The story of personal computers is quietly repeating itself — and almost nobody is noticing.
There is a peculiar comfort in déjà vu. When something radically new arrives, it tends to feel unprecedented — a rupture with everything that came before. Right now, artificial intelligence carries that feeling. Every week, another headline announces that AI will change everything. Some of those headlines are breathless. Some are frightened. Nearly all of them treat AI as something the world has never encountered.
But it has. Not in its technology, perhaps. But in its shape — in the way it disrupts industries, creates winners and losers, confuses ordinary people, and gradually, almost reluctantly, becomes part of daily life — this story is familiar. We lived through a version of it forty years ago.
It was called the personal computer.
What follows is not a technical argument. It is a historical one. It is an attempt to show that the map we are trying to draw for AI already exists — worn, creased, and surprisingly accurate — folded somewhere in the story of the PC revolution.
The Foundation Nobody Sees
When you use your laptop today, you are not thinking about the operating system running beneath everything you do. You are thinking about your email, your spreadsheet, and your music. The operating system — Windows, macOS, or Linux — is invisible. It is the stage on which the play happens. You do not notice it unless something goes wrong.
In the early 1980s, this was not yet the case. The operating system was new, contested, and confusing. Companies like Apple, IBM, Commodore, and Atari were each building their own. They did not speak to each other. A program written for one machine would not run on another. Choosing a computer meant committing to an entire ecosystem — a nerve-wracking decision because nobody knew which one would survive.
IBM made a fateful decision in 1981. Needing an operating system quickly, they struck a deal with a small company called Microsoft. Microsoft did not even have the software IBM needed — they bought it from a third party, repackaged it, and licensed it as MS-DOS. IBM, confident in its hardware brand, did not think to retain exclusive rights over the software.
This was one of the most expensive mistakes in corporate history.
Microsoft went on to license MS-DOS to every IBM clone manufacturer that followed. Within a decade, Microsoft — not IBM — owned the economic heart of the personal computer.
Now look at what is happening with artificial intelligence today.
The equivalent of the operating system in AI is the large language model — the enormous piece of software that understands human language and generates responses. ChatGPT, Claude, Gemini, Llama — these are today's competing operating systems. They are the invisible foundation beneath everything else. Most people who use AI products — a customer service chatbot on a bank's website, an AI writing tool in Microsoft Word, a medical summary generator in a clinic — do not know or care which model is running underneath, just as you do not think about Windows when you are writing an email.
The battles being fought right now over which AI model will dominate, which company will set the standards, and who will control the licensing — these are the same battles that were fought over DOS and Windows in the 1980s. The technology is different. The drama is identical.
At a Glance: Who Plays Which Role?
The Programs and the Agents
Once the operating system existed, something remarkable happened: people started building things on top of it.
A program called VisiCalc arrived in 1979. It was the world's first spreadsheet — a grid of rows and columns you could fill with numbers, and which would automatically calculate totals, percentages, and projections. Before VisiCalc, this was done by hand, by accountants with pencils and paper. VisiCalc made it instant. It was so useful that people bought personal computers specifically to run VisiCalc. The software was the point. The computer was merely the vessel.
Then came word processors — software that let you type, edit, move, and print text without the glacial tyranny of typewriters and correction fluid. Then databases, then design programmes, then accounting software. Each programme did one thing well. Each required you to learn its particular logic — its menus, its commands, its way of thinking. You worked within the programme's rules.
This is where we are in the history of AI applications, but something is shifting. The AI equivalent of a programme is called an agent. Like a programme, an agent does tasks on your behalf. But unlike a programme, you do not need to learn its menus or remember its commands. You simply say what you want — in ordinary language — and the agent figures out how to do it.
You might say: "Look at all the invoices in my email from last month, add them up, and flag any that are overdue." An AI agent can read your emails, identify the invoices, perform the arithmetic, and return a list — without you clicking through seventeen menus or remembering where anything is stored.
This is a genuinely significant change. It is the difference between driving a car with a manual gearbox and one with an automatic. The destination is the same. The effort required is not.
