Nobody ever got fired for buying IBM.

On 16 June 1911 the financier Charles Flint merged a handful of unglamorous companies into one. They made meat scales, factory time clocks and the punch-card tabulators Herman Hollerith had built to count the 1890 United States census. Flint called the result the Computing-Tabulating-Recording Company, and for its first few years it was not obviously going anywhere.

In 1914 it hired Thomas J. Watson, a salesman recently pushed out of the National Cash Register Company. Watson brought a one-word motto, THINK, and a belief that a company could be a kind of family: dark suits, company songs, clubs for employees, and the understanding that if you were loyal, you stayed. In 1924 he gave the business a name to match his ambitions: International Business Machines.

The business model was quietly brilliant. IBM rarely sold its tabulators. It rented them out, and it sold the cards they used up, so customers paid every month for as long as they kept counting. It is the first appearance of an idea that runs through this whole story: once your records live on IBM machines, leaving becomes a project that takes years.

In 1935 the United States created Social Security and suddenly needed a wage record for almost every working adult in the country. Through the worst of the Depression, Watson had kept IBM’s factories building machines that nobody had ordered, a gamble that looked reckless at the time. In September 1936 it paid off. IBM won the contract to keep the records of 27 million workers and 3.5 million employers on 400 of its own collators, and became the machine room of the American state.

IBM came late to electronic computers and then took them over. Its first, the 701, shipped in 1952. (The famous line that Thomas Watson saw a world market for “maybe five computers” has never been traced to him. The likeliest root is his son telling shareholders in 1953 that IBM had expected orders for five 701s and came back with eighteen.) Within five years IBM had given the world the hard disk drive and FORTRAN, the first programming language most scientists ever learned.

On 7 April 1964 it announced System/360: six computers of different sizes that all ran the same software. Until then a company that outgrew its computer had to rewrite its programs for the next one. With the 360 it could grow without starting over, provided it stayed with IBM. The project cost about $5 billion over four years, at a time when IBM’s annual sales were $3.2 billion. Customers ordered more than a thousand in the first month.

Five years later IBM was on its way to the Moon. A computer it built guided the Saturn V rocket, five System/360s worked the numbers at Mission Control in Houston, and about 4,000 IBM employees worked on Apollo.

By the early 1980s IBM was so dominant that the United States government had spent thirteen years trying to prove it was a monopoly. The case began in 1969 and was dropped in 1982 as “without merit”, after more than 104,000 pages of trial transcript.

In 1985 IBM sold $50 billion of goods and services, earned $6.6 billion, employed 405,535 people and was the most valuable company in the world. One dollar in every sixteen invested in the S&P 500 was invested in IBM. Its laboratories had produced the hard disk, the floppy disk, the modern memory chip, the relational database and the barcode, and they were about to win two Nobel Prizes in two years.

IBM’s personal computer was built in a hurry. A small team in Boca Raton, Florida, was given about a year, and to move that fast it bought almost everything from outside suppliers: the processor from Intel, the operating system from a young Seattle company called Microsoft. Neither deal was exclusive.

The IBM PC went on sale on 12 August 1981 for $1,565 IBM shipped 13,533 of them before the year was out and about 750,000 in 1983. The only piece IBM had written itself was the BIOS, the small program that starts the machine. In 1982 Compaq rebuilt it from scratch without copying a line, and from 1983 anyone could make a computer that ran the same software. They were sold as “IBM compatible”. IBM kept the name on the standard; Intel and Microsoft kept the money.

The mainframe paid for everything, and at the end of the 1980s its customers began to need fewer of them. Networks of cheaper machines, built from the same parts as the PC, were taking over work that used to need one big computer. In 1991 IBM lost money for the first time in its history: $2.8 billion. In 1992 it lost $4.97 billion. In January 1993 the board cut the dividend and the chief executive, John Akers, left.

1993 was worse. In July IBM took an $8.9 billion charge to restructure and announced 35,000 more job cuts, on top of tens of thousands already gone. It lost $8.1 billion that year. The shares, which had traded at about $175 in 1987, fell to around $41.

The full-employment tradition did not survive it. Between 1985 and 1994 IBM’s workforce fell from 405,535 to 219,839.

The plan Gerstner inherited was to break IBM into a set of independent companies, which the press called the Baby Blues. He stopped it. His argument was that IBM’s rare asset was the ability to make a customer’s whole tangle of systems work together, and that only a whole IBM could sell that.

