IBM Computing Legacy, The Complete Story From 1890 To Quantum Computing

The IBM computing legacy is usually told as a story about machines, and it is really a story about a company repeatedly destroying its own best business before somebody else did. IBM sold punch card tabulators when tabulators were the industry, then bet itself on mainframes, then built the personal computer that took the mainframe’s place at the centre of the story. The same pattern runs through its quantum programme. IBM put a five-qubit processor on the public internet in 2016, years before there was a market for one, and it funds that work today from the enterprise business those machines are eventually meant to succeed. What follows traces the IBM computing legacy from the 1890 United States census to the processors now installed in Kobe and San Sebastian. Before the computer, the punch card The story begins before there was a computer to speak of, with a machine built to hit a deadline. The 1880 United States census had taken more than eight years to complete and the population was still climbing, so the Census Office ran a competition for something faster. Herman Hollerith, who had worked there, won it with a system that recorded each person as holes punched into a card and counted those cards electrically. The 1890 count came in far faster than its predecessor, and Hollerith turned the machines into a business. In 1911 it was merged into the Computing-Tabulating-Recording Company, an amalgam of three firms that between them made punched-card equipment, weighing scales and time clocks, and Thomas J. Watson arrived to run it in 1914. He renamed the company International Business Machines in 1924, a name several sizes larger than the thing it then described. The punch card outlasted the tabulator that read it. It established that information could be held in a machine-readable form which survived any particular machine, and IBM was still selling cards by the billion when it began building computers to read them. That is the first pattern in the IBM computing legacy, a format outliving by decades the machinery that created it. The Hollerith tabulator and its card sorter. The dials counted; the sorter beside it dropped cards into pockets so the same deck could be counted again a different way. Credit: IBM. Dehomag and the German records Any honest account of the punch card era has to deal with what the machines were used for in Germany. IBM’s German subsidiary, Dehomag, supplied and serviced Hollerith equipment through the 1930s and into the war, and that equipment was used by the Nazi state for censuses and population registration, the administrative groundwork of identifying and tracking people. The contested question is what New York knew and controlled. Edwin Black’s 2001 book IBM and the Holocaust argued that the American parent maintained oversight of the German business and profited from its work, drawing on company correspondence and subsidiary records. IBM’s published response has been that its German subsidiary came under the control of Nazi authorities before and during the war, as happened to hundreds of foreign-owned companies operating in Germany. It adds that most of the relevant documentation was lost or destroyed, and that it disputes the picture of headquarters direction that Black draws. Its published statement sets out the company position in full. No court has ruled on the substance of the allegations. A class action filed in the United States in 2001 was withdrawn by the plaintiffs within months, and a later suit brought in Switzerland was dismissed as time-barred rather than on its merits. Historians have not settled it either. The argument is about intent and control rather than about whether the machines were used, and reviewers who accepted Black’s documentary record still divided over how much responsibility it establishes. What is not in dispute is that the technology which made IBM a company was well suited to that kind of record keeping, because a punch card system is indifferent to what it counts. That is the uncomfortable half of the argument this article opens with, that IBM’s real product was never a machine but a way of organising information. The 1950s and 1960s, becoming a computer company IBM entered the 1950s selling tabulators and left the 1960s having defined what a computer company was. The machines in this period were paid for partly by the American military and sold to businesses that had never owned a computer. The IBM computing legacy that matters commercially begins in this decade rather than in 1911. The air defence money There is a public subsidy underneath all of this, because the machines that made IBM a computer company were largely paid for by the United States Air Force. Through the 1950s IBM built the computers at the centre of SAGE, the continental air defence system, working from the Whirlwind design developed at MIT. Each AN/FSQ-7 weighed 250 tons, a figure MIT Lincoln Laboratory and MITRE both give, and MITRE puts the floor area at around 7,500 square feet per machine, so roughly half an acre for the duplexed pair that filled a blockhouse floor. IBM says SAGE work accounted for 80 per cent of its computing revenue between 1952 and 1955, a figure the company has never broken out in an audited filing. The money mattered less than what the work taught. SAGE required computers that responded to radar in real time rather than processing a batch overnight, driving displays that operators watched continuously. It pushed IBM into magnetic core memory and long-distance data transmission years before there was a commercial reason to go there. By the end of the decade that capability was sitting inside a company with a national sales force and sixty years of experience selling machines to organisations that did not want to understand them. The commercial computers of the 1960s were built on an education the Air Force had funded. The 701 IBM’s entry into electronic computing began with the IBM 701, announced in May 1952 and delivered in 1953. It was called the Defense Calculator during development, which tells you who was expected to buy it. Built from vacuum tubes, IBM rated it at more than 16,000 additions or subtractions and more than 2,000 multiplications or divisions a second, and