We are at an early and imperfect stage of this. Today's agents are capable but fragile — they make mistakes, get confused by ambiguous instructions, and sometimes confidently produce wrong answers. Early word processors crashed frequently, could not handle long documents, and had no concept of an "undo" button. The trajectory is what matters, not the current imperfection.
At a Glance: Old Software vs New AI Agents
The Killer App Question
In the history of personal computing, there is a concept called the killer app — the single application so useful, so obviously necessary, that it makes the entire platform indispensable.
For the personal computer, the killer app was the spreadsheet. Businesses could not afford to be without it. Once accountants, analysts, and managers had tasted the ability to recalculate an entire financial model in seconds, there was no going back. The spreadsheet is why computers entered offices before they entered homes.
For the internet, the killer app was email — and later, search.
For the smartphone, the killer app was the combination of maps, camera, and messaging in a single pocket-sized device.
For AI, nobody has yet identified the killer app with certainty. ChatGPT created enormous excitement, but excitement is not the same as indispensability. Millions of people have tried it; far fewer have restructured their working lives around it. It is still, for most users, a novelty with occasional bursts of genuine usefulness.
The killer app for AI is likely to emerge in a specific domain where improvements in speed and quality are so dramatic that returning to the old method becomes unthinkable. Medical diagnosis support, legal document review, and personalised education are all candidates. The spreadsheet did not replace accountants — it made them vastly more productive. The AI killer app will probably follow the same pattern: not replacing professionals, but making certain tasks so fast and accurate that those who use it gain an overwhelming advantage over those who do not.
We are waiting for that moment. It is coming. It may already be arriving in the corners of medicine, law, and engineering, quietly, without fanfare, the way VisiCalc first arrived in small accounting firms before anyone wrote about it in newspapers.
At a Glance: The Killer App Across Technology Eras
The Lock-In Trap
Here is something the personal computer era teaches that deserves more attention than it receives.
Once you chose an operating system in the early 1980s, you were not just choosing a piece of software. You were choosing an entire world. The programmes you bought, the files you saved, the skills you learnt — all of it assumed that operating system. Moving to a different one meant starting over. Companies invested in Apple systems and found they could not share files with colleagues on IBM machines. Entire departments became islands.
This is called lock-in. It is not unique to technology — it happens with currencies, languages, and infrastructure — but in technology, it is particularly acute.
We are watching the same thing happen with AI, and most people are not yet aware of it. The way you learn to use one AI assistant shapes your habits and expectations. The internal systems that companies build around one AI platform become deeply embedded. The files, structures, and workflows created with one tool require significant effort to migrate to another.
This matters because the company that achieves lock-in gains extraordinary power. Microsoft's dominance through the 1990s and 2000s was not because Windows was always the best operating system. It was because switching away from Windows was expensive, disruptive, and frightening. Microsoft could raise prices, slow innovation, and bundle mediocre products into the package — and customers would swallow it, because the cost of leaving was higher than the cost of staying.
The AI companies that achieve deep integration into business workflows — into the daily habits of knowledge workers — will gain similar power. Whoever becomes the Windows of AI will have captured something extraordinarily valuable, not just for years but for decades.
At a Glance: How Lock-In Looks in Practice
Open Versus Closed
Not everyone in the PC era accepted the logic of proprietary, closed systems.
Apple, under Steve Jobs, believed in tight control. Apple designed its own hardware, wrote its own operating system, and refused to let others copy its technology. The result was a machine of unusual elegance and reliability — but also one that cost more, ran fewer programmes, and appealed to a smaller audience. Apple nearly went bankrupt in the 1990s. It survived, ironically, partly because Microsoft invested in it.
IBM, by contrast, built the PC with open, standardised components that anyone could copy — and they were copied, aggressively. Within a few years, dozens of companies were making IBM-compatible machines at lower prices. This expanded the market dramatically and made personal computing affordable for ordinary people. But it also meant IBM could not control the ecosystem it had created. The value flowed to Microsoft, which owned the software that ran on all those machines.
