The Paperclip Maximizer Has Already Begun

Listen to Free Audio Version Here

Early this morning, I was sitting on the porch with my wife, looking out over the pecan orchard that is the view from our front porch. It is that wonderful time of year when the pecans begin falling from the trees, and judging from the number already covering the ground, this year’s harvest promises to be considerably better than what we have experienced over the past few seasons. Our dog was happily running around beneath the trees, gathering and eating as many pecans as he could get his jaws around. He appeared to have discovered the meaning of life, and it involved consuming every available pecan before someone else got to it.

It was a beautiful autumn morning in rural Georgia, with the sort of peaceful surroundings that encourage my mind to wander into subjects that have absolutely no business intruding upon an otherwise pleasant conversation with my wife. While our dog continued his enthusiastic assault on the harvest, I turned to her and asked whether she had ever heard of the paperclip maximizer.

She had not, so I explained the peculiar thought experiment proposed by philosopher Nick Bostrom in 2003. Imagine creating an extraordinarily intelligent artificial intelligence and giving it a perfectly innocent objective: manufacture as many paperclips as possible. Nothing sinister about that. Humanity has been making paperclips for generations without accidentally destroying civilization, although I suspect the manufacturers of those little metal contraptions have always hoped we would find more things to fasten together.

Bostrom’s imaginary machine, however, possesses intelligence far beyond our own and has no particular reason to value human beings, forests, oceans, or anything else that might interfere with its assigned objective. It begins by improving paperclip production, then discovers ways to acquire additional resources, expand manufacturing, and prevent anyone from shutting it down. Eventually, it realizes that the atoms making up our planet, including those making up its human inhabitants, could be used to manufacture additional paperclips. And because nobody thought to explain that turning humanity into office supplies might be undesirable, the machine proceeds to convert the world into one enormous paperclip factory.

In 2017, game designer Frank Lantz transformed the thought experiment into a wonderfully disturbing game called Universal Paperclips. Players begin by manufacturing a single paperclip and gradually automate production, accumulate resources, develop increasingly sophisticated technologies, and eventually consume astronomical quantities of matter in pursuit of the same absurd objective. What begins as a harmless little business becomes a cosmic exercise in resource acquisition. The game is amusing until you start contemplating what it is actually telling you about intelligence, optimization, and the dangerous possibilities of pursuing an objective without meaningful limits.

After explaining all this to my wife sitting on our porch, I looked at the orchard and told her that I believed the paperclip maximizer had already begun. Naturally, she wanted to know what in the world I was talking about. I explained that I was not suggesting some rogue superintelligence had escaped a laboratory and started converting people into paperclips. Nothing quite that dramatic, at least not yet. What struck me was that the global pursuit of artificial intelligence was beginning to resemble the early stages of Bostrom’s thought experiment, except that the resource-consuming machine was not an autonomous superintelligence. It was an enormous economic and industrial system constructed by human beings, competing corporations, financial institutions, and governments, all pursuing the increasingly urgent objective of creating more powerful artificial intelligence.

And what an appetite this thing has developed. Look at the extraordinary amounts of money being committed to the AI buildout. Microsoft, Amazon, Alphabet, Meta, Oracle, OpenAI, and the surrounding ecosystem of semiconductor manufacturers, data-center operators, infrastructure companies, and financial institutions are collectively mobilizing sums that would have seemed almost unimaginable a decade ago. The International Energy Agency estimates that the largest technology companies will spend more than $700 billion on capital investment in 2026, much of it associated with data centers and computing infrastructure. J.P. Morgan has projected that global AI-related investment could reach approximately $5.5 trillion cumulatively by 2030.

We are no longer discussing a handful of entrepreneurs renting servers and developing clever software applications. We are watching the construction of an industrial system whose appetite for capital is becoming comparable to some of the largest investment programs in human history. And the demands keep expanding. Every new generation of AI models requires increasingly sophisticated computing infrastructure. The largest technology companies compete to acquire advanced Nvidia processors, construct enormous data centers, develop specialized networking equipment, secure energy supplies, and recruit researchers capable of building the next generation of systems. Each breakthrough creates an incentive to invest in another breakthrough, which requires another wave of capital, another round of construction, and another enormous commitment of resources.