nineteen were installed, mostly at aerospace firms and government agencies. Memory came from 72 Williams tubes, each three inches across, holding 2,048 words of 36 bits at a cycle time of 12 microseconds. It ran scientific calculations 25 to 50 times faster than the machines it replaced, which is what moved IBM from selling accounting equipment to selling computation. The 650, workhorse of modern industry The 701 sold to nineteen customers who could afford it, and the 650 was built for everybody else. Announced in July 1953 and first shipped in December 1954, it became known as the workhorse of modern industry. IBM projected it would sell fifty units and built nearly two thousand by 1962, which is why the company describes it as the first mass-produced computer. It held 20,000 digits on a magnetic drum and could be programmed in SOAP, the Symbolic Optimal Assembly Program, rather than in raw machine code. IBM also placed around a hundred of them in universities on the condition that computing courses were taught, which is how a generation of computer scientists, Donald Knuth among them, first met a computer. RAMAC and FORTRAN Two things IBM built in the same decade as the 701 outlasted every machine in this article. In 1956 the company introduced RAMAC, which IBM’s own archive calls the first computer to use a random-access disk drive, storing five million characters across fifty spinning disks. Before it, data was read in the order it had been written, off tape or cards, and RAMAC meant a program could reach any record directly without winding through everything in front of it. The 305 RAMAC production line in San Jose. Each machine carried fifty spinning disks, and this is the ancestor of every hard drive that followed. Credit: IBM. The second was a language. John Backus and his group delivered FORTRAN in 1957 after arguing for years that programmers should not have to write in the machine’s own instruction set. The compiler had to produce code fast enough that experienced programmers would not simply ignore it, and it did, which is why scientific computing still carries FORTRAN in its foundations nearly seventy years later. Neither was a machine, and that is the point. IBM’s habit of building the thing underneath the product, the storage format or the language or the interface, is what let it keep selling to customers whose hardware it had made obsolete. The 1401 reaches the back office The machine that actually put IBM into ordinary offices was not the 701 or the System/360. The 1401, announced in 1959, was a transistorised computer aimed at businesses running punch card accounting, and it could be leased for a few thousand dollars a month rather than bought outright. More than ten thousand were installed, a figure no computer had approached before, and for a great many companies the 1401 was the first computer they owned. It also did something subtler, reading and punching the same cards the customer already used, so a business could move onto a computer without reorganising everything on the first day. System/360, the bet that remade the industry On 7 April 1964, IBM announced the System/360, a family of compatible computers that reset the industry. Development cost $5 billion across new processors, new peripherals and a new operating system, which is why the programme is remembered as the point where IBM staked the company on a single product line. Nothing else in the IBM computing legacy cost as much or repaid the outlay as completely. A System/360 Model 65. The point of the family was that this machine and the entry-level Model 30 ran the same software, which had not been true of any computer line before it. Credit: IBM. Compatibility was the central idea, because every model in the family ran the same software. Six processors were announced that day, from the Model 30 upward, and IBM put the span between smallest and largest at roughly fiftyfold; the Model 75 followed in April 1965. The line also fixed the 8-bit byte as the standard unit of computer memory. Its modular architecture meant a customer could move up that range without rewriting anything. That is what made the family worth announcing as a family rather than as six machines. IBM did not publish a price per model. It quoted the range instead, from $2,700 a month for a basic configuration to $115,000 for a large multisystem one, which tells you how wide a span one architecture was being asked to cover. The family also created an industry made of other people’s products. Firms appeared selling plug-compatible disks, tape drives and memory that ran on IBM’s architecture without IBM’s permission, which is the first recognisable platform business in computing. System/360 installations ran everything from airline reservations to the Apollo programme, and IBM’s share of the market it had defined was large enough to attract the antitrust suit that followed. OS/360 runs late System/360’s hardware arrived roughly on schedule. Its operating system did not. OS/360 ran late and over budget on a scale nobody had seen before, and it became the example that taught the industry software does not scale the way manufacturing does. The IBM computing legacy includes its failures, and this is the one that produced a book still set for students fifty years later. Fred Brooks, who managed the project, wrote up the lesson in 1975 as The Mythical Man-Month. Its central observation, that adding people to a late software project makes it later, is still quoted because it is still true, and the book effectively founded software project management as a discipline. IBM produced it by failing publicly and at enormous expense. Unbundling, 1969 On 23 June 1969 IBM began pricing software and services separately from the hardware they ran on. Until then a computer arrived with its programs and its engineers included, which meant nobody else could sell software to an IBM customer, because the software was already free. The operating system itself stayed free; what acquired a price was the layer above it, the compilers, utilities and applications. IBM had announced the review that produced it on 6 December 1968, six weeks before the United States Department of Justice filed its antitrust suit in January 1969. The case then ran for thirteen years before it was dropped in 