Computer Timeline: From the Abacus to AI PCs

By Victor Hoang · Updated Oct 21, 2026

In 1821 Charles Babbage and the astronomer John Herschel sat checking tables of numbers by hand. They found error after error. As the Computer History Museum records it, Babbage said: "I wish to God these calculations had been executed by steam."

The computer timeline starts long before electricity, with stone counters on marked boards and a Greek geared calendar device from the late second century BCE. It runs through Babbage's engines, the room-sized machines of the 1940s, the microprocessor in 1971, the PC in 1981, the cloud in 2006 and today's AI PCs. The throughline is a steady build. Each step made calculation cheaper, smaller or easier to reach. Navigators, engineers and insurers had long depended on printed tables that people copied by hand. This page is the dated reference. For the manager's view, read The History of Computers.

Two questions get their own pages. Who gets the title of inventor is argued in Who Invented the Computer?, and the five-generation classification of hardware is in Generations of Computers. This page just gives you the dates.

Timeline at a Glance

The Counting Boards (300s BCE-1670s)

Stones on lines, then gears, then drums

Picture stone counters moved along lines scratched on a table. The Computer History Museum says early abacuses were probably stone markers placed on lines in the dirt, and that others used beads on rods or beans on a lined board or cloth. The oldest surviving example it shows is the marble Salamis tablet, found on the Greek island of Salamis, which the Museum dates to around the 4th century BCE and notes may be Babylonian. The abacus then lasted. The same page calls it perhaps the oldest continuously used calculating tool aside from fingers, and recounts a 1946 contest in Tokyo in which a Japanese postal worker with a soroban beat a US soldier with an electric calculator in four of five rounds.

The ancient world also built something far stranger. The Antikythera Mechanism is described in a 2006 Nature paper as a geared device constructed around the end of the second century BCE. The paper says it calculated and displayed celestial information such as the phases of the moon and a luni-solar calendar, and that it predicted lunar and solar eclipses using Babylonian arithmetic-progression cycles. The authors found in its gearing a mechanical version of Hipparchos's theory of the Moon's uneven motion, and called that an unexpected degree of technical sophistication for the period.

The next steps came in the 1600s, and they were about arithmetic. Napier's rods, which the Computer History Museum catalogs as a palm-sized calculator designed in 1617, turned multiplication into reading numbers off a set of rods. In 1623 Wilhelm Schickard wrote to the astronomer Johannes Kepler about a machine that, per the Museum's exhibit, combined Napier's rods for multiplication and division with a toothed-wheel system to add and subtract. His design survived only in sketches and letters that stayed lost until 1935.

In the 1640s, Blaise Pascal built an adding machine for his father, a tax collector. The same exhibit says he built about 50 Pascalines, the first adding machine made in even modest quantity, though he sold only about 15, and that the fragile carry mechanism limited its success. In 1673 Gottfried Leibniz built what the Museum calls the first true four-function calculator. One gear could represent any digit from 0 to 9 in a single turn, and the drum-shaped design underpinned calculators for roughly 275 years. Only one of Leibniz's versions survives.

The problem this era solved was speed and accuracy for one calculation at a time. The problem it left was that every sum still needed a person to start it, and nobody had a way to give a machine a list of instructions.

For teams today: the first computing tools sat with the person doing the work. If a tool is simple enough to carry in a pocket and check by eye, people will use it for thousands of years.

Babbage's Engines (1801-1930s)

Cards, cranks and a census turn calculation into data

In 1801, in Lyon, Joseph Marie Jacquard demonstrated a loom that ran on a chain of cards punched with holes. The holes decided which warp cords were raised on each pass of the shuttle, so the loom could reproduce complex silk patterns on its own. It was a program, stored on a card, long before anyone used the word that way.

Babbage knew the idea. According to the Museum, he described plans in 1837 to use punched "number cards" to feed programs and data into his Analytical Engine. He had started in 1821 with the Difference Engine, and the Museum's history says he conceived the Analytical Engine in 1834 as a general-purpose programmable machine whose design marks the transition from mechanized arithmetic to general-purpose computation. The Difference Engine, by contrast, is strictly a calculator that works by repeated addition. In 1843, Ada Lovelace published her Notes on the Analytical Engine, which included descriptions of the steps it would take to solve certain problems, the first published descriptions of what we'd now call programs. She also speculated that numbers could stand for letters or notes of music, so a machine might manipulate symbols and not only quantities.

Babbage never finished his engines. The Museum notes that a complete Difference Engine No. 2, built from his drawings with 8,000 parts and weighing five tons, was finished in London in 2002, 153 years after the design. The first big customer for punched cards was a government, not a scientist. The US census had a processing crisis, and Hollerith won the contest to solve it. According to the Museum, about 60 million cards were punched for the 1890 census, an average operator could process around 7,000 cards a day, roughly ten times faster than by hand, and the Census Bureau was still using essentially the same system seventy years later. Hollerith founded the Tabulating Machine Company in 1895.

The era solved the volume problem: a machine could now count a whole nation. It created a labor problem that sounds familiar. Somebody had to prepare, punch, stack and feed millions of cards, and the work moved from thinking to feeding the machine.

