only a mesh of networks that agree to carry each other’s traffic.

TECHNICAL30.9.5 the engineer’s version#

1. NSFNET began in 1986 with six backbone sites on leased 56 kbit/s links. The T-1 backbone, 1.544 Mbit/s across 13 nodes, was in place by July 1988, operated by Merit Network with IBM and MCI. The T-3 backbone at 44.736 Mbit/s was deployed in November 1991.
2. The NSFNET Backbone Service transition completed and the fibre backbone was decommissioned on 30 April 1995.
3. What replaced it is described by business relationship, not technology.

4. Those relationships are enforced in BGP, the Border Gateway Protocol, currently RFC 4271. BGP decided the path the reader’s packets took towards Microsoft.
5. An autonomous system number, an ASN, identifies each independently routed network. Roughly 75,000 to 80,000 ASNs are visible in the global routing table in recent years, and the IPv4 table passed 1,000,000 routes in 2025.
6. An exchange point is physically a large Ethernet switching fabric in a neutral building. Members take a port, commonly 10, 100 or 400 Gbit/s, and set up BGP sessions with whichever other members agree.
7. DE-CIX Frankfurt had more than 4,300 connected networks by the end of 2025.
8. India’s National Internet Exchange, NIXI, was registered on 19 June 2003 specifically so Indian-to-Indian traffic would stop being hauled via Singapore, London or the United States. It has reported peak traffic of about 393 Gbit/s inbound.
9. DE-CIX opened its first Indian exchange in Mumbai in 2018 and now operates in several Indian cities including Delhi, Chennai, Bengaluru, Hyderabad and Kolkata.

WORDS30.9.6 remember these#

1. NSFNET — the American research backbone of the late 1980s — the NSF-funded TCP/IP backbone, 1986 to 30 April 1995.
2. Acceptable use policy — the rule saying what the network may be used for — a funding condition restricting traffic to research and education.
3. Peering — two networks swapping traffic directly — a BGP interconnection, usually settlement-free and carrying customer routes only.
4. Internet exchange point — a building where many networks meet — a shared layer 2 fabric enabling many-to-many BGP peering.
5. Autonomous system — one network under one routing policy — an independently administered routing domain identified by an ASN.

30.10 The World Wide Web#

PLAIN30.10.1 in simple words#

1. By 1989 the internet existed and worked, but there was no web.
2. To read a document on another machine you had to know it existed, know which machine held it, log in, find the directory and fetch the file.
3. Tim Berners-Lee was a software engineer at CERN, the particle physics laboratory near Geneva, where thousands of scientists came and went.
4. In March 1989 he wrote a proposal called “Information Management: A Proposal”, suggesting a linked information system. He wrote a second in May 1990, and a formal one with Robert Cailliau on 12 November 1990.
5. By the end of 1990 he had written a server and a browser, both running on a NeXT computer, and the first website was live at info.cern.ch.
6. On 6 August 1991 he posted about the project publicly, and the software went out to the whole internet on 23 August 1991.
7. On 30 April 1993 CERN put the web software into the public domain, free for anyone to use with no fee and no permission. That may be the most valuable giveaway in the history of technology.

PLAIN30.10.2 a picture in your head#

1. Think of a library with millions of books and no catalogue.
2. Now each book gains two things. First, a printed address on its spine that identifies it uniquely worldwide.
3. Second, wherever a book mentions another book, that mention becomes a button. Press it and the other book appears in your hands.

Where this comparison breaks: real links break. A book cannot vanish from your hands mid-sentence, but a web page can be deleted and the link becomes a dead end. Berners-Lee chose one-way links precisely because two-way links need permission from both ends, and requiring permission would have stopped the whole thing growing.

PLAIN30.10.3 a worked example#

1. Take one address apart. Every part does one job.

https://info.cern.ch/hypertext/WWW/TheProject.html
\___/   \__________/\_____________________________/
scheme      host                path

GET /hypertext/WWW/TheProject.html HTTP/1.1
Host: info.cern.ch

HTTP/1.1 200 OK
Content-Type: text/html

2. The scheme says which protocol to speak, the host says which machine, and the path says which document on it.
3. The response starts with a number. 200 means here it is, 404 means no such document, 301 means it moved. The body is HTML: text with tags marking headings, paragraphs and links.

4. None of the three was new on its own. The combination was: a global name, a trivial fetch, and a format that contains the next global name.

