A bronze disk the size of a dinner plate lies under harsh museum lights, its corroded gears frozen after two thousand years underwater. Divers found it near a Greek island, tangled with amphorae and statues. When scientists scanned it, they saw a mechanical sky coded in metal. It predicted eclipses, tracked planets, and modeled cycles of time. It is the Antikythera mechanism, and it reframes a deceptively simple question. What counts as the first computer? If the goal is to compute, the ancient world was doing it long before electricity. The abacus, a frame with beads, performed arithmetic reliably and fast in practiced hands. The Antikythera mechanism went further. It encoded astronomical theory into gear ratios so that turning a hand crank produced future positions of the Sun, Moon, and known planets. It output dials for calendars, lunar phases, and even predicted eclipses with a Saros cycle scale. It embodied a model of the heavens and let a user run that model through time. That is not just calculating. That is simulation. Yet the Antikythera mechanism lacks one quality many of us associate with computers. It cannot be reprogrammed. Its knowledge is baked into teeth counts. You cannot change the rules without rebuilding the machine. If reprogramming is essential, then the title of first computer moves forward, passing through centuries of astronomical clocks, automata, and looms. In the medieval Islamic world, scholars created ingenious astronomical devices with geared trains and mechanical calendars. In Europe, cathedral clocks animated celestial spheres and automata that displayed the phases of the Moon and the motions of planets. These devices turned theory into mechanism and time into motion. They did computation physically by exploiting the geometry of gears. But they shared the same limit. They computed one thing, in one way.

A new thread enters with weaving. In early nineteenth century France, Joseph Marie Jacquard built a loom controlled by punched cards. The cards encoded the pattern of threads. By changing the deck, a weaver changed the output without rebuilding the machine. This is a crucial idea. Instructions stored outside the machine guide its operation in general ways. The Jacquard method inspired a British mathematician named Charles Babbage. In the eighteen twenties he proposed the Difference Engine, a mechanical calculator to tabulate polynomial functions, with the goal of eliminating human error in navigational tables. It never reached full completion then, but it established a path. Babbage then imagined the Analytical Engine, a much more ambitious machine. He designed it with a mill to compute, a store to hold numbers, punched cards to encode instructions, and mechanisms to branch based on conditions. He proposed separate cards for operations and for data, a memory addressed by position, and conditional jumps. This is a template for a general purpose computer. A person could change the program and compute new things with the same hardware. Ada Lovelace, a collaborator and correspondent of Babbage, understood the implications. She described how the Analytical Engine could manipulate symbols according to rules and generate compositions beyond arithmetic. In an extended note, she sketched how to compute Bernoulli numbers using a sequence of punched cards. This is often called the first published computer program. The Analytical Engine never materialized as a functioning device in their time. Without a working machine, the world had to wait for technology to catch up with the idea of general purpose computation. While Babbage and Lovelace were wrestling with mechanics, a quieter revolution was unfolding in logic and algebra. George Boole proposed an algebra of logic with operators like and and or. Gottlob Frege, Giuseppe Peano, and later Bertrand Russell and Alfred North Whitehead developed formal systems where symbols followed precise rules. This laid a foundation for programmable reasoning. By the early twentieth century, mathematicians like David Hilbert proposed that every mathematical statement could be decided by a finite procedure. This sparked the Entscheidungsproblem, the decision problem, and led to a conceptual device called a Turing machine. Alan Turing described an abstract machine in nineteen thirty six that operated on an infinite tape with a read write head, a finite set of states, and a table of rules that specified actions based on current state and symbol. He proved that a universal Turing machine could emulate any other Turing machine if given a description of it. Church and Turing showed that many different formulations of computation were equivalent in power. From this came the notion of general purpose computing and the limits of what procedures can decide. There exists no algorithm that solves the halting problem in all cases. Computation has boundaries. The conceptual groundwork was in place. The question becomes one of engineering. Who built the first machine that embodied general purpose programmability in real time? Here, murk and secrecy complicate the answer. Several contenders deserve attention, each satisfying different criteria. Your conclusion depends on how you weigh requirements like programmability, stored programs, electronic implementation, practical execution, and reconfigurability. Consider Konrad Zuse in Berlin. Between nineteen thirty six and nineteen forty one he built a series of machines labeled Z one through Z three and beyond. The Z three, completed in nineteen forty one, used electromechanical relays and implemented binary floating point arithmetic. It read punched film strips that encoded programs and data. It could run sequences of instructions including loops. It was