For teams today: representing work as structured data was the breakthrough, and re-keying data between systems is still the punched-card job in modern clothing.

The Room-Sized Machines (1940-1970)

War funds electronics, then business buys the result

This era overlaps two shifts, the electronic machines of the 1940s and the mainframes that followed.

War and the stored program (1940s)

Picture a roll of punched paper tape running on pulleys through a Colossus at Bletchley Park in 1944, a machine designed by the engineer Tommy Flowers. The Computer History Museum timeline records that the first Colossus became operational that year to break the Lorenz ciphers, that ten were eventually delivered with as many as 2,500 vacuum tubes each, and that Colossus cut the time to break Lorenz messages from weeks to hours. Its existence was not made public until the 1970s. In Pennsylvania, John Mauchly and J. Presper Eckert began ENIAC in 1943 at the University of Pennsylvania's Moore School. It was unveiled in 1946 and, being electronic, was over 1,000 times faster than any earlier computer, according to the same timeline.

Two changes from this decade mattered as much as speed. On December 16, 1947, John Bardeen and Walter Brattain achieved the first successful semiconductor amplifier at Bell Labs, amplifying a signal up to 100 times. And in 1948 the Manchester "Baby," developed by Frederic Williams, Tom Kilburn and Geoff Toothill, ran the first computer program on a stored-program machine. The Museum calls Cambridge's EDSAC, led by Maurice Wilkes, the first practical stored-program computer to provide a regular computing service.

Mainframes (1951-1970)

The first UNIVAC 1 went to the US Census Bureau in 1951. The Museum's timeline says it used 5,200 vacuum tubes, weighed 29,000 pounds, and was eventually sold 46 times at more than $1 million each, to customers including utilities, insurers and the military. It also records that the Lyons tea company in England built a machine, LEO, to schedule the production and delivery of cakes to its tea shops. Business computing had begun. IBM followed with the 701, which the same timeline says shipped 19 units and rented for $15,000 a month.

Smaller parts were already coming. In 1959, Jack Kilby of Texas Instruments and Robert Noyce, working separately, developed the integrated circuit. DEC's PDP-8 in 1964 was, per the Museum, the first commercially successful minicomputer, priced at $18,000, bringing a computer within reach of a lab or factory.

Then came the compatibility bet. IBM's history page says the System/360, launched on April 7, 1964, replaced all five of IBM's existing computer product lines with one strictly compatible family, and describes the project as a $5 billion gamble over four years. IBM's SPREAD task force had met in secret in 1961 to work out the plan. Customers could add capacity or switch models without rewriting their programs.

The era solved reliability and scale. It created distance: the machine was centralized and costly, so people queued behind specialists to use it.

For teams today: compatibility is a product feature. The enterprise software you buy now, including ERP, still sells the promise System/360 made, that growth won't force a rewrite.

The Microprocessor Age (1971-1984)

A CPU shrinks to a chip and lands on the desk

In 1971, Ted Hoff and Stanley Mazor conceived Intel's first integrated CPU, the 4004, and Federico Faggin, assisted by Masatoshi Shima, squeezed its roughly 2,300 transistors into a 16-pin package. The chip was developed for the Japanese calculator maker Busicom, according to the Museum's timeline. A computer had turned into a part you could design into other products.

Hobbyists moved first. In its January 1975 issue, Popular Electronics ran a cover story on the MITS Altair 8800, a kit that, per the Computer History Museum, sold for $297 (or $395 with a case) and shipped with 256 bytes of memory. Bill Gates and Paul Allen licensed their BASIC interpreter to MITS for it. Apple's Apple II followed in 1977 as a complete, ready-to-use machine, and the Museum says millions sold through 1993.

What turned a hobby into a business purchase was software. VisiCalc, a 1979 spreadsheet from Software Arts, was the "killer app" that spurred Apple II sales; many customers bought an Apple specifically to run it. The standard arrived on August 12, 1981, when IBM's Don Estridge unveiled the IBM PC at New York's Waldorf Hotel. IBM says it was priced at $1,565, had 16 kilobytes of RAM and no disk drive, and shipped with two programs, VisiCalc and EasyWriter. The Museum adds that Compaq later legally reverse-engineered IBM's BIOS software, launching a market for IBM-compatible computers that by 1996 held an 83 percent share of the PC market. Apple's Lisa in 1983 was the first commercial personal computer with a graphical interface, and the Macintosh in 1984 carried that idea to a wider audience, inspired in part by the Xerox Alto.

The era solved access. It created fragmentation, since every desk had its own machine, its own files and its own numbers.

For teams today: standards beat features. The IBM PC won less on its specifications than on being a platform others could build for, which is the same logic behind every integration marketplace.

The Network and Pocket Years (1969-2015)

Computers link up, then shrink into a phone and a rented server

This era starts earlier than the last one ends, because the network was running while the PC spread. The Computer History Museum says the ARPANET was switched on in late October 1969 and was the first large-scale, general-purpose computer network to connect different kinds of computers. The full story of what grew from it is in The History of the Internet.