PLAIN30.10.4 what is really happening inside#

1. The web is an application that runs on the internet. It is not the internet.
2. The internet is the delivery system: addresses, routing, packets. Email, video calls, games, software updates, DNS and messaging apps all use the internet without using the web at all.
3. Three design choices made the web spread faster than its rivals.
4. Links are one way. You may link to any page without telling the owner, so the graph grows with no coordination.
5. Broken links are tolerated. A 404 is a normal answer, not a failure of the system. Designs that guaranteed link integrity could not scale.
6. It was free of ownership. CERN’s 1993 release meant no company could charge for the protocol or block a competing browser.
7. The honest version: the web was not the best hypertext system designed. It was much weaker than several research systems of the 1980s. It won because it was the one that required permission from nobody.

TECHNICAL30.10.5 the engineer’s version#

1. Protocol versions, with dates, because the differences matter in practice.

2. URLs were specified in RFC 1738 in 1994. The current generic URI syntax is RFC 3986 from 2005.
3. HTTP/3 uses QUIC (RFC 9000, 2021), which runs over UDP rather than TCP, because TCP’s ordering guarantee causes head-of-line blocking when a single packet is lost.
4. To see the raw protocol yourself, run curl -v against any site. It prints the request and response headers exactly as described above.

WORDS30.10.6 remember these#

1. Web — pages with links that you read in a browser — a distributed hypertext application layered on the internet.
2. URL — the full address of one document — a Uniform Resource Locator, syntax defined in RFC 3986.
3. HTTP — the request and answer language — the stateless application protocol for retrieving and submitting resources.
4. HTML — the way pages are written — a markup language describing structure and hyperlinks, now a living standard with no version number.

30.11 The browser wars#

PLAIN30.11.1 in simple words#

1. Early web browsers showed text and made you open images in a separate window.
2. In January 1993 Marc Andreessen and Eric Bina at the National Center for Supercomputing Applications released an alpha of a browser called Mosaic. Version 1.0 came on 21 April 1993.
3. Mosaic showed pictures inline with the text. That sounds small. It changed who wanted to use the web.
4. Andreessen left and, with Jim Clark, founded a company on 4 April 1994 that became Netscape. Its browser was released as Mosaic Netscape 0.9 on 13 October 1994 and took over the web almost immediately.
5. Netscape went public on 9 August 1995. Shares were offered at 14 dollars, doubled to 28 before trading opened, touched 75 on the first day and closed at 58.25, valuing a company with little revenue at about 2.9 billion dollars.
6. Microsoft responded by giving its browser away and putting it inside Windows. Netscape lost, was bought by AOL, and released its source code, which became Mozilla and eventually Firefox.
7. Then Google entered with Chrome in 2008 and won more completely than anyone before it.

PLAIN30.11.2 a picture in your head#

1. Imagine two companies selling door locks for a new kind of house.
2. Company B also sells houses, and starts fitting its own lock to every house it builds, free and already installed.
3. Company A cannot compete with free, or with pre-installed, and goes out of business. Later, a third company that owns the road everyone drives on offers a new lock and mentions it on every journey.

Where this comparison breaks: locks do not improve when more people use them, but browsers do. Popularity made web authors test against a browser, which made it work better, which made it more popular. That loop is why browser markets tip so hard toward one winner and are so difficult to reverse.

PLAIN30.11.3 a worked example#

1. Follow the market share and notice how fast each turn happened.

2. The July 2026 worldwide figures from Statcounter page-view measurement were Chrome 68.22, Safari 16.47, Edge 5.37, Firefox 3.34, Samsung Internet 2.06 and Opera 1.88 percent.
3. Be careful with such numbers. They measure page views from participating sites, not installed copies, and different measurement firms disagree by several points.
4. Edge and Opera are built on the same engine as Chrome, so engine share is far more concentrated than browser share.

PLAIN30.11.4 what is really happening inside#

1. A browser is several programs in one, and the important one is the rendering engine.
2. The rendering engine takes HTML, CSS and images and works out exactly what pixels go where. Beside it sit a JavaScript engine and a networking layer.
3. Building a rendering engine is enormously expensive. It must handle every badly written page on the web and match the leaders bug for bug. That is why the number of independent engines keeps falling.
4. Three engines survive in general use in 2026.
5. Blink, used by Chrome, Edge, Opera, Brave and Samsung Internet, forked from WebKit by Google in 2013.
6. WebKit, used by Safari, forked by Apple from KHTML, from the KDE project, in 2001.
7. Gecko, used by Firefox, descends directly from the Netscape code released in 1998. So the code Netscape published when it lost is still running today, which is a fair result for the loser of the first browser war.