1982 without a judgment. The independent software industry grew in the space unbundling opened, and the largest companies in computing today sell precisely what IBM stopped giving away. The 1970s and 1980s, computing leaves the machine room The following two decades moved computing out of the raised-floor room and onto desks. Most of what IBM invented in this period ended up making money for other companies. Much of the IBM computing legacy in this period is a record of inventing things and then watching other companies sell them. System/370 moves to silicon The System/370 arrived in 1970 and kept every System/360 program running while changing what the machine was made of. The Model 145 replaced magnetic core memory with silicon, the first IBM mainframe to do so, which is the point at which memory stopped being wound by hand and started being printed. The System/370 family ran through the decade and into the 1980s, each generation faster than the last. Codd’s relational model In 1970 Edgar Codd, a mathematician at IBM’s San Jose laboratory, published a paper proposing that data be organised as relations rather than as the hierarchies then in use. The idea was that a user should be able to ask for what they wanted without knowing how the records were physically stored, which sounds unremarkable now because Codd won the argument. IBM built System R to test it, and out of that project came a query language called SEQUEL, written by Donald Chamberlin and Raymond Boyce, later renamed SQL. IBM was slow to commercialise the work and Oracle reached the market first with a relational product. SQL remains the language almost every business system in the world is queried in. The floppy disk escapes its purpose IBM’s San Jose laboratory, which had produced RAMAC, introduced the eight-inch diskette in 1971. It was built for an unglamorous reason, to load microcode into System/370 machines, and it was read-only at first. The format escaped its purpose almost immediately. Removable, cheap and standardised, the floppy became the way software was distributed for more than two decades, including the software for the personal computers that took IBM’s own market. The company invented the medium its competitors shipped their products on, which is IBM’s whole history in a single object. John Cocke argues for doing less The work began in 1974 on a cancelled telephone-switching controller, and by 1980 John Cocke’s team at Yorktown had a prototype machine called the 801. The argument behind it was that processors had grown baroque, and that a smaller instruction set executed quickly would beat a large one executed slowly. The idea became known as reduced instruction set computing, and Cocke took the Turing Award in 1987, its citation naming compilers, large-system architecture and RISC together. IBM was characteristically slow to build a business on it, shipping the ROMP processor in a RISC workstation only in 1986, by which time others were arriving at the same idea. The POWER architecture in the RS/6000 is the direct descendant of the 801, and the same reasoning shaped processors from other companies for the following thirty years, including the designs in most phones today. Dennard’s memory cell Robert Dennard, at IBM’s Yorktown laboratory, worked out in 1966 that a single transistor and a capacitor could hold one bit, and patented it in 1968. Dynamic random access memory replaced designs needing six transistors per bit, and the density that followed is why memory became something bought by the gigabyte rather than the kilobyte. Dennard is remembered a second time for the 1974 paper describing how transistors behave as they shrink, the scaling rule that governed chip design for three decades until power density broke it in the mid 2000s. Both results came out of a research division permitted to work on problems with no product attached, and it is the same arrangement funding the quantum programme now. The PC IBM gave away IBM announced the Personal Computer on 12 August 1981, a year after a small team in Boca Raton, Florida, started work on it. It is the most famous machine the company ever built and the one it kept least of. The machine ran an Intel 8088 processor at 4.77 MHz with 16KB of memory, expandable to 256KB, and it shipped with IBM Personal Computer DOS, adapted with Microsoft, alongside CP/M-86 and the UCSD p-System as options. Its open architecture, more than any single component, is what let a clone industry grow up around it. It started at $1,565, about $5,700 in 2026 money on the US consumer price index. That price put a machine carrying IBM’s name on desks that had never held a computer before. What IBM did not own was the processor or the operating system, and that omission decided the next twenty years. Microsoft licensed MS-DOS to everyone, Intel sold chips to everyone, and the clone makers took the market IBM had opened, leaving IBM with the x86 platform it had created and a shrinking share of it. The success was so overwhelming that Time named the personal computer, as a category rather than any one model, its Machine of the Year for 1982. The IBM Personal Computer, 1981. IBM owned neither the processor nor the operating system, and the clone makers took most of the market it opened. Credit: IBM. Zurich and the Nobel Prizes IBM’s laboratory outside Zurich produced two Nobel Prizes in physics in consecutive years. Gerd Binnig and Heinrich Rohrer took half the 1986 prize for the scanning tunnelling microscope, an instrument that let researchers see and eventually move individual atoms, the other half going to Ernst Ruska for the electron microscope. The following year J. Georg Bednorz and K. Alex Müller took the whole of the 1987 prize for finding superconductivity in a copper oxide ceramic, twelve degrees above a record that had stood since 1973, in a class of material nobody expected to superconduct at all. Both awards came out of the same corporate research operation that now builds superconducting qubits, though those qubits rely on conventional low-temperature superconductors rather than on the ceramics Bednorz and Müller found. Two physics Nobels won inside its own laboratories sit in the IBM computing legacy alongside the products. The 1990s, near collapse and reinvention The decade opened with IBM’s