Linking PCs together was the next problem, and the answer was a program anyone could use. At CERN, Tim Berners-Lee submitted proposals starting in March 1989, neither approved, and by Christmas 1990 he had prototyped "WorldWideWeb" in about three months on a NeXT computer, with a server, HTML, URLs and the first browser, according to the same timeline. In 1991 the National Science Foundation changed its policy and the Internet became a publicly accessible network with no commercial restrictions.

Then computing became a service. When AWS launched on March 14, 2006, it offered just one service, S3 storage, with EC2 servers following a few months later. Before that, Amazon notes, launching a tech startup meant buying machines and forecasting capacity. On January 9, 2007, Apple introduced the iPhone, which its announcement described as three products in one: a mobile phone, a widescreen iPod with touch controls and an Internet communications device, with US availability set for June 2007. The Museum notes it ran downloadable apps. At the top end, IBM's Roadrunner became the first supercomputer to reach 1 petaflop per second on the TOP500 list of June 2008.

The era solved reach: software could follow a person anywhere, and capacity became rentable. It created sprawl. Teams ended up with dozens of subscriptions and scattered data, and the question of who owns the record behind all those tools became harder.

For teams today: renting capacity turns a capital purchase into a monthly cost, and the mobile phone is the main computer for anyone who doesn't sit at a desk. Plan workflows for the phone first.

The Accelerator Age (1999-present)

Graphics chips train AI, then neural chips move into the PC

This era overlaps the cloud years, because its key part appeared in 1999. In August of that year NVIDIA launched the GeForce 256, which its own corporate timeline calls the industry's first graphics processing unit. A chip built for parallel graphics math turned out to be useful for much more. The CHM historian Hansen Hsu notes in IEEE Spectrum that neural networks needed vast training data from the Web and the computing power of GPUs, and that NVIDIA's CUDA programming system, released in 2007, made GPUs programmable beyond graphics. In 2012, Alex Krizhevsky and Ilya Sutskever's AlexNet put the two together. The surge in AI work after that is why AI chips and AI data centers now shape hardware plans; the software side is in the AI timeline.

The newest step moves part of that capability back into the personal computer. On December 14, 2023, Intel launched its Core Ultra mobile processors, which Intel says ushered in the first generation of the AI PC, with dedicated AI acceleration across the CPU, the GPU and a neural processing unit, or NPU. On May 20, 2024, Microsoft introduced Copilot+ PCs as a new category of Windows PCs designed for AI, built on a system architecture that brings together the CPU, GPU and an NPU capable of 40+ TOPS (trillion operations per second). Microsoft describes features that run on the device itself, such as live captions that can translate audio offline. The idea that a model can run on the machine in front of you, without a round trip to a data center, is covered in what on-device AI is.

The era solved the compute bottleneck for AI. It created a placement question: whether a given task should run in the cloud, on the device or inside an application.

For teams today: treat the NPU as a procurement and privacy question as well as a speed one. Decide which tasks can stay on the device before you decide which model to buy.

Key People

Key Facts

Key Facts: Computer Timeline

• The Antikythera Mechanism was built around the end of the second century BCE and predicted lunar and solar eclipses.
• About 60 million punched cards were used for the 1890 US census, and an operator could process about 7,000 a day.
• ENIAC was over 1,000 times faster than any earlier computer.
• IBM's System/360 cost $5 billion over four years and replaced five incompatible product lines.
• The IBM PC, unveiled on August 12, 1981, cost $1,565 with 16 kilobytes of RAM and no disk drive.
• AWS launched on March 14, 2006 with one service: S3 storage.

The Pattern Behind the History

Across the six eras, the same four moves keep repeating. We call the sequence the Shrink-and-Spread Loop. It's our synthesis of the milestones above, not a research finding.

1. A narrow tool does one job. Counters on a board, a cog adder, a card tabulator each handled one kind of work.
2. A standard turns it into a platform. Punched cards, System/360 compatibility and the IBM PC's open design let others build on top.
3. The parts shrink and the price falls. Vacuum tubes gave way to transistors, then the integrated circuit, then the microprocessor.
4. Reach widens and a new bottleneck appears. More people use the tool, and the next constraint shows up: card feeding, queues, scattered files, subscription sprawl, where AI should run.

What It Teaches Teams Today

• Ask which step you're in. A tool that's still a narrow specialist needs specialists. One that has become a platform can be handed to everyone with guardrails.
• Check compatibility before features. Every lasting platform on this timeline won on what could plug into it.
• Expect a new bottleneck. Cheaper access shifts the problem, whether to data entry, data quality or tool sprawl.
• Match the tool to where the work happens. From the abacus to the phone, adoption followed the machines that sat closest to the person doing the job.
• Decide where intelligence runs. AI PCs with NPUs reopen the old question of central versus local, and the answer will differ by task.

The computer timeline reads like a long argument for reach. Every step that mattered put calculation in the hands of someone who couldn't have used it the step before. When you weigh the next tool, ask who it brings in, and what it will need from the people already there.