TECHNICAL30.11.5 the engineer’s version#

1. NCSA Mosaic 0.5 alpha was announced on 23 January 1993, version 1.0 came on 21 April 1993, and NCSA ended support in January 1997.
2. Internet Explorer 1.0 shipped on 24 August 1995 in the Microsoft Plus pack for Windows 95. IE peaked above 90 percent usage share in 2001.
3. United States v. Microsoft was filed in 1998. The district court found a Sherman Act violation and in 2000 ordered a breakup; the appeals court reversed the breakup in 2001, and a settlement with conduct remedies rather than divestiture was approved in 2002.
4. Netscape announced on 22 January 1998 that its source would be free and developed openly as Mozilla. AOL announced the acquisition on 24 November 1998 at 4.2 billion dollars in stock, worth about 10 billion when it closed on 17 March 1999.
5. Firefox 1.0 was released 9 November 2004. Google Chrome was released 2 September 2008.

6. Microsoft retired its own Trident and EdgeHTML engines when Edge moved to Blink, released in January 2020. That took the world from four independent engines to three.
7. The concentration matters technically, not only commercially. When one engine has over two-thirds of usage, “it works in Chrome” starts to replace “it follows the specification” as the working definition of correct.
8. This is a live disagreement. One side argues consolidation makes the web more consistent and lets features ship faster. The other argues it hands one company effective control of a public standard. Serious people hold both positions.

WORDS30.11.6 remember these#

1. Browser — the program that shows web pages — a user agent implementing HTTP, HTML, CSS and scripting.
2. Rendering engine — the part that decides what the page looks like — the layout and paint engine turning markup and styles into pixels.
3. Bundling — shipping it already installed — including a product with an operating system, the core issue in the Microsoft antitrust case.

30.12 The dot-com boom and bust#

PLAIN30.12.1 in simple words#

1. In 1995 three things happened at once and together they lit the fire.
2. Windows 95 put a usable computer in ordinary homes, Netscape made the web easy, and the NSFNET ban on commercial traffic was gone.
3. Suddenly anyone could sell anything to anyone, and nobody knew how big that would be.
4. Netscape’s listing on 9 August 1995 taught investors a lesson they took too far: a company with almost no profit could be worth billions.
5. Money poured in. The accepted wisdom was “get big fast”: collect users now and worry about profit later. Users were called eyeballs.
6. The NASDAQ Composite index peaked on 10 March 2000 at 5,048.62, then fell to about 1,114 on 9 October 2002, a drop of roughly 78 percent.
7. Roughly 5 trillion US dollars of market value was destroyed by the end of 2002. Some companies survived and became the largest in the world. Most did not.

PLAIN30.12.2 a picture in your head#

1. Imagine a gold rush in a valley where gold has genuinely been found.
2. The gold is real. That matters. This was not a fantasy.
3. Eventually money stops arriving, and the diggers with no gold cannot pay wages and stop within weeks. The few who found real seams keep working and in twenty years own the valley.

Where this comparison breaks: gold is finite and the internet market was not. The survivors were building businesses whose costs fell as they grew, which is the opposite of mining. The bubble was in the pricing, not in the underlying opportunity, and telling those two apart is the hardest judgement in investing.

PLAIN30.12.3 a worked example#

1. Pets.com is the standard example. Founded November 1998, it ran a Super Bowl advertisement in January 2000 costing about 1.2 million dollars, went public in February 2000 at 11 dollars per share raising 82.5 million, and shut down on 9 November 2000, about nine months later.
2. The fatal detail: shipping heavy pet food cost more than the margin on it, so every sale lost money and growth made the loss bigger.
3. Webvan, founded 1996, raised 375 million dollars at its November 1999 listing and reached a valuation over 4.8 billion. In 2000 it had 178.5 million dollars of sales against 525.4 million of expenses, and filed for bankruptcy in July 2001 having lost over 800 million.
4. Boo.com, founded 1998 in London, spent 135 million dollars of venture capital in about 18 months and went into receivership on 18 May 2000.

5. Amazon’s share price fell around 90 percent from its peak and it survived anyway, largely because it had raised bond money before the market closed.
6. That is the uncomfortable truth: survival depended heavily on how much cash was in the bank when the funding stopped.

PLAIN30.12.4 what is really happening inside#

1. Separate four things that get blended together in the retelling.
2. The technology was real. Internet use grew right through the crash and never stopped.
3. The business models often were not. Free delivery of heavy goods at a loss does not become profitable at scale.
4. The valuations were unmoored. Companies were priced on visitor counts because there were no earnings to price on.
5. The financing was fragile. Most needed a new funding round every year, so when funding stopped they died within months.
6. Alan Greenspan used the phrase “irrational exuberance” about asset prices in December 1996, more than three years before the peak. Being early is the same as being wrong, until it is not.
7. The honest lesson is not that the internet was overhyped. Amazon in 2026 is larger than any 1999 forecast. The forecasts were not too optimistic about the internet. They were too optimistic about who would capture the value, and how soon.

