The Five Generations of Computers

By Victor Hoang · Updated Oct 21, 2026

In February 1946, The New York Times put a machine on its front page. The Computer History Museum records that it was ENIAC, the Electronic Numerical Integrator and Computer. The United States Army had built it to calculate ballistics tables.

The five generations of computers sort machines by their switching technology. They run from vacuum tubes in the 1940s and 1950s through transistors from the 1950s, integrated circuits from the late 1950s and microprocessors from 1971. A fifth generation, named by Japan's government in 1982, aimed at knowledge processing on parallel machines and is summarized as AI. The labels were built after the fact, so sources disagree on where each one ends. ENIAC lost roughly one vacuum tube every day or two, and this article follows that component. For the dated chronology, see Computer Timeline. For the argument over who built the first one, see Who Invented the Computer?.

Timeline at a Glance

Gears and Relays (c. 300 BCE-1940s)

Before the numbered generations, calculation was mechanical

Picture a market in the Greek world around 300 BCE: a merchant pushes pebbles across lines scratched into a board to total a bill. The oldest surviving counting board is the Salamis tablet, a marble slab found on the island of Salamis in 1846 and dated to around 300 BCE, according to Toronto Metropolitan University's abacus history. The same page notes that these boards were aids to a human doing the arithmetic, not calculators in the modern sense, and that the Chinese suan-pan abacus was first chronicled around 1200 CE.

Mechanism came next. The Antikythera mechanism, recovered from a shipwreck, was an analog device with at least 30 gear wheels, and ASME describes it as used to forecast lunar and solar eclipses, keep the calendar and set the dates of the Panhellenic Games. ASME dates it to the second century BCE. Then, in the 1800s, Charles Babbage designed automatic computing engines, and the Computer History Museum says he "invented computers but failed to build them." The museum finished his Difference Engine No. 2 in London in 2002, 153 years after the design.

Nobody called any of this a generation. These were precursors to the concept, and the generation scheme starts only once electricity does the switching. But the 1940s gave a bridge. The Computer History Museum timeline describes the Zuse Z3 of 1941 as a machine with 2,300 relays that performed binary floating point arithmetic, and the Harvard Mark 1, conceived by Howard Aiken and built by IBM, as a room-sized, relay-based calculator with 3,500 relays. Relays are electromagnetic switches, so they clicked and wore out. The problem they left behind was speed: a mechanical contact can only move so fast.

For teams today: the abacus and the relay were each "good enough" until someone needed a thousand times more throughput. If your process works only because it's slow and forgiving, ask what happens at ten times the volume.

The Vacuum Tube Age (1940s-1950s)

Glowing glass makes calculation electronic, and fragile

Walk into the room that held ENIAC and you'd see racks of glowing tubes and cables everywhere. PBS lists 17,468 vacuum tubes, 70,000 resistors and 1,500 relays, and says changing the program meant rewiring it with plugboards and switches, which could take a team two days. The Computer History Museum says John Mauchly proposed it in 1942, that it was built from 1943 to 1945 for wartime ballistics, and that crews eventually learned to fix a fault in about 15 minutes. (The two sources count tubes slightly differently. Both land near 18,000.) The same museum's timeline says the first Colossus at Bletchley Park became operational in 1944, designed by Tommy Flowers to break Lorenz ciphers, with as many as 2,500 tubes in each of the ten machines delivered.

Tube machines became products at the start of the 1950s. The Computer History Museum records that the US Census Bureau received the first UNIVAC I on March 30, 1951. It could do 1,905 operations per second and stored data on magnetic tape. Retrospective textbooks call all of this the first generation, and it ran on government, military and university money.

This era solved speed, because tubes switch electronically. It created two new problems. Reliability was one, since tubes burned out constantly. Programming was the other, because instructions lived in wiring and in raw machine code. IBM's FORTRAN history says programmers hand-coded in rows of zeros and ones, spent up to half their time debugging, and that FORTRAN, created from 1954 under John Backus and released commercially in 1957, could turn a 1,000-instruction job into 47 lines.

For teams today: the first generation shows what happens when only specialists can operate the tool. If a system needs a "two-day rewire" every time the work changes, that's a signal to simplify the interface, not to hire more specialists.

The Transistor Age (1950s-1960s)

A solid-state switch replaces the tube and the headaches

In December 1947 at Bell Labs, a small germanium device amplified a signal. The Computer History Museum says John Bardeen and Walter Brattain achieved the first successful semiconductor amplifier on December 16, 1947, and that the aim was to replace unreliable vacuum tubes and electromechanical switches in the telephone system. They shared the 1956 physics prize with William Shockley, and the Nobel Prize cites "their researches on semiconductors and their discovery of the transistor effect."