business model coming apart, and the company posted the worst results in its history. What it did in response set the shape it still has. The IBM computing legacy came closer to ending in 1993 than at any point before or since. The AS/400 The AS/400, launched in 1988, reshaped mid-range computing by integrating the database into the operating system. A company buying one did not buy a database separately, or hire anyone to install it, because OS/400 already contained one. A layer called the Technology Independent Machine Interface sat between the software and the hardware, which is why programs compiled in the 1980s still run on IBM i machines built decades later. RS/6000 and POWER IBM entered the UNIX workstation market in February 1990 with the RS/6000 and the POWER architecture. The machines undercut the established UNIX workstation vendors on price for the performance delivered, which is how IBM took a share of the scientific and engineering market. POWER outlived the workstation it launched in. Variants of the architecture ran the Summit supercomputer and, in PowerPC form, the Xbox 360, the PlayStation 3 and the Wii. The year IBM nearly went under In 1993 IBM reported a net loss of $8.1 billion, one of the largest annual losses any American company had posted, exceeded only by General Motors’ $23.5 billion the year before. The mainframe business it depended on was being eaten by cheaper distributed systems, and the strategy in place was to break the company into increasingly autonomous business units, an emerging federation of companies. Louis Gerstner arrived as chief executive that April from a food and tobacco conglomerate, with no computing background, and decided not to split it up. His argument was that the one thing customers could not get anywhere else was a single supplier who would make all of it work together, which turned IBM into a services company that also made hardware. System/390 hits 1,000 MIPS The mainframe was supposed to be finished by the late 1990s, and in 1998 IBM shipped its fastest one yet. The 1998 S/390 G5 Parallel Enterprise Server 10-way Turbo model reached 1,000 MIPS, a figure IBM presented at the time as the most powerful mainframe in the world. IBM presented that as a record, and figures of this kind came from the company rather than from an independent count. Deep Blue and Watson In May 1997 Deep Blue beat Garry Kasparov over six games, the first time a reigning world chess champion had lost a match to a computer under tournament conditions. The machine was a special-purpose chess engine rather than anything general, and IBM retired it immediately afterwards, its two racks going to the Smithsonian and the Computer History Museum. It had done what it was built to do, which was to put the company back in the public mind. Watson repeated the trick in February 2011 by winning at Jeopardy against Ken Jennings and Brad Rutter, this time using natural language rather than brute-force search. The healthcare business IBM tried to build on Watson went considerably worse than the demonstration, and in January 2022 IBM sold the Watson Health data and analytics assets to Francisco Partners without disclosing terms. That gap between a spectacular result and a sellable product is worth remembering when reading any quantum roadmap, IBM’s included. Public demonstrations run right through the IBM computing legacy, from Deep Blue to the quantum roadmap, and each one was aimed at customers rather than at chess or quiz audiences. The 2000s and 2010s, selling the hardware IBM spent these two decades deciding what it no longer wanted to make. The pattern of the exits explains what the quantum programme is being built inside. For these two decades the IBM computing legacy is defined by what the company stopped selling rather than by what it built. z/Architecture and Power Systems In 2000, IBM introduced the z/Architecture, extending mainframe computing to 64-bit addressing. The z-series that followed kept the mission-critical workloads, with the z13, announced in 2015, capable of processing 2.5 billion transactions per day. Announcing the z16 in April 2022, IBM said the platform was valued by 45 of the world’s top 50 banks, which is a claim about who rates it rather than who runs on it. IBM has not restated the figure in its z17 material, it names no ranking source for the top 50, and numbers of this kind come from the company rather than from an independent count. In 2008, IBM unified its server lines into Power Systems, combining the AS/400 and RS/6000 heritage into a single platform. Running the same POWER chips, they support IBM i, AIX and Linux, which let one platform carry workloads that had needed three. Why the mainframe survived The mainframe was declared obsolete in the early 1990s and it is still shipping. IBM released the z16 in 2022 and the z17 in 2025, and the business remains one of the more profitable parts of the company. The institutions running payments, insurance and airline reservations on those machines have found migrating away more expensive than staying. That durability is the quiet argument underneath the quantum programme. IBM does not attribute research funding to any one business line, but the enterprise business as a whole is what carries the cost. IBM’s pitch to a bank is not that a quantum computer will replace what it has, but that IBM will still be there to run both. That is precisely the argument Gerstner made in 1993, and the one Watson made about tabulators before him. Selling the PC business IBM announced in December 2004 that it was selling its personal computer division to Lenovo, closing the business it had opened in 1981. In October 2014 it agreed to pay GlobalFoundries $1.5 billion to take its semiconductor manufacturing operations, a deal that closed the following July. Paying somebody to accept a business is the reverse of a normal sale, and IBM had decided that owning fabrication no longer paid for itself. What it kept was research, software, services and the mainframe. Reading those exits alongside the quantum programme explains the shape of the company today. The quantum machines are not sold as boxes so much as rented as capacity, which is the model IBM has been moving toward for twenty years. From Gerstner to Krishna The company that emerged