TECHNICAL30.12.5 the engineer’s version#

1. NASDAQ Composite closing peak: 5,048.62 on 10 March 2000. Trough: about 1,114 on 9 October 2002, roughly 78 percent below the peak, giving up all the gains of the bubble.
2. From 1995 to March 2000 the index rose about 600 percent. In 1999 alone it rose 85.6 percent, while more listed stocks fell than rose.
3. The index did not regain its March 2000 closing level until 2015.
4. The infrastructure overbuild had a long tail. Enormous amounts of optical fibre were laid and left unlit. That dark fibre was bought cheaply later and is part of why bandwidth prices fell so far in the 2000s.
5. WorldCom, a major carrier of the era, collapsed in an accounting fraud, was renamed MCI in 2003 and acquired by Verizon in 2006.
6. A useful measure for judging any such period is burn rate: cash spent per month, divided into cash on hand, gives runway in months. Most failed dot-coms had a runway under 12 months at all times.
7. The pattern generalizes. Real technology, unclear business models and stretched valuations appear in later cycles too, and reasonable people disagree in 2026 about whether current AI spending is a repeat, an overshoot, or neither. Treat that as an open argument, not a settled fact.

WORDS30.12.6 remember these#

1. Eyeballs — a count of visitors used instead of profit — audience metrics used as a valuation proxy when there are no earnings.
2. Burn rate — how fast the money is spent — net cash outflow per month, which divided into reserves gives runway.
3. Dark fibre — cable laid but not switched on — installed optical fibre with no active transmission equipment attached.

30.13 Broadband and the physical internet#

PLAIN30.13.1 in simple words#

1. For most of the 1990s, home internet came through a modem over an ordinary phone line. A modem turns bits into sounds, because the phone network only carried sound.
2. That is why dial-up made a noise. The two modems were negotiating out loud: testing the line, agreeing a speed and measuring the echo.
3. The noise stopped once agreement was reached, because the speaker was switched off, not because the sounds stopped.
4. Then came ways to use the same copper without occupying the voice channel, which is what ADSL does. ADSL gives more download than upload speed on purpose, which is what the A means.
5. Cable television networks were adapted to carry data, and eventually fibre was run all the way into homes.
6. Underneath all of it sits the part almost nobody sees: over a million kilometres of cable on the sea floor, carrying about 99 percent of intercontinental data. Satellites carry a tiny fraction.

PLAIN30.13.2 a picture in your head#

1. Think of the phone line into a house as a wide pipe that was only ever used at one end.
2. Voice needs the frequencies a human ear cares about, roughly 300 to 3,400 cycles per second. The copper can carry far more than that.
3. Dial-up squeezed data into the voice part of the pipe, so it competed with speech and was capped near 56 thousand bits per second.
4. ADSL puts data in the unused higher frequencies and leaves the voice part alone, with a small filter at each phone keeping them apart.
5. The download side gets the bigger share because that is what people mostly do, and because upstream signals are weakest exactly where interference is worst. Fibre replaces the pipe with glass, where the usable range is thousands of times wider.

Where this comparison breaks: a water pipe’s capacity does not fall with length, but copper’s does, badly. ADSL speed drops steadily with distance from the exchange and is usually unusable past about 4 to 5 km. Fibre does not care about a few kilometres, which is why the industry moved to it rather than squeezing copper further.

PLAIN30.13.3 a worked example#

1. Follow one home connection through its generations.

2. Take a 4-megabyte photograph, which is 32 million bits.
3. On a V.90 modem at 56 kbit/s that is about 570 seconds, so nearly ten minutes. On ADSL2+ at 20 Mbit/s it is about 1.6 seconds. On a 300 Mbit/s fibre link it is about 0.1 seconds.
4. Now the ocean. A single modern transatlantic cable such as MAREA is rated at about 224 terabits per second.
5. That one cable could move that photograph roughly 7 million times per second, which is why cables and not satellites carry the world’s traffic.