Computers took a while to follow. The museum's Silicon Engine timeline shows a transistorized computer prototype emerging in 1953, and its mainframe overview says "second generation" transistor machines were replacing vacuum-tube ones in the late 1950s. The machine that made the switch visible to ordinary buyers was the IBM 1401. IBM announced it on October 5, 1959, calling it one of the first computers to run on transistors instead of fragile tubes, with more than 5,200 orders in the first five weeks and over 10,000 systems installed by the mid-1960s. A 1401 could be leased from $2,500 a month, against $10,000 a month for comparably equipped machines and more than $1 million to buy a large mainframe outright.

Reliability and cost fell together, and software became a business. Compilers like FORTRAN let engineers write programs without a machine-code specialist. The new problem was incompatibility. Each model had its own instruction set, so moving to a faster machine meant rewriting your programs.

For teams today: a better component only helps if the surrounding process can use it. The transistor existed in 1947, and the 1401 reached the mass market in 1959, so budget for a long gap between "invented" and "installed."

The Integrated Circuit Age (1958-1970s)

Whole circuits grow on one piece of silicon

In 1958, Texas Instruments engineer Jack Kilby built a circuit whose resistors, capacitors and transistors were all made from semiconductor material. The Computer History Museum dates the demonstration of his "solid circuit" to 1958, and Robert Noyce's patent for a manufacturable monolithic integrated circuit to 1959. Kilby later shared the 2000 physics prize, which the Nobel Prize awarded to him "for his part in the invention of the integrated circuit." The museum notes that the first computer uses were in aerospace in 1962, because the size and weight savings justified the higher price, and that large manufacturers announced chip-based mainframes in 1965.

The machine that stands for this generation is IBM's System/360, launched on April 7, 1964. IBM's history calls it a $5 billion gamble, replacing five product lines with one strictly compatible family, so programs written for one model ran on any other. Six processor models covered a fiftyfold performance range, and it introduced the 8-bit byte. Orders passed 1,000 in the first month. IBM describes it as "the beginning of a new generation of computers." That's the closest IBM's own history page comes to the numbered label.

Here sources start to disagree. System/360 is routinely taught as the third generation, but the Computer History Museum explains that its Solid Logic Technology modules were hybrids, small ceramic units carrying transistor chips and other parts, because monolithic chips couldn't yet meet the cost and speed demands of large computers. So the flagship "integrated circuit" computer of 1964 used a bridge technology. The same timeline dates Gordon Moore's prediction about transistor density to 1965, and it became the industry's planning yardstick.

The problem this age created was lock-in. Compatibility was a promise that kept customers on one vendor's family for decades, and the next generation had to undercut it by getting small and cheap enough to escape the data center.

For teams today: compatibility is a feature you pay for twice, once in the purchase and once when you want to leave. Before choosing a platform, ask what migrating off it would cost.

The Microprocessor Age (1971-1990s)

A whole CPU on one chip reaches desks

The 1971 chip that began this era first went into a calculator. The Computer History Museum credits Ted Hoff and Stanley Mazor with conceiving Intel's 4004 and Federico Faggin, assisted by Masatoshi Shima, with squeezing its roughly 2,300 transistors into a 16-pin package. It was a 4-bit CPU built for Busicom's calculator. A computer's brain had become a part you could buy.

Hobbyists noticed. In 1975 the Altair 8800 kit appeared on the cover of Popular Electronics, and the Computer History Museum says MITS sold it for $297, or $395 with a case, with 256 bytes of memory, and that Bill Gates and Paul Allen licensed their BASIC to it. The reason a business would buy one arrived in 1979: the Museum calls the VisiCalc spreadsheet "the 'killer app'" that spurred Apple II sales. Then came the standard. IBM unveiled the Personal Computer on August 12, 1981, at $1,565, with an Intel 8088, 16 kilobytes of RAM and an operating system from Microsoft.

Textbooks call this the fourth generation and tie it to large-scale integration, meaning ever more circuits per chip. The shift here wasn't a new kind of switch. The switch was already a transistor. What changed was where the machine lived and who owned it. A computer moved from a computing center into a person's office, and programming style moved with it: a BASIC interpreter or a spreadsheet instead of a job deck handed to an operator.

The new problem was fragmentation. When every desk has its own computer, files and spreadsheets scatter, and the work of connecting them starts.

For teams today: the tools that win are the ones the person closest to the work can use without a specialist. That was true of the spreadsheet in 1979 and it's the test for any AI tool now.

The Fifth Generation Question (1982-present)

A government names the next leap, and the scheme starts to strain

This era starts with a government program rather than a chip. In 1982, Japan's Ministry of International Trade and Industry started the Fifth Generation Computer Project, according to the Information Processing Society of Japan's computer museum. It called the project "Fifth Generation" to signal technology that was creative and pioneering, with a technical goal of "research and development on new computer technology oriented toward knowledge-based information processing." The Institute for New Generation Computer Technology (ICOT) led the work. The same page says five parallel inference machine models were built, including the PIM/p with 512 processors, that about 54 billion yen was invested over 11 years, and that the project came to an end in 1992.