from 1993 sold services more than machines. Sam Palmisano succeeded Gerstner in 2002 and pushed further into software and consulting, selling the PC division on the way. Ginni Rometty followed in 2012 and made the largest bet of the modern era. IBM acquired Red Hat for $34 billion in a deal that closed in 2019, giving it the open-source layer that enterprise cloud runs on. Arvind Krishna, a principal architect of the Red Hat acquisition, became chief executive in April 2020, and the following year made the company smaller. In November 2021 IBM spun off its managed infrastructure business as Kyndryl, which took about 90,000 staff and roughly $19 billion of annual revenue with it. That was not a disposal of a failing unit so much as a decision about what IBM wanted to be. What is left is software, consulting, the mainframe and research. IBM reported revenue of $67.5 billion for 2025 and 264,300 employees at the end of it, against roughly 346,000 five years earlier. IBM does not disclose Red Hat’s revenue separately. Hybrid Cloud, the revenue category inside the Software segment that IBM identifies with Red Hat, reported $7,327 million in 2025, against $3.36 billion for Red Hat alone in its last full independent year to February 2019. The two figures do not cover the same set of products, so the comparison shows the change in what is being counted as much as the change in the business. When this article says the quantum programme is paid for by the enterprise business, that is the business it means. The quantum programme, from thermodynamics to 2029 IBM runs its quantum programme the way it once ran the mainframe business, as a published roadmap that customers can plan against. It has shipped a named processor almost every year since 2019 and offers more than a dozen quantum computers over the cloud. The quantum programme is the part of the IBM computing legacy that is still unfinished, and the only part whose value is not yet settled. Landauer, Bennett and the prehistory IBM’s quantum work did not start with a qubit. In 1961 Rolf Landauer, at Yorktown, showed that erasing a bit of information carries an unavoidable thermodynamic cost, establishing that information is a physical quantity rather than an abstraction. In 1973 his colleague Charles Bennett showed that computation itself need dissipate arbitrarily little energy if it is made reversible. That line runs directly into the modern field. Bennett and Gilles Brassard published the BB84 protocol in 1984, giving quantum cryptography its foundation, and in 1993 Bennett was one of six authors on the paper describing quantum teleportation. IBM was working on quantum information theory for decades before any of it became a product. It ran Shor’s algorithm on a seven-qubit NMR machine at Almaden in 2001, factoring 15. The 2016 decision to put a superconducting processor on the public internet is better understood as that research finally acquiring an audience. Putting the machine on the internet In May 2016 IBM put the IBM Quantum Experience online, placing a five-qubit processor on the public internet. Anyone with a browser could queue a circuit on real hardware, which had not been possible before. IBM says it was the first to open such a service. Rigetti put its Forest platform into public beta in June 2017 and D-Wave launched Leap in 2018, which its own filings call the first real-time publicly accessible quantum cloud service. Both statements can stand, because Leap opened access to annealing hardware rather than to a gate-model processor. What is not in dispute is that a generation of researchers learned the machine on IBM’s quantum cloud. In March 2017 IBM released Qiskit, the open-source framework that became the field’s most widely used quantum development kit. Version 1.0 arrived in 2024 built for circuits of a hundred qubits and more, and the software has mattered as much as the hardware in keeping researchers on IBM machines. The processors, year by year IBM publishes all of this in advance on its quantum roadmap, and the processor line since 2019 has run in two directions, first chasing qubit count and then chasing quality. That change of direction is the most important thing in the table below, because it is the point at which IBM stopped competing on the number everyone else was quoting. The 433-qubit Osprey processor in exploded view, with the qubit layer on top and the readout and wiring layers beneath. IBM introduced this multi-level wiring on Eagle in 2021 and carried it to more than three times the qubit count here. Credit: Connie Zhou for IBM. YearProcessorQubitsAvailabilityWhat it was for2016IBM Q Experience5Free cloudFirst quantum computer on the public internet2019Falcon27CloudSet IBM’s successive Quantum Volume records2020Hummingbird65CloudScaled the readout electronics2021Eagle127CloudFirst IBM processor past 100 qubits, multi-level wiring2022Osprey433CloudLargest count of the scale-first era2022Egret33CloudIntroduced the tunable couplers that define the Heron generation2023Condor1,121Never offered to usersScale and yield milestone, 50% higher qubit density2023Heron r1133CloudFewer qubits, far lower error rates2024Heron r2156CloudThe workhorse behind the System Two deployments2025Loonnot disclosedResearch test chipProving the wiring qLDPC error correction needs2025Nighthawk120Cloud, from Jan 2026218 tunable couplers, aimed at verified quantum advantage2026Kookaburranot disclosedplannedFirst module to hold data in a qLDPC memory2029Starlingnot disclosedplanned200 logical qubits running 100 million operationsDelivered hardware to 2025, and IBM’s stated plans beyond it. Kookaburra and Starling are targets on IBM’s published quantum roadmap, not machines that exist. The utility experiment The most closely examined result IBM has produced in quantum computing is a 2023 paper in Nature, titled Evidence for the utility of quantum computing before fault tolerance. Youngseok Kim was first author and Abhinav Kandala the senior author. Running on a 127-qubit Eagle processor, the team measured expectation values for circuits large enough that brute-force classical simulation was out of reach. It used error mitigation rather than error correction to get usable answers out of a noisy machine. What happened next is the part that matters. Within