PLAIN30.13.4 what is really happening inside#

1. A submarine cable is thinner than you expect: about 25 mm across in deep water, weighing roughly 1.4 tonnes per kilometre.
2. Inside are a few pairs of glass fibres, a steel strength member, a copper conductor and layers of insulation and water barrier.
3. The copper is not for data. It carries direct current at several thousand volts to power repeaters on the sea floor, roughly every 60 to 100 km.
4. A cable ship lays it slowly, and in shallow water a plough buries it about a metre under the seabed to protect it from anchors and trawlers.
5. Cables break often. About 200 faults happen worldwide each year, and about two-thirds are caused by fishing gear and ships’ anchors.
6. Repair means sending a ship, grappling the cable up, cutting it, splicing in a new length and lowering it back. That takes days to weeks depending on weather, permits and ship availability, which is why redundancy matters more than repair speed.
7. Satellites are different. A geostationary satellite sits about 35,786 km above the equator, so light takes about 240 milliseconds for the round trip before any processing. Low earth orbit constellations fly around 550 km up, which cuts that delay enormously at the price of needing thousands of satellites.

TECHNICAL30.13.5 the engineer’s version#

1. Modem standards, with years: Bell 103 at 300 bit/s in 1962; V.22bis at 2,400 in 1984; V.32 at 9,600 in 1984; V.32bis at 14,400 in 1991; V.34 at 28.8 then 33.6 kbit/s from 1994 to 1996; V.90 at 56 down and 33.6 up in 1998; V.92 in 2000, raising upload toward 48 kbit/s.
2. The dial-up handshake is a defined training sequence: answer tone, echo canceller disabling tone, probing signals to measure the channel, then rate negotiation. It is audible only because the monitor speaker is on by default, controlled by the ATM and ATL commands.
3. ADSL, ITU-T G.992.1, uses discrete multitone modulation. In Annex A, 26.075 to 137.825 kHz carries upstream and 138 to 1104 kHz downstream. ADSL2+ is G.992.5, reaching about 24 Mbit/s down. VDSL2 is G.993.2.
4. Asymmetry has two causes: crosstalk is worst at the exchange end where the upstream signal is already weakest, and demand is genuinely asymmetric. The first is engineering, the second is marketing. Both are real.
5. ISDN basic rate is 2B+D: two 64 kbit/s bearer channels plus a 16 kbit/s signalling channel, so 128 kbit/s bonded.
6. Submarine facts as of early 2026: more than 600 active and planned cables, over 1.5 million km in service, carrying about 99 percent of international data traffic. Fault rates fell from about 3.7 per 1,000 km per year between 1959 and 1979 to about 0.44 after burial became standard in 1985.

7. Starlink reported approximately 10,400 satellites in orbit and more than 12 million subscribers as of June 2026, with typical latency around 25 to 35 ms.
8. India specifically. Public internet began on 15 August 1995, provided by Videsh Sanchar Nigam Limited. TRAI reported about 918 million internet subscribers as of 30 September 2023, of which about 881 million were wireless and about 37 million wired.
9. India is served by roughly 17 international submarine cables landing at about 5 main landing points, chiefly Mumbai and Chennai, including SEA-ME-WE 3, Bay of Bengal Gateway and Europe India Gateway. Reliance completed the IAX and IEX systems in 2024.
10. That concentration is the real chokepoint. A small number of landing stations in two cities serves 1.4 billion people, and cable damage in the Red Sea and near Egypt has repeatedly degraded Indian and Gulf connectivity because so many routes share one corridor.
11. The reader’s own traceroute shows this geography directly: Delhi, then Mumbai, then Pune, all inside Microsoft’s backbone, because that is where the capacity and the landing stations are.
12. Speedtest Global Index figures for India in November 2023 were about 94.62 Mbit/s mobile download and 58.62 Mbit/s fixed download. Mobile being faster than fixed is unusual worldwide and reflects how India built out.

WORDS30.13.6 remember these#

1. Modem — the box that turned bits into sound — a modulator and demodulator converting digital data to and from an analogue channel.
2. ADSL — fast down, slower up, over phone copper — asymmetric digital subscriber line, ITU-T G.992 series, using discrete multitone modulation.
3. Repeater — the thing on the sea floor that keeps the light strong — an optical amplifier, usually erbium-doped, powered over the cable’s copper.
4. Landing station — where a sea cable comes ashore — the terminal facility that powers the cable and connects it to terrestrial networks.
5. Chokepoint — a place where too much depends on too little — a corridor or facility whose failure degrades connectivity for a whole region.

30.14 Who actually runs the internet#

PLAIN30.14.1 in simple words#

1. Nobody owns the internet, and that is not a slogan. There is no company or government that can switch it off.
2. But some things must be agreed globally or nothing works.
3. Names must be unique, so two people cannot own the same domain. Addresses must be unique, so a packet has one destination. Protocols must be shared, so machines built by strangers can talk.
4. Three kinds of body handle those three jobs, and none controls the network.
5. ICANN coordinates names and, through IANA, the address and number pools.
6. The IETF writes the protocol specifications, published as RFCs. The W3C writes web standards, and five regional registries hand out address blocks in their own parts of the world.
7. Everything else is contracts between companies, and national law.