programmable, though not stored program in the modern sense. It survives as one of the earliest working program controlled computers. Zuse later explored the idea of a Plankalkül, a programming language, anticipating array operations and structured data. Meanwhile in the United States, George Stibitz at Bell Labs demonstrated complex number calculations on a relay based device called the Model K in nineteen thirty seven, leading to the Complex Number Calculator in nineteen forty. It accepted input via teletype from remote locations and processed operations automatically. It was a calculator rather than a general purpose machine, but it showed how electromechanical logic could be scaled. John Atanasoff and Clifford Berry at Iowa State built an electronic device called the Atanasoff Berry Computer around nineteen forty two. It used vacuum tubes and capacitors to solve systems of linear equations. It implemented binary arithmetic and electronic logic. It lacked a stored program and was specialized to one class of problems. A later court case would cite their ideas in a patent dispute. It sits at another edge of the definition. It is electronic and digital but not general purpose. In Britain, Tommy Flowers and colleagues at the Post Office Research Station designed Colossus for codebreaking during the war. Colossus ran in nineteen forty four. It used thousands of vacuum tubes to perform high speed Boolean operations on encrypted teleprinter traffic. It read punched paper tape in loops and executed fixed sequences of logical tests configured by switches and plugs. It did not store programs in memory, yet it was an electronic digital machine performing programmable logical operations. Its existence was secret for decades, which distorted early histories of computing. Then comes the Electronic Numerical Integrator and Computer, better known as ENIAC, developed at the University of Pennsylvania and unveiled in nineteen forty six. ENIAC used almost eighteen thousand vacuum tubes, consumed large amounts of power, and filled a room. It could perform thousands of additions per second. ENIAC was a general purpose machine in principle but was programmed by setting switches and wiring plugboards, a task that could take days. Later, a new control system let it execute sequences stored on function tables, moving it closer to a stored program concept. Some of ENIAC’s earliest and most important programmers were women whose contributions were underrecognized for years. They translated mathematical procedures into control flows and parallel operations on the machine’s units.

The stored program architecture, where both instructions and data reside in the same read write memory, is often treated as the turning point. John von Neumann, drawing on discussions with colleagues including J. Presper Eckert, John Mauchly, and others at the Moore School and the Institute for Advanced Study, circulated a report in early nineteen forty five describing a logical design for a stored program electronic computer. It formalized ideas about memory, a processing unit, and a control unit fetching and executing instructions. This became called the von Neumann architecture, though it grew from group contributions. Who implemented it first? The answer splits again by criteria. In June nineteen forty eight, at the University of Manchester, Tom Kilburn and Freddie Williams ran a short program on the Small Scale Experimental Machine, nicknamed the Baby. It used a cathode ray tube storage device to hold both data and instructions, and it executed a program to find the highest factor of a number. The Manchester Baby was a test bed, but it proved the stored program principle in electronic hardware. Also in nineteen forty eight, Maurice Wilkes and his team at the University of Cambridge completed the Electronic Delay Storage Automatic Calculator. EDSAC used mercury delay line memory and executed a practical instruction set. It ran the first programs for real scientific problems in May nineteen forty nine. It had a defined assembly language and subroutine library, establishing modern practices in programming and software reuse. Another claimant is the BINAC, built by Eckert and Mauchly after leaving the University of Pennsylvania. BINAC ran test programs in nineteen forty nine and influenced the UNIVAC line, which became the first commercial computer delivered in the United States. At the same time in Britain, the Manchester Mark One led to the Ferranti Mark One, often cited as the first commercially available general purpose computer in the world. These milestones mark the transition from prototypes to sustained computation in organizations. Look back at this arc and you will see multiple answers to the question. If you define a computer as any device that mechanizes a model to predict and compute, then the Antikythera mechanism stands as the first known analog computer, built in the first century before the common era. If you require programmability in a general sense, Jacquard’s loom introduces external instruction control in the early nineteenth century, and Babbage’s Analytical Engine defines a blueprint for a general purpose digital computer decades before usable hardware existed. If you require an electronic, digital, general purpose, stored program machine actually running, the Manchester Baby in nineteen forty eight is the first to execute such a program, while EDSAC is the first to provide sustained practical computing with software practices that look modern. There is another candidate that sits nearby but deserves its own category. If the criterion is a general purpose automatic calculator that operated at scale under real time industrial constraints, ENIAC’s mid nineteen forties achievements