Linux, a free and open operating system built collaboratively by programmers around the world, proved that another model was possible. Linux is today the most widely deployed operating system on earth — it runs most of the internet's servers, most of the world's smartphones (via Android), and most supercomputers — though ordinary consumers rarely see its name.
These three models — closed and premium (Apple), open standard with one dominant supplier (Microsoft/Windows), and truly open and free (Linux) — are all reappearing in AI.
Anthropic (the company behind Claude) and OpenAI pursue something close to the Apple and Microsoft models, respectively — proprietary, controlled, and sold as a service. Meta (the company behind Facebook) has released its AI model, called Llama, as open-source — available for anyone to download, modify, and use freely, much as Linux was. China has its own models. Europe has its own.
The outcome of these competing models will shape who profits from AI, who controls it, and how widely it is distributed. History suggests there is no single winning model — all three approaches survived in computing, serving different needs. The same is likely true for AI.
At a Glance: Three Ways to Own a Technology Platform
Rivals Who Fund Each Other
This is perhaps the strangest chapter in the history of technology — and it is repeating itself with a fidelity that should make us pause.
In 1997, Apple was close to bankruptcy. Its share price had collapsed. Its products were outdated. The press had largely written it off. And then, at a company conference, Steve Jobs walked on stage and announced that Microsoft — Apple's fiercest competitor, the company Apple employees considered a mortal enemy — had written Apple a cheque for $150 million.
The crowd booed.
The logic was not sentimental. Microsoft needed Apple to survive. If Apple disappeared, Microsoft would be the only computing platform in existence — and regulators would likely treat that monopoly with considerable hostility. Beyond that, Microsoft sold its Office software (Word, Excel, PowerPoint) on Apple computers and made good money doing so. Keeping Apple alive was, from Microsoft's perspective, simply good business.
It was an uncomfortable truth: sometimes you need your rival to survive, because the alternative — winning everything — creates problems of its own.
Fast forward to today, and this dynamic has returned — this time with sums of money that make the original Microsoft-Apple deal look like loose change.
Microsoft has invested $13 billion in OpenAI, the company that makes ChatGPT. Google has invested roughly $2 billion in Anthropic, the company that makes Claude. Amazon has also committed up to $4 billion to Anthropic. Each of these technology giants has its own artificial intelligence products that compete directly with the companies it funds.
Microsoft is simultaneously one of OpenAI's biggest investors and one of its biggest competitors — its own AI assistant, Copilot, is built on top of OpenAI's models. Still, Microsoft is also developing independent AI capabilities that could one day make OpenAI redundant to them.
Google funds Anthropic's Claude while trying to make its own Gemini model the industry standard.
Amazon pours billions into Anthropic whilst building its own AI services on Amazon Web Services — the cloud platform on which Anthropic also runs.
This is not hypocrisy. It is a strategy. The pattern from the PC era tells us exactly what is happening: companies invest in rivals they cannot afford to let fail, to ensure they have a seat at whatever table emerges, while hedging against the risk that their own bets do not pay off. In a period of genuine uncertainty about which technology will dominate, putting money into multiple competing horses is simply rational.
At a Glance: When Giants Fund Their Rivals
What the table above does not capture is the psychological dimension. In 1997, Apple employees watched their CEO accept money from the company they despised. In 2023, engineers at Google and Amazon are building AI tools that compete with the companies that their employers fund. This is not comfortable. But it is the nature of a platform war, where the outcome remains unresolved and the stakes are enormous.
The PC era teaches one other thing about these cross-investments: they do not last. Microsoft eventually stopped needing Apple to survive, and the $150 million became a historical footnote. Today's AI investments will likely follow a similar trajectory. As the technology matures and winners emerge, the strategic rationale for funding rivals will diminish. The investments are a symptom of uncertainty — and uncertainty, in technology, is always temporary.
From Specialists to Everyone
Perhaps the most important parallel of all concerns who gets to use the technology.
In 1975, using a personal computer required genuine expertise. You needed to understand programming — how to give a computer instructions in a language it could follow. The machine was unforgiving. A misplaced character in a command would cause it to do nothing, or crash, or produce nonsense. Computers in 1975 were for hobbyists, engineers, and the technically brave.