PLAIN30.14.2 a picture in your head#

1. Think of the world’s postal systems. No single body runs them.
2. Someone must agree country codes, or letters cannot be sorted. Someone must agree envelope sizes and address layout so machines can read them. Small, boring, essential jobs.
3. The internet has the same shape: a tiny agreed core and an enormous uncoordinated periphery.

Where this comparison breaks: postal agreements are made between governments in a treaty organization. Internet standards are made largely by engineers and companies, with governments as one voice among several. That is the multistakeholder model, and several governments have objected to it for decades, wanting state control instead. The argument is unresolved.

PLAIN30.14.3 a worked example#

1. Here is how a proposal actually becomes an internet standard.
2. Someone writes an Internet-Draft and posts it. Anyone may do this, with no membership required.
3. A working group discusses it on a public mailing list, archived and readable by anyone. Implementations are expected, because working code carries more weight than argument.
4. When the group reaches rough consensus, meaning no serious unanswered objection rather than a majority vote, it goes to the steering group.
5. If approved it is published as an RFC with the status Proposed Standard. If it proves itself in deployment it can be raised to Internet Standard, and most useful protocols never bother.

Idea -> Internet-Draft -> working group list ->
running code -> rough consensus -> IESG review ->
RFC (Proposed Standard) -> deployment ->
Internet Standard

6. Meetings famously test consensus by humming rather than voting, so volume rather than headcount is judged and nobody is formally counted.

PLAIN30.14.4 what is really happening inside#

1. Keep the layers of governance apart, because they are often confused.
2. Unique identifiers. ICANN coordinates the domain name root and the number pools, and decides which top-level domains exist.
3. Protocol specifications. The IETF decides how TCP, HTTP and DNS behave. It has no power over who uses them.
4. Everything else. Which cables exist, who peers with whom, what content is allowed and what is censored is decided by companies and states.
5. That is why “who runs the internet” has an uncomfortable answer: for the things people care about most, it is your government and your internet provider. Net neutrality is exactly the argument about that last point.

TECHNICAL30.14.5 the engineer’s version#

1. ICANN was incorporated in California on 30 September 1998, taking over functions previously performed by Jon Postel at USC-ISI. It does not control content, spam, or what any network carries.
2. The IANA functions contract with the US National Telecommunications and Information Administration ended on 1 October 2016, transferring stewardship to the global multistakeholder community.
3. The IETF held its first meeting on 16 January 1986, with 21 US federal-government-funded researchers attending.
4. RFC 1 was published on 7 April 1969 by Steve Crocker of UCLA, titled “Host Software”. Jon Postel edited the series from 1969 until his death in 1998.
5. The process is defined in RFC 2026, “The Internet Standards Process, Revision 3”. RFC 6410, October 2011, cut the maturity ladder from three levels to two by retiring Draft Standard.
6. “Rough consensus and running code” comes from David Clark’s 1992 IETF presentation, where he said the group rejects kings, presidents and voting, and believes in rough consensus and running code.

7. The IANA free pool of IPv4 addresses was exhausted on 3 February 2011, when the last five blocks went to the five registries. APNIC, which covers India, effectively ran out in April 2011, earliest of all.
8. The W3C was founded in October 1994 by Tim Berners-Lee at MIT. Since 2019 the HTML and DOM specifications are maintained by the WHATWG with W3C endorsement, after years of the two bodies publishing rival versions.
9. Net neutrality in the United States, in dates: the FCC voted for Title II common-carrier treatment on 26 February 2015, effective 12 June 2015; repealed it on 14 December 2017, effective 11 June 2018; voted to restore it on 25 April 2024; and the Sixth Circuit struck that down in January 2025. As of 2026 there is no federal rule, though state laws such as California’s SB 822, signed 30 September 2018, remain.
10. Both sides, stated fairly. For: without a rule a provider can slow or charge for competing services, and small entrants cannot negotiate. Against: broadband needs heavy investment, traffic genuinely differs in requirements, and common-carrier rules designed for telephones discourage building. Serious economists hold both views.
11. India took a strong position. TRAI prohibited discriminatory pricing of data services in February 2016, ending zero-rating schemes, and recommended full net neutrality rules in November 2017.