two weeks, classical simulation groups came for the result. Tomislav Begušić and Garnet Kin-Lic Chan at Caltech reproduced the quantum answers using a method called sparse Pauli dynamics, running on a single core of a laptop, faster than the quantum hardware. A tensor-network approach from Joseph Tindall and colleagues, published in PRX Quantum in 2024, did the same to greater accuracy than the quantum processor had achieved. IBM’s central result was not shown to be wrong. It was shown to be unnecessary, which for a claim about utility is nearly as damaging. A follow-up from the same Caltech group went further, identifying inaccuracies in the zero-noise extrapolations IBM had used to clean up its data. The episode is the clearest illustration available of why the July 2026 advantage announcement should be read with patience. The pattern in this field is that a quantum claim stands until somebody writes a better classical algorithm, and somebody usually does. System Two IBM unveiled Quantum System One in 2019, the first integrated quantum computing system it sold as a product rather than a laboratory rig. IBM Quantum System Two, operational since 2023, is the modular cabinet those later processors sit in, holding several chips with high-speed links between them. It has since left IBM’s own sites. RIKEN took delivery of the first System Two outside the United States in June 2025, installed alongside the Fugaku supercomputer, and Europe’s first went to the IBM-Euskadi centre in San Sebastian. IBM Quantum System Two. The cabinet is modular by design, holding more than one processor with links between them, which is how IBM intends to grow past the limits of a single chip. Credit: IBM. The roadmap separates two goals that are routinely confused. Quantum advantage, meaning a verified result classical hardware cannot match, is what Nighthawk is aimed at, and IBM has said it expects to reach that by the end of 2026. Fault tolerance is the 2029 target, when the Starling system is meant to deliver 200 logical qubits capable of 100 million operations. Both are company projections rather than achievements, and IBM has moved its own dates before. IBM told investors in February 2025 that its quantum business was approaching $1 billion in signings since the first quarter of 2017. Signings are booked contract value rather than recognised revenue, and IBM does not report quantum revenue separately anywhere, so no audited figure for it exists.

Its Quantum Network passed 250 member enterprises, universities and research institutions, which is the number IBM itself publishes rather than the 300 Fortune 500 companies sometimes attributed to it. What a logical qubit is The processors in that table are superconducting circuits, patterned in aluminium and niobium on a silicon wafer and cooled to roughly fifteen thousandths of a degree above absolute zero. Each qubit uses a Josephson junction to make the circuit’s energy levels uneven, so that two of them can be addressed as a qubit and the rest left alone. The qubit counts in the table above are physical qubits, and physical qubits are unreliable. Every operation on one carries an error rate and errors accumulate through a circuit, so a circuit has a depth budget before the output becomes noise. IBM puts that budget at around 5,000 two-qubit gates on Heron, and only with error mitigation applied on top. This is why a 1,121-qubit processor was not a thousand times more useful than a one-qubit one, and why IBM stopped competing on the number. A logical qubit is many physical qubits operated together so that errors can be detected and corrected faster than they appear. The cost is brutal, because the correcting code needs a large number of physical qubits for each logical one. That is why IBM’s Starling target of 200 logical qubits in 2029 is a far harder problem than the 1,121 physical qubits it already built in 2023. This is what makes the July 2026 figures readable. Seventy logical qubits running 2,415 two-qubit operations is a claim about depth as much as width. The interesting question is not how many qubits took part, but how long the machine stayed coherent enough to keep computing. The qLDPC codes that Loon and Kookaburra exist to test are IBM’s bet on making that ratio cheaper. What still has to be proved Everything before this point is history and can be checked. The quantum section is a roadmap, and roadmaps are promises. IBM has been unusually specific about its own, which makes it unusually easy to hold to account. The two dates that matter are a verified advantage result by the end of 2026 and a fault-tolerant Starling system in 2029. IBM said on 30 July 2026 that it had met its own advantage criteria in work with the University of Chicago, a claim examined at the end of this piece. Starling remains four years out. The advantage claim in particular has to survive the response IBM itself made when Google claimed quantum supremacy in 2019, which was that a better classical algorithm had closed the gap. Any result IBM publishes will face the same test from Google, and that adversarial checking is the healthiest feature of the field. Krishna puts a date on it IBM’s chief executive has been unusually willing to attach numbers and years to all of this, which makes the company easy to hold to account later. Speaking to CNBC on 30 July 2026, the day of the advantage announcement, Arvind Krishna said quantum would have a measurable impact on IBM’s top and bottom line “in 2028 or 2029”. By the end of the 2030s, he said, the company was “pretty convinced this is a trillion dollars of value”. He had already told the Think conference earlier in the year that IBM believed advantage would be reached during 2026. Two distinctions are worth keeping straight, because the headlines rarely do. The trillion-dollar figure is a claim about the value of the market, not about IBM revenue, and the 2028 date is about revenue becoming measurable rather than material. Krishna has also compared IBM’s position in quantum to Nvidia’s early lead in GPUs, which is a comparison that flatters the speaker and assumes the demand arrives. Set against the rest of this article, the claims are recognisable. IBM said something similar about Watson, and the healthcare business built on it did not arrive. The