The world paid attention. In 1983 Edward Feigenbaum and Pamela McCorduck published The Fifth Generation (Addison-Wesley), and a Christian Science Monitor review of it reported the authors' warning that the US was "squandering our lead at the rate of one day per day." The authors of one outside evaluation wrote that many research programmes and institutes were conceived in reaction to ICOT. That evaluation by Seif Haridi and Siwert Sundström of the Swedish Institute of Computer Science, published in ICOT's final evaluation report, takes a mixed view. It says the project made solid progress in applying parallel processing to symbolic computing and that its internal technical results exceeded expectations, but also that "the project remains unfinished," and it recommends porting the software to widely available machines. The authors say ICOT did more to influence other research than it absorbed.

So what did the fifth generation achieve? A new class of research prototypes and a body of work on parallel logic programming, yes. A commercial machine class that displaced the PC, no. The label also differed from the earlier four in kind. Tubes, transistors, chips and microprocessors were named after the components inside. "Fifth" was named for an ambition, and the machines that delivered large-scale AI came largely from graphics chips and the cloud.

Those two arrived in the 2000s and 2010s. Amazon says AWS launched publicly on March 14, 2006 with one service, S3 for storage. In 2012 the AlexNet team of Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained a neural network on two Nvidia cards using CUDA code, according to IEEE Spectrum. The hardware story since then, from AI chips to AI data centers, is about parallel arithmetic at huge scale. Parallel hardware aimed at AI was Japan's bet too, built on different parts. The longer software story is in The History of AI and the AI Timeline, and the definition is in what artificial intelligence is.

Where does the scheme stop fitting? Some tutorials call GPUs, cloud or quantum machines a "sixth generation." That is usage, not a standard. There's no single new switch to point at, since today's machines still run on transistors packed far more densely, and the biggest changes are in architecture (parallel chips, rented data centers) and in how people instruct machines. The generation labels work best as a map of the first four jumps. After that, they describe a direction more than a component.

For teams today: when a vendor calls its product "next generation," ask what actually changed: the component, the architecture, or just the interface. Judge it on your workflow's measured results.

The Generations Side by Side

The "Approx. years" column is approximate on purpose. The transistor appeared in 1947, a prototype computer used it in 1953, and the 1401 shipped in 1959, so "second generation" can start anywhere across twelve years depending on which event an author counts. The third generation has a similar spread (1958 demonstration, 1959 patent, 1964 System/360, 1965 chip-based mainframes).

Key People

Key Facts

• ENIAC used about 18,000 vacuum tubes and lost one every day or two.
• IBM launched System/360 on April 7, 1964, expected to cost $5 billion over four years, and replaced five product lines with one compatible family.
• System/360 used hybrid Solid Logic Technology modules, not monolithic chips, because chips could not yet meet cost and speed demands for large computers.
• The 4004 packed about 2,300 transistors into a 16-pin package in 1971.
• Japan's Fifth Generation project spent about 54 billion yen over 11 years and ended in 1992.
• The IBM PC launched on August 12, 1981 at $1,565.

The Pattern Behind the History

Across the five generations, the same four steps repeat. We call it the Switch, Stall, Spread, Label sequence. It's our synthesis of the history above, not a published study.

1. Switch. A lab invents a better switch: the transistor in 1947, the integrated circuit in 1958 and 1959, the microprocessor in 1971.
2. Stall. Machines adopt it slowly, or in a half-step. The 1401 came twelve years after the transistor, and System/360 used hybrid modules before chips were ready.
3. Spread. Cost falls and the machine moves closer to the user, from government room to company data center to desk.
4. Label. Writers name the generation afterward, and the label arrives with fuzzy edges. The fifth generation reversed the order by naming the ambition first.

What It Teaches Teams Today

• Judge tools on workflow results, not generation labels. "Next generation" is a marketing phrase, as it was in 1964.
• Expect a lag between invention and payoff. Budget years, not months, between a technology appearing and your team using it well.
• Pay attention to compatibility. System/360's compatibility promise created whole ecosystems, and also lock-in.
• Be wary of ambition-first naming. A fixed plan named for an outcome can miss a faster route that arrives from another direction, as GPUs and the cloud did for AI.
• Put the capability near the person doing the work. Each generation spread because the tool moved closer to the user. The same test applies to on-device AI and to any new platform. For the full story of that shift, return to The History of Computers.

The generations are a useful map, but the lesson underneath them is simpler: each time the machine got smaller and cheaper, the people who benefited most were the ones nearest the work. Use that as your test the next time a vendor promises a new generation.