WORDS30.14.6 remember these#

1. ICANN — the body that coordinates names and numbers — the corporation overseeing the DNS root, gTLD policy and the IANA functions.
2. RFC — the document a standard is written in — a numbered publication of the RFC series, of which only some are standards-track.
3. Rough consensus — no serious unanswered objection — the IETF decision rule, explicitly not a majority vote.
4. Net neutrality — the rule that providers must not favour some traffic — the principle that carriers should not block, throttle or paid-prioritize.

30.15 Web 1.0 to now#

PLAIN30.15.1 in simple words#

1. The first web was files. Someone wrote an HTML file, put it on a server, and everyone who asked got the same file.
2. Then pages started being built when you asked for them, from a database, so a shop could show your basket rather than a fixed page.
3. From about 2004 the phrase Web 2.0 described sites where the users made the content: blogs, wikis, photos, video and comments.
4. Social networks turned that into the main use of the web, and smartphones changed it again. From around 2016 more web traffic came from phones than from computers.
5. Apps took a large share of that time back from the web, inside private platforms with their own rules.
6. Running your own server became rare, and most services now run on cloud infrastructure owned by a handful of companies.
7. So traffic and attention both concentrated, on a network originally designed to have no centre.

PLAIN30.15.2 a picture in your head#

1. Imagine a town that starts as thousands of small shops, each owner-built.
2. Then a few large market halls open. They are cleaner, easier and free to enter, so almost everyone moves inside.
3. The streets still exist and the roads are still public, but nearly all the trade happens in five buildings, and each landlord sets the rules inside.
4. Nobody forced anyone. Each move was individually sensible.

Where this comparison breaks: a shop that leaves a market hall can reopen on the street. On the web, leaving usually means losing the audience, because the audience is a feature of the hall rather than of the shop. That lock-in has no clean equivalent in a market town.

PLAIN30.15.3 a worked example#

1. Compare the same task across eras: publishing a photograph for friends.

2. In 2026 anyone can do it in four seconds, and cannot control who sees it, how long it lasts, or whether it is shown at all.
3. That is the trade in one table: an enormous gain in access, and a real loss of control.

PLAIN30.15.4 what is really happening inside#

1. Say plainly what was gained. Publishing became free and instant for billions of people who had no voice before. Search made the whole archive findable. Encryption became normal, so most traffic is private in transit.
2. Say plainly what was lost. Independence: most content now lives on servers the author does not control.
3. Durability: pages vanish when a platform closes, and nobody is obliged to keep them.
4. Interoperability: an early web page could be read by any client, and a modern app feed usually cannot.
5. Privacy in a different sense: the cost of free services is measurement of behaviour at a scale nobody imagined in 1993.
6. Resilience: a design meant to have no single point of failure now leans heavily on a few cloud and content-delivery providers, so single configuration errors have caused worldwide outages.
7. None of that was forced by the technology. It followed from economics: scale is cheaper, and defaults decide behaviour.

TECHNICAL30.15.5 the engineer’s version#

1. The phases in technical terms: static file serving; then server-side generation with CGI (RFC 3875) and later application servers; then asynchronous updates in the browser using XMLHttpRequest, which made single-page applications practical from around 2005.
2. The term Web 2.0 was popularized by the O’Reilly Media conference of that name held from 2004. It described a business and design pattern, not a protocol version. There was never a Web 1.0 specification.
3. Mobile passed desktop in worldwide web page views around late 2016 by Statcounter’s measurement, and has stayed the majority since.
4. Encryption in transit became the default rather than the exception. Let’s Encrypt began issuing free certificates in 2015, and the share of page loads over HTTPS in Firefox telemetry rose from under 30 percent in 2014 to above 80 percent by 2020.
5. Content delivery networks now terminate a large share of connections close to users, so your packets often never reach the origin server’s country. That is exactly what the reader’s trace shows, ending inside Microsoft’s own network rather than at a distant data centre.
6. The honest assessment: the internet’s core design remains decentralized and works. The application layer above it has centralized heavily. Those are two separate facts, often wrongly merged into one complaint.

WORDS30.15.6 remember these#

1. Web 2.0 — sites built from what users post — a design and business pattern from around 2004, not a technical standard.
2. Content delivery network — copies of a site kept near you — distributed caching and connection termination close to end users.
3. Lock-in — hard to leave without losing something — switching costs created by data, audience or formats that are not portable.