He was right, and the turnaround was real. IBM rebuilt itself around services and software. It bought Lotus in 1995, made “e-business” a household phrase, and in 2002 bought PricewaterhouseCoopers’ consulting arm with its 30,000 consultants. Revenue climbed back above $88 billion by 2000. In May 1997 an IBM computer called Deep Blue beat the world chess champion, Garry Kasparov, and it would be fourteen years before IBM led the evening news like that again.

The rescue also moved the company’s centre of gravity. More and more of IBM’s money came from running and fixing other organisations’ computers, and less from building the machines people wanted next. A services business grows with its clients’ budgets. It rarely gets to decide where computing goes.

One by one, IBM sold the businesses that made physical things for ordinary buyers. Each sale made sense on its own terms, because the margins were thin and the buyers were glad to have them. Together they meant that by 2015 there was almost nothing left that a person could buy from IBM and hold in their hands.

In February 2011 an IBM system called Watson beat the two greatest champions of the American quiz show Jeopardy! on national television. For a while IBM owned the public idea of artificial intelligence, more than a decade before ChatGPT.

What followed was a brand more than a product. In 2014 IBM put a billion dollars and about 2,000 people into a Watson business unit and pointed it at medicine. From 2015 it spent more than $4 billion buying health-data companies. At the University of Texas MD Anderson Cancer Center a Watson project cost $62.1 million and was shelved without ever being used on patients. In 2018 the health news site STAT reported internal IBM documents showing Watson for Oncology giving “unsafe and incorrect” treatment advice in testing.

In 2022 IBM sold the health business to the private equity firm Francisco Partners. It did not announce the price; its annual report later recorded $1.065 billion.

In the 1960s IBM invented the virtual machine, the trick of making one computer behave like many. Forty years later that trick became cloud computing, and the company that turned it into a business was Amazon, which opened its cloud service in 2006.

IBM answered in 2013 by buying a hosting company, SoftLayer, for about $2 billion. The same year it lost a CIA cloud contract to Amazon, challenged the award, and lost in court.

Today companies spend more than $140 billion a quarter on cloud infrastructure. Amazon, Microsoft and Google take about 63% of it. IBM is no longer in the top five.

In 2010 IBM published a promise: at least $20 of operating earnings per share by 2015. The promise was about a number per share, and there are only three ways to raise one: grow the business, cut costs, or reduce the number of shares. IBM leaned hardest on the last two.

Between 2000 and 2019 IBM spent $168.6 billion buying back its own shares. In the decade from 2010 alone it handed $134 billion back to shareholders, more than any one of Amazon, Microsoft or Alphabet spent in those years on buildings, servers and equipment.

IBM did not so much shrink as stand still while everything around it grew. In 1990 it was worth about $65 billion, more than Microsoft, Apple, Intel and Oracle put together. In 2018 Apple and then Amazon passed a trillion dollars. On 26 October 2026 Nvidia was worth about $5.2 trillion and IBM about $240 billion, the 71st most valuable company in the world.

In 2019 IBM paid $34 billion for Red Hat, the largest purchase in its history and, by most accounts, a good one. In 2020 Arvind Krishna, the engineer who had argued for the deal, became chief executive. In 2021 IBM spun off its IT infrastructure services business as Kyndryl, and about 90,000 people went with it. Since then IBM has bought HashiCorp for $6.4 billion and Confluent for about $11 billion.

What remains sold $67.5 billion of software, consulting and hardware in 2025 and employed 264,300 people. It is profitable and pays a steady dividend. More than 80% of its revenue comes from clients that buy from all three of its businesses, most of them banks, insurers, airlines, retailers and governments in more than 175 countries.

The Watson that won Jeopardy! was taken apart. Its health business became Merative. Its name came back in May 2023, in lower case, as watsonx: a set of tools for building, running and policing generative AI, and since 2026 a system called watsonx Orchestrate that manages AI agents made by IBM and by others.

IBM never reported revenue from AI. It reported a “book of business” instead, a running total of AI-related software sales and consulting signings. It passed $12.5 billion at the end of 2025, more than $10.5 billion of it consulting. Then IBM stopped publishing it.

When ChatGPT arrived at the end of 2022, the company that had once owned the idea of AI had no large language model of its own to show. IBM’s answer, in September 2023, was a family of models called Granite.

Granite does not try to be the smartest model in the world. It is meant to be cheap to run, safe to put inside a bank or a ministry, and open: since 2024 the models have been free to download under the Apache 2.0 licence. IBM backs them with things only a cautious buyer cares about: certification to the ISO/IEC 42001 standard for AI management, which IBM says no open model family had before, cryptographically signed model files, and legal cover for customers who use them through watsonx. Stanford’s Foundation Model Transparency Index gave IBM 95 out of 100 in December 2025, the highest score of any developer.