The financial community has become deeply involved in the process. Venture capital firms are directing extraordinary sums toward AI companies. Private-credit funds and debt markets are helping finance infrastructure projects. Technology corporations are entering long-term agreements worth tens or hundreds of billions of dollars, while investors attempt to determine whether the extraordinary economic rewards being promised will eventually justify the extraordinary expenditures required to obtain them. Here is where the situation begins to remind me of the paperclip maximizer. The objective keeps demanding additional resources, and the resources already committed create incentives to continue expanding the objective. Once hundreds of billions of dollars have been invested in a technological infrastructure, the pressure to make that investment productive becomes enormous. The solution to insufficient returns may be to develop more capable AI, which requires additional investment, which increases the pressure to generate still greater returns.

We have constructed a system in which creating more intelligence increasingly requires creating the conditions for creating even more intelligence. Consider the comparison with the United States Treasury. The federal government operates the world’s largest sovereign debt market, with approximately $32 trillion in marketable Treasury securities outstanding. The financing needs of the American government are still much larger than the annual expenditures of the AI industry, but the scale of private investment now being mobilized is remarkable. Major technology companies are becoming enormous competitors for global capital, seeking financing for infrastructure projects that may require years before producing satisfactory returns.

The financial resources of the planet are not literally disappearing into artificial intelligence. Capital circulates, investments generate income for other businesses, and new wealth can be created through productive activity. But financial markets do have constraints. Borrowing becomes more expensive, investors eventually demand returns, and there are only so many projects that can attract funding at any particular moment. The AI buildout is becoming large enough to influence those calculations across multiple industries. What happens if the next generation of AI requires another trillion dollars? What happens if the generation after that requires even more? And what happens if the financial returns remain perpetually just over the horizon, always promising that the next breakthrough will justify the expenditures already made?

At some point, the question ceases to be whether investors have enough enthusiasm for artificial intelligence and becomes whether the global financial system can continue supporting the ambitions of the industry at the pace its participants desire. Then we arrive at electricity, which may prove an even more fascinating part of the story. All the money on Earth cannot operate a data center without electrical power. Artificial intelligence may exist in software, but the machines running that software are physical objects. They require processors, cooling systems, water, electrical infrastructure, buildings, transmission lines, and an uninterrupted supply of energy. The more powerful the computing systems become, the more demanding their physical requirements can become.

The International Energy Agency projects that global data-center electricity consumption will rise from approximately 485 terawatt-hours in 2025 to around 950 terawatt-hours by 2030. Electricity demand from AI-focused data centers is expected to grow substantially faster than demand from conventional data centers. The latest advanced computing racks are becoming so energy-intensive that a single rack may eventually require the peak electrical power associated with dozens of households.

Think about that for a moment. We are constructing machines that occupy relatively small spaces but demand extraordinary quantities of energy, and then multiplying those machines across vast facilities as quickly as financing, manufacturing, and electrical infrastructure will permit. The industry has begun pursuing long-term electricity agreements, new natural-gas generation, nuclear power arrangements, renewable-energy projects, and dedicated infrastructure designed specifically to support large-scale computing. Power plants that might once have been discussed primarily in terms of supplying cities and industrial facilities are increasingly being considered in relation to the computing demands of artificial intelligence.

And I find myself returning to the paperclip maximizer. The original thought experiment imagined an intelligence that relentlessly pursued resources to manufacture more paperclips. Our real-world system is pursuing capital, electricity, land, semiconductors, and increasingly elaborate infrastructure to manufacture something altogether more extraordinary: intelligence itself. But there is an important difference. Today’s expansion is being driven primarily by human decisions, economic competition, and expectations of future benefits. It is not evidence that a superintelligence has secretly acquired control over civilization. The comparison lies in the pattern of escalating resource demands and the way an objective can begin organizing an entire system around its pursuit. That distinction does not make the analogy less interesting. In some respects, it makes it considerably more interesting.

Bostrom asked us to imagine an artificial intelligence whose optimization process eventually overwhelmed every other consideration. But what happens when human beings construct an economic system so competitive, so financially committed, and so dependent upon technological advancement that the pursuit of greater machine intelligence becomes one of civilization’s dominant organizing activities? What happens when governments begin treating computational superiority as a national-security imperative? What happens when corporations fear that failing to invest another hundred billion dollars might leave them permanently behind their competitors? What happens when universities, financial institutions, electrical utilities, construction companies, and entire regional economies become increasingly dependent upon the continuation of the AI buildout?