cannot be ignored. It executed complex artillery trajectory tables, weather calculations, and nuclear research computations. Though it did not begin as a stored program machine, its later modifications blurred the line. It is fair to say that ENIAC proved that fully electronic computing was operationally viable. What if your definition stresses universality at the level of logic rather than engineering details? Then Alan Turing’s universal machine from nineteen thirty six is the first computer in the definitional sense. It is not a device you can touch, but it tells you what a device must do to count as computation. It draws a boundary around the computable. From here, we can extract a more useful understanding. The character of a computer is not one thing. There are analog computers, which represent quantities by continuous physical states. There are digital computers, which represent quantities by discrete symbols. Programmability can be external, like a deck of cards or cables, or internal, like instructions stored in read write memory. Universality is the ability to simulate any other effective procedure given time and memory. Speed and scale are engineering, not essence, but they determine usefulness. With these distinctions, we can retell the story without forcing a single first. The Antikythera mechanism and sophisticated astrolabes are early analog computers. Pascal’s Pascaline and Leibniz’s stepped reckoners are early mechanical calculators. Jacquard’s loom introduces input encoding that controls a machine’s behavior without rebuilding it. Babbage’s Analytical Engine is the first design for a general purpose programmable digital computer, though unbuilt at the time. Zuse’s Z three is the first working program controlled general purpose computer using relays. Colossus is the first large scale electronic digital machine used in operation, though it is specialized. ENIAC is the first large scale electronic general purpose computer to run a range of problems, though initially not stored program. The Manchester Baby is the first to execute a stored program electronically in practice. EDSAC is the first practical stored program computer delivering sustained service and a software ecosystem. Ferranti Mark One and UNIVAC bring general purpose computing to the market. There is value in asking why the Antikythera mechanism matters in this narrative. It shows that computation did not begin with electronics. It reveals that scientific theory can be embodied in matter so thoroughly that a turn of a crank performs a calculation. It also shows that complexity in gear work was within reach in the ancient Mediterranean, then lost or not reproduced for centuries. When later generations built astronomical clocks, they reinvented an idea that had been forgotten. Knowledge can be fragile. There is also a lesson in the Jacquard loom. Stored instructions originated as solutions to practical problems in weaving. The loom’s punched cards traveled to computing through the mind of Babbage, who saw how a deck of instructions could control a machine. Later, Herman Hollerith used punched cards for data processing in the eighteen nineties United States census. Tabulators read and sorted cards with holes that represented categories. This created a data processing industry that became International Business Machines. By the time electronic computers arrived, an ecosystem of card based workflows and operators already existed. Software culture, with its schedules, formats, and meticulous checking, grew from that lineage.

A third lesson sits with the codebreakers. Colossus and its team proved that electronics could be pushed hard in real time for sustained periods. It also showed that secrecy can erase credit. For decades, histories of computing left out British advances because the machines were classified. When details became public, the story shifted. History is shaped as much by archives and declassification as by innovation. Try to establish a one line answer and nuance pushes back. The Antikythera mechanism is older than punched cards and vacuum tubes, yet it is persuasive when you hold the question to simulation. Zuse’s Z three proves program control in hardware, yet it relies on relays and external media. ENIAC demonstrates a wide range of tasks electronically, yet it begins as a wired program machine. The Manchester Baby executes a stored program first, yet EDSAC turns that principle into sustained practice with a programming culture. Each of these is a legitimate first under a different lens. The focus on firsts can obscure more than it clarifies. A better lens asks about capabilities, constraints, and transitions. When did we gain reprogrammability without rebuilding hardware? When did we gain universal computation in a machine that actually ran? When did we gain speed and reliability to tackle large scale scientific and commercial work? When did users gain programming abstractions like subroutines and assemblies? When did a community form around creating, sharing, and maintaining code? Using that lens, the story becomes a set of thresholds. Threshold one is programmable control of mechanical processes, achieved by the Jacquard loom. Threshold two is a design for a general purpose programmable digital machine, achieved by Babbage in concept. Threshold three is a working program controlled computer, achieved by Zuse with the Z three. Threshold four is high speed electronic computation at scale, achieved by Colossus and ENIAC for specialized and general tasks respectively. Threshold five is a stored program architecture realized in hardware, achieved by the Manchester Baby. Threshold six is practical, repeatable, user centered computing with software practices, achieved