30.16 A master internet timeline#

1. One table, from the semaphore to today. Every date here is discussed somewhere in this chapter.

30.98 Common wrong ideas#

1. Wrong: the internet was built to survive a nuclear war. Right: Paul Baran’s RAND work had a survivability motive, but ARPA’s director stated plainly that the ARPANET was not started for that. It was built for resource sharing among scarce research computers.
2. Wrong: Al Gore claimed he invented the internet. Right: in a 1999 interview he said he took the initiative in creating the internet, meaning legislative and funding support. He sponsored the High Performance Computing Act of 1991, and Vint Cerf and Bob Kahn publicly defended his record. He never claimed to have invented it.
3. Wrong: the internet and the web are the same thing. Right: the internet is the network. The web is one application on it, created eighteen years later. Email, video calls and app traffic use the internet without the web.
4. Wrong: the internet is wireless now. Right: your last few metres may be wireless, but traffic goes onto fibre almost immediately, and about 99 percent of international traffic runs through cables on the sea floor.
5. Wrong: satellites carry most long-distance internet traffic. Right: they carry a very small fraction. Cables are far cheaper, far faster and far lower in delay.
6. Wrong: one organization runs the internet and could switch it off. Right: ICANN coordinates names and numbers, the IETF writes specifications, and nobody operates the whole network. Governments can cut off their own country, which is a different thing.
7. Wrong: Tim Berners-Lee invented hypertext. Right: hypertext ideas date to Vannevar Bush in 1945 and Ted Nelson in the 1960s. He combined URLs, HTTP and HTML into something anyone could deploy without permission.
8. Wrong: the dot-com crash proved the internet was hype. Right: usage kept growing right through the crash. The pricing and the business models were wrong, not the direction.
9. Wrong: Leonard Kleinrock invented packet switching. Right: he did important queueing analysis and later ARPANET performance work, but primary sources and historians credit Baran and Davies with the independent invention, and Kleinrock acknowledged in 2023 that his early published work was about message switching.
10. Wrong: TCP/IP won because it was technically the best design. Right: it won because it existed, worked, cost nothing, shipped in Berkeley Unix, and had a guaranteed government market while OSI was still being drafted.

30.99 Chapter summary in 20 lines#

1. Long-distance messaging began by making a signal visible from the next hill, then giving it parts so it could spell anything.
2. Chappe’s towers, from 1792, cut Paris to Lille from two days to about 32 minutes, and introduced encoding, relaying and a shared code book.
3. Electricity removed the need for line of sight. Cooke and Wheatstone patented in May 1837 and Morse’s line ran on 24 May 1844.
4. The 1858 Atlantic cable managed about a word every ten minutes and died in weeks. The 1866 cable worked and shrank the world permanently.
5. The telephone, patented by Bell on 7 March 1876, needed exchanges, and Strowger automated them from 1891.
6. Circuit switching books a path for a whole call. That suits speech and wastes almost everything on a bursty computer conversation.
7. Packet switching splits messages into addressed pieces that travel independently and are reassembled at the far end.
8. It was invented three times: Baran at RAND from 1960 with a survivability motive, Davies at NPL in 1965 who named the packet, and Kleinrock supplying queueing analysis.
9. The ARPANET was built for resource sharing, not nuclear survival. Its first message, on 29 October 1969, was “LO” before the far end crashed.
10. The IMP was the first router, and its key idea was that hosts never had to agree with each other, only with the box beside them.
11. Email was not in the plan. Tomlinson joined two programs in 1971, chose the @ sign, and messaging became the network’s main use.
12. Cerf and Kahn’s May 1974 paper solved internetworking by putting a common protocol over dissimilar networks instead of making them agree.
13. TCP and IP were split in 1978, giving a thin waist: many technologies below, many applications above, one narrow protocol between.
14. On 1 January 1983 the ARPANET switched to TCP/IP in a single flag day. Nothing that size can be done that way again, which is why IPv6 has taken thirty years.
15. TCP/IP beat X.25, Minitel, private online services and OSI because it had running code, no fees, US government demand and simplicity.
16. NSFNET carried the research internet from 1986, banned commercial traffic, then lifted the ban and shut down on 30 April 1995, leaving a mesh of commercial networks and exchange points.
17. The web was proposed in March 1989, working by the end of 1990, and put into the public domain on 30 April 1993. It is an application on the internet, not the internet.
18. Mosaic in 1993 and Netscape’s 9 August 1995 listing started both the browser wars and the dot-com boom. The NASDAQ peaked on 10 March 2000 at 5,048.62 and fell about 78 percent.
19. Physically the internet is fibre: over 600 submarine cables and more than 1.5 million kilometres, carrying about 99 percent of international traffic, with India served by a handful of landing points.
20. The core stayed decentralized and the applications did not, which is the central tension of the internet in 2026.