difference this time is that the roadmap has been public for years and the dates are specific enough to be checked, which is more than most of the industry offers. The company’s record suggests it will get some of this wrong and ship something anyway. It missed the market it created twice, with the relational database and the personal computer, and it built two of the most-watched demonstrations in computing while failing to turn the second into a business. What it has not done, in a hundred and thirty years, is stop building the next machine. The research division A pattern runs through everything above. Landauer, Bennett, Codd, Dennard, Cocke, Backus, Binnig, Rohrer, Bednorz and Müller were all employed by the same research division, and much of what they produced was given away, lost to competitors, or commercialised by somebody else. On the IFI CLAIMS ranking, which is a commercial count rather than an official one, IBM led United States patent grants for 29 consecutive years from 1993 to 2021 before Samsung overtook it in 2022. IBM said the drop reflected a deliberate shift, since 2020, away from chasing the ranking. That is a strange record to hold for a company that so often failed to convert its own inventions, and it is the best evidence for how this organisation actually works. It has treated fundamental research as a fixed cost of being IBM rather than as a pipeline that must pay for itself each quarter. That run of patents is the most quantified thing in the IBM computing legacy, and also the least informative about what the company can now build. The pattern, and what it predicts The distance from the 701 to a Heron is easier to state than to picture, and the more useful comparison is with the competition. Google has pursued error correction on smaller, higher-quality chips. D-Wave took a different route again and built quantum annealers, a distinct architecture aimed at optimisation problems rather than at the circuit model, and it has had machines in commercial use for longer than any gate-model vendor. What distinguishes IBM here is the bet that error correction and a published schedule matter more than raw qubit count. One habit recurs across the whole of the IBM computing legacy, and it explains the quantum bet better than the roadmap does. Today IBM’s quantum machines are used mostly for chemistry and materials problems, and for benchmarking work that tests where an advantage might appear. No commercially useful result has yet been produced that a classical computer could not match, which is the honest position across the industry and not only at IBM. Its standing in the field rests on the roadmap being met, not on results already banked. The IBM computing legacy is consistent enough to state plainly. The company has repeatedly built the product that undercut its own best business. It could afford to be early because a research division was carried by whatever happened to be profitable at the time, which is IBM’s own stated model rather than an outside assessment. That dependency was tested in the open on 14 July 2026. IBM preannounced a second quarter it had been due to report a week later, the shares fell just over 25 per cent in a single session and closed near $217, and roughly a quarter of a $272 billion company went before lunch. It was the worst day in at least fifty-eight years of the company’s trading records, worse than Black Monday in 1987, and unlike Black Monday there was no market-wide collapse to be swept up in. Quantum had nothing to do with it. Adjusted earnings came in at $2.93 a share against a $3.02 consensus, on revenue of $17.2 billion against an expected $17.86 billion, and Arvind Krishna told investors that enterprise clients had swung their capital spending away from software and mainframes towards servers, storage and memory in the closing weeks of June, ahead of expected price rises. His letter carried the sentence “this quarter we faltered”. Five months earlier the stock had dropped 13 per cent in a day after Anthropic demonstrated a tool for modernising COBOL, the language a good deal of IBM’s installed base still runs on. The point is not that the quantum programme caused any of this, because against $17.2 billion of quarterly revenue it does not register. The point is that the business paying for the programme turned out to be the exposed one, while the programme itself shipped Nighthawk R2 on schedule the day after the crash. Nothing in the IBM computing legacy has ever been funded any other way, which is why those figures are worth reading beside the roadmap rather than separately from it, and we set out the full argument in our feature on IBM’s record single-day plunge. The quantum programme is that same wager placed again. The company that told its salesmen to THINK now sells time on machines whose commercial value is still unproven, and pays for them out of an enterprise business those machines are meant to succeed. Whether that wager comes off is the open question hanging over the IBM computing legacy, and unlike the mainframe bet it will be settled in public against a published date. The line from the 1890 census to Nighthawk is a century and a third of the same behaviour, and it is documented in the IBM computing legacy archive the company keeps of its own history. On 30 July 2026 IBM said work with the University of Chicago had met its own criteria for quantum advantage, reporting 70 logical qubits and 2,415 logical two-qubit operations in about fifteen minutes. Jay Gambetta, Director of IBM Research and an IBM Fellow, said the field was “now firmly in the quantum advantage era”. That claim is IBM’s, the classical comparison is exactly the ground IBM itself contested when Google claimed supremacy in 2019, and the fault-tolerant Starling system on its quantum roadmap remains four years out. On this company’s record the useful question is not whether the machine works, but who other than IBM is in a position to confirm that it does. Legal disclaimer Quantum Zeitgeist does not provide personal investment or financial advice, and does not act as a personal financial, legal, or institutional investment adviser. We do not individually advocate the purchase or sale of any security or investment, or the use of any particular financial strategy, and all