by EDSAC and then commercial systems like the Ferranti Mark One and UNIVAC. Each threshold answers a different sense of first. Recognizing thresholds also explains why debates persist. People choose different thresholds based on their values and the problem they are trying to solve. A historian of science may prize the embodiment of theory in the Antikythera mechanism. A computer scientist may prioritize universality and stored programs. An engineer may focus on reliability and throughput. A business historian may look for the first commercially viable product. All are coherent choices for different questions. This reframing is not wordplay. It helps when you evaluate new technologies. Ask whether a new device crosses a real threshold or only moves within one. There are many inventions that add speed but not capability. Others introduce a new capability at modest scale. When you understand the thresholds, you can place devices and ideas without forcing them into a single linear race. Let us return briefly to Ada Lovelace. Her insight about symbolic manipulation extends beyond number crunching. She imagined that when a machine operates on symbols according to rules, it could handle notation for music or graphics as easily as arithmetic, provided the symbols were well defined. That is prophetic. When we load a drawing program and manipulate shapes, we exercise that idea. When we compose music using software, we rely on it. Computation is not about numbers only. It is about rule governed transformation of representations. Look again at the Antikythera mechanism with that lens. It manipulates a representation of celestial cycles inscribed in gear ratios. It computes not numbers on a screen but positions on dials and phases engraved in metal. That is still computation. The gears do what software would later do in abstract, but they do it through physical constraints. The device is a simulation of the heavens as understood by Greek astronomers, turned into a machine that a traveler or scholar could use. We might ask whether any earlier devices qualify if we loosen the definition further. Water clocks that equalize flow and adjust gears to measure hours in varying length days encode astronomical knowledge in cam shapes. Orreries that model planetary motions with tokens on rotating arms compute positions periodically. These are analog computers in a broad sense. They accept as input the rotation of a crank or the passage of time and produce as output a configuration that stands in for the sky. They lack reprogrammability, but they are richer than pure clocks. The temptation to crown a single victor arises from tidy storytelling. Resist it. There is no singular first that satisfies every criterion without caveat. Instead, there is a family tree with several roots. Ancient simulation devices. Early programmable control. Conceptual designs for general purpose machines. Working program controlled relays. Electronic high speed special purpose logic. General purpose electronic giants. Stored program pioneers. Practical systems and commercial rollouts. Each branch matters. If you need a crisp answer for a particular purpose, choose your threshold. For the first known analog computer, point to the Antikythera mechanism. For the first design of a general purpose programmable digital computer, point to Babbage. For the first working program controlled general purpose machine, point to Zuse’s Z three. For the first electronic digital machine used in operational settings, point to Colossus. For the first large scale general purpose electronic computer, point to ENIAC. For the first electronic stored program execution, point to the Manchester Baby. For the first practical stored program computer delivering computing as a service, point to EDSAC. For the first commercial general purpose computers, point to the Ferranti Mark One and UNIVAC. There is value in memorizing those names and dates, but there is more value in memorizing the logic of the distinctions. When a debate about firsts erupts, translate the question into capabilities and thresholds. Then the answers become clear and complementary rather than competitive.

The broader lesson is about technological progress. It rarely proceeds in a straight line from one genius to a finished product. It emerges from overlapping attempts, partial successes, and conceptual leaps awaiting material support. Babbage’s designs needed precision engineering and reliable power sources that his era could not provide. Turing’s universal machine needed memory technologies like delay lines and electrostatic tubes. ENIAC needed teams of people who could reframe mathematics into networks of operations. EDSAC needed organized software practices. Commercial machines needed standard interfaces, maintenance routines, and a market ready to pay. Once you see this, you can apply it to new fields. When someone claims a first in artificial intelligence, ask whether it is a conceptual first, an operational first, a practical first, or a commercial first. When you hear about a breakthrough in quantum computing, ask whether it crosses a capability threshold or a performance threshold. Precision about criteria cuts through hype. So what was the first computer? The answer is not what you might expect. It is many firsts, each honest in its own category, and one especially old bronze device that forces us to expand our notion of what computing can be. The Antikythera mechanism stands at the beginning of a long habit of turning models into machines. The rest of the story is a steady accumulation of flexibility, speed, and generality until the machines became universal, programmable, and useful in daily work. That is the path worth remembering.