The graphical interface changed everything. Apple's Macintosh, released in 1984, replaced typed commands with pictures. Instead of typing "DELETE FILE REPORT.TXT", you dragged a small picture of a document to a small picture of a dustbin. This was obvious. This was human. This required no technical knowledge whatsoever.
The Macintosh democratised computing. It was no longer a tool for specialists. It was a tool for secretaries, designers, teachers, doctors, and children.
We are approximately 1981 in the history of AI. The tools are powerful but still require a certain kind of comfort with technology to use well. Knowing how to ask an AI model a question effectively — what is sometimes called prompt engineering — is today's equivalent of knowing how to type a command correctly into DOS.
The graphical interface for AI is currently being built. It will look different from anything we expect, but it will almost certainly involve a move away from typed instructions toward something more intuitive — voice, gesture, visual cues, or perhaps AI that understands your intent before you have fully articulated it.
When that interface arrives, AI will stop being a tool for the technically confident and become a tool for everyone. That moment, whenever it comes, will be as significant as the Macintosh.
At a Glance: The Journey from Specialists to Everyone
The Things That Are Different
Analogies are useful until they are not. There are three ways in which AI is genuinely unlike the personal computer revolution, and they deserve to be stated clearly.
The first is speed. The personal computer took roughly thirty years to go from hobbyist curiosity to mass consumer device. AI appears to be moving faster — far faster — though it is wise to be cautious about this. Every generation of technology believes it is moving at an unprecedented speed.
The second is energy. Moore's Law powered the personal computer revolution — the observation that computer chips roughly doubled in power every two years while halving in cost. This made the whole thing self-funding: the technology became cheaper as it improved. AI models require enormous amounts of electricity to train and run. The energy costs are high and growing. Whether this will limit the AI revolution, or simply redirect it toward more efficient methods, is genuinely uncertain.
The third — and most important — is agency. A word processor never had goals. A spreadsheet never makes decisions. The software of the personal computer era was, in a deep sense, passive. It did exactly and only what you told it to do. Current AI systems are beginning to do things their creators did not specifically programme them to do, in ways that are not always fully understood. This introduces questions about oversight, control, and safety that have no equivalent in the personal computer era. They are not reasons to abandon the technology, but they are reasons to think carefully about it in a way that was never necessary when the only risk was a spreadsheet crashing before you had saved your work.
At a Glance: Where the Analogy Holds — and Where It Breaks
What History Suggests
The personal computer story ended differently than most people expected at the beginning.
IBM, which launched the revolution, did not dominate it. Microsoft, a small company making a modest operating system, captured the most durable value. Apple, which came closest to getting the experience right, nearly died before becoming the world's most valuable company, through a product (the iPhone) that nobody anticipated.
The companies that dominated the early PC era — Wang, Commodore, Atari, Digital Equipment Corporation — are gone or marginal. The companies that dominate today were built on the foundations of that era without being bound by its assumptions.
This suggests a few things worth holding in mind as AI develops.
The companies that seem most powerful today may not be the ones that end up mattering most. The breakthrough that defines the AI era may come from a direction nobody is currently watching. The question of who captures the operating system layer — and who controls the applications built on top of it — will determine where the value accumulates, as surely as it did with personal computing.
And the extraordinary investments that rivals are making in one another — Microsoft in OpenAI, Google and Amazon in Anthropic — are not signs of confidence. They are signs of anxiety. Nobody knows who will win. That is why everyone is hedging. The PC era felt the same way in 1983, before Windows, before the internet, before anyone knew what a personal computer would eventually become.
Ordinary people — those who use these tools without understanding how they work, just as most people use computers without understanding how transistors function — will benefit from the competition between those trying to win their attention. The battles of the 1980s over operating systems, clunky as they were, eventually produced the iPhone.
Something similarly improbable is waiting at the other end of the AI revolution.
We have been here before. The terrain is familiar, even if the landscape looks different. And that is, on balance, a reassuring thing to know.