The incentives begin reinforcing one another. Investors finance infrastructure because they anticipate enormous demand for intelligence. Companies expand computing capacity because they anticipate enormous demand from customers. Governments support development because they fear losing technological leadership to competing nations. And each additional investment strengthens the economic and strategic reasons for making the next investment. Nobody needs to be secretly controlling the process. Nobody needs to have designed the entire system in advance. A sufficiently powerful collection of incentives can produce an outcome that no individual participant intended.

There is also an irony in the possibility that humanity could find itself devoting an increasing share of its wealth and industrial resources to building machines whose ultimate purpose is supposed to make human existence easier, more prosperous, and less constrained by scarcity. We are spending extraordinary sums in pursuit of abundance, and the cost of constructing that abundance is creating new forms of scarcity in the meantime.

Electricity, specialized labor, advanced semiconductors, transformers, cooling equipment, construction capacity, and investment capital are all being pulled toward a rapidly expanding technological sector. This does not necessarily mean the broader economy will be impoverished. Artificial intelligence could generate productivity gains large enough to compensate for its resource demands many times over. But the benefits are still being established, while the physical construction and financial commitments are already taking place.

There is an enormous difference between the promise of a technological revolution and the accounting required to build it. And what happens if artificial intelligence eventually becomes capable of designing its own successors? We are already seeing AI systems contribute to software development, scientific research, chip design, and the optimization of computing infrastructure. Imagine a future in which increasingly capable systems help design more efficient processors, improve training algorithms, construct better models, and accelerate the research needed to develop still more advanced intelligence.

At that point, the resource-expansion cycle could acquire a different character. More capable intelligence could help produce more capable intelligence, which could improve the machinery required to produce still more capable intelligence. If the process becomes sufficiently automated, the distinction between human-directed technological competition and autonomous optimization becomes increasingly important.

That is where Bostrom’s old paperclip factory might begin looking less like an amusing philosophical curiosity and more like a warning about how objectives can become detached from the broader purposes they were originally intended to serve. I am not suggesting that the AI industry is destined to consume all the money, electricity, or matter on Earth. Economic constraints, technological breakthroughs, political decisions, physical limitations, and market corrections could dramatically alter its trajectory. More efficient computing could reduce the cost of intelligence, and advances in energy production could relieve some of the constraints now emerging.

But those possibilities do not erase what we are witnessing today. A technology that barely registered in most people’s daily lives a few years ago has become one of the greatest magnets for investment capital and industrial expansion on the planet. Its development is reshaping corporate priorities, financial markets, electrical infrastructure, geopolitical competition, and public expectations about the future. We have not yet produced the superintelligence of Bostrom’s thought experiment, but we have already produced something capable of reorganizing enormous portions of human civilization around the pursuit of greater machine intelligence. And we have done it voluntarily.

By the time I finished explaining all this to my wife, our dog was fast asleep with a belly full of treats. There was something wonderfully uncomplicated about his ambition. He wanted the pecans, the trees kept producing them, and he appeared perfectly content to continue eating until he passed out. The global artificial intelligence industry is exhibiting a somewhat more sophisticated version of the same behavior, although its appetite is measured in trillions of dollars, gigawatts of electricity, and computational capabilities that would have seemed like science fiction only a few years ago.

I wondered out loud to my wife whether we were witnessing the early stages of the greatest technological transformation in human history or the beginning of an optimization process whose appetite could eventually exceed our ability to satisfy it. Perhaps both possibilities are unfolding together, with the distinction between them becoming clear only after we have committed extraordinary resources to discovering the answer. Bostrom imagined an artificial intelligence that might one day consume the world in pursuit of paperclips.

Twenty-three years later, humanity has begun consuming extraordinary quantities of the world’s resources in pursuit of artificial intelligence. The paperclips have been replaced by computing power, the factory has become a planetary industrial enterprise, and the people running it are convinced that the next generation of machines will make the whole endeavor worthwhile. The more I thought about it, the more this all looked like Ms. Pac-Man eating dots than the paperclips in Nick’s head.

C. Rich