investment strategies carry the risk of loss for some or even all of your capital. Before pursuing any financial strategy discussed here, or relying on information within this website, you should always consult a licensed financial adviser. Any analysis we provide is for informational purposes only, does not take your circumstances into account, and should not be treated as an individualised recommendation, since the securities mentioned may not be suitable for all investors. Questions and answers What is the IBM computing legacy in one sentence?A century and a third of setting the standard and then losing the market for it. The IBM computing legacy runs from the punch card through System/360, the relational database, the disk drive and the PC to the quantum roadmap. The habit that recurs is that IBM defines the format everyone adopts while somebody else sells the volume. What did IBM make before it made computers?Punch card tabulators. The company was assembled in 1911 as the Computing-Tabulating-Recording Company around Herman Hollerith’s machines, which had counted the 1890 United States census electrically, and it was renamed International Business Machines in 1924. It was still selling cards by the billion when it started building computers to read them. Why does System/360 matter so much?Because it made software portable across a family of machines. Announced in April 1964 after a $5 billion development programme, every model ran the same software, so a customer could move up the range without rewriting anything. It also fixed the 8-bit byte as the standard unit of memory. Did IBM invent the relational database?Yes, and then lost the market. Edgar Codd published the relational model at IBM’s San Jose laboratory in 1970, IBM built System R to test it, and the query language that came out of that project became SQL. IBM was slow to commercialise it and Oracle reached customers first with a relational product. What happened to the IBM PC?IBM opened the market and then lost it, because it owned neither the processor nor the operating system. Intel sold chips to everyone and Microsoft licensed MS-DOS to everyone, so clone makers took the business. IBM sold the division to Lenovo in a deal announced in December 2004. Has IBM actually achieved quantum advantage?IBM says so, and the claim is very new. On 30 July 2026 the company said work with the University of Chicago had met its criteria, reporting 70 logical qubits and 2,415 logical two-qubit operations. IBM made two such claims the same day, the second with the software firm Algorithmiq on a materials simulation run on a Heron processor. Both are IBM’s own claims, and the history of this field is that quantum results stand until somebody writes a better classical algorithm. What is a logical qubit?Many physical qubits operated together so errors can be corrected faster than they appear. Physical qubits are unreliable and errors accumulate through a circuit, so raw qubit counts are misleading. Error correction is expensive in physical qubits, which is why IBM’s target of 200 logical qubits in 2029 is harder than the 1,121 physical qubits it built in 2023. When does IBM expect a fault-tolerant quantum computer?2029, with a system called Starling intended to deliver 200 logical qubits capable of 100 million operations. That is a company projection rather than an achievement, and IBM has moved its own dates before. Arvind Krishna has separately said quantum should have a measurable impact on IBM’s revenue in 2028 or 2029. Does IBM still make mainframes?Yes, and they remain profitable. The z16 arrived in 2022 and the z17 in 2025, and the institutions running payments, insurance and reservations on them have found migration more expensive than staying. IBM does not break out research funding by business line. Its Infrastructure segment, which contains IBM Z, ran a 22 per cent segment profit margin in 2025, behind Software on 33 per cent. The machines, in order The tables below set out the IBM computing legacy in order of announcement, first the classical machines and then the quantum processors. Dates are the year each system was announced rather than first shipped, which in several cases was a year or more later. Classical machines, 1952 to now SystemYearsClassWhat it introducedIBM 7011952-1955Scientific MainframeFirst IBM commercial scientific computer, 72 Williams tubes, vacuum tubesIBM 6501954-1962Business ComputerFirst mass-produced computer, ~2000 units sold, magnetic drumIBM 7041954-1960Scientific MainframeFirst with FORTRAN, magnetic core memory, floating pointIBM 14011959-1971Business ComputerMost popular computer of early 1960s, transistorisedSystem/360 Model 301964-1970sBusiness MainframeEntry-level S/360, 8-64KB memory, microcodeSystem/360 Model 401964-1977General PurposeMid-range S/360, 16-256KB memorySystem/360 Model 501964-1970sGeneral PurposeHigher performance, 64-512KB memorySystem/360 Model 651965-1970sEnterprise MainframeLarge system, 128KB-1MB memorySystem/360 Model 751965-1970sHigh-End MainframeTop S/360, hardwired logic, no microcodeSystem/370 Model 1451970-1977Business MainframeFirst with silicon memorySystem/370 Model 1551970-1977Enterprise MainframeHigh-performance, virtual memory supportSystem/370 Model 1681972-1980High-End MainframeTop-end S/370, multiprocessor capableIBM PC 51501981-1987Personal ComputerOpen architecture, Intel 8088, MS-DOSAS/4001988-2000Midrange BusinessIntegrated DB2 database, object-basedRS/60001990-2000UNIX WorkstationPOWER architecture, AIX OSSystem/390 90211990-1994Enterprise MainframeWater-cooled, bipolar technologyS/390 G51998-2000Enterprise MainframeFirst to reach 1000 MIPSz9002000-2003z/ArchitectureFirst 64-bit z/Architecturez10 EC2008-2010z/Architecture4.4 GHz quad-core processorz132015-2019z/Architecture5 GHz, 8-core processor, 10TB memoryz142017-2021z/Architecture5.2 GHz, 10-core, pervasive encryptionz152019-2023z/Architecture5.2 GHz, 12-core, data privacy featuresz162022-presentz/Architecture7nm process, AI acceleration, Telum processorEvery IBM machine named in this article, in order of introduction. Performance figures have been left out because IBM did not publish comparable ones across seventy years of architectures. Stay currentSee today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals. Tags: