Reader Edition

READER EDITION

Who Builds the Future?

AI, Infrastructure, Power, and the Age of Abundant Intelligence

For the dreamers, the curious, and anyone wondering what kind of world is being built beneath the chatbot.

The Future Has a Basement

Artificial intelligence arrives on our screens looking almost weightless. A question goes in. An answer appears. A paragraph, an image, a computer program, a plan, a translation—produced in seconds. It can feel as though intelligence has somehow escaped the physical world.

But somewhere behind that clean little box on the screen, the future has a basement.

There are factories etching microscopic patterns into silicon. There are memory chips and networking switches. There are buildings the size of warehouses packed with machines that turn electricity into computation and computation into words, images, decisions, and predictions. There are substations, transmission lines, cooling systems, fiber-optic cables, transformers, construction crews, financing agreements, utilities, regulators, and enormous pools of capital.

The deeper we looked, the stranger the story became. The age of artificial intelligence may be remembered not only as a software revolution, but as the moment software became heavy industry.

That was not where this investigation began. It began with people.

Five Men and an Irresistible Pattern

Look at the technology world from a distance and a pattern seems almost impossible to ignore. Elon Musk appears across electric vehicles, batteries, artificial intelligence, reusable rockets, satellite communications, robotics, tunneling, and brain-computer interfaces. Jeff Bezos sits behind an empire spanning commerce, logistics, cloud computing, satellites, and a separate space company. Sam Altman’s world of AI increasingly touches data centers, electricity, manufacturing, chips, and financing. Mark Zuckerberg connects social networks, personal data, AI, wearables, fiber, custom computing hardware, and energy procurement. Jensen Huang’s NVIDIA has moved far beyond graphics chips into networking, AI systems, simulation, digital twins, and robotics.

From the present, it is tempting to draw a dramatic picture: a handful of extraordinary individuals assembling rival technological kingdoms, each building a private stack that might someday become part of the operating system of civilization.

It is a compelling story. It is also too neat.

Chronology broke it apart. SpaceX existed long before xAI. Blue Origin predates the current AI boom. Facebook began as a social network, not as one piece of a grand AI-energy machine. OpenAI began as an artificial-intelligence research organization, not as an industrial infrastructure company. NVIDIA’s path grew out of graphics and parallel computing. Many of the pieces now fitting together were born separately, for different reasons, in different decades.

The investigation therefore changed its mind. The evidence did not support a secret master blueprint designed years ago by a few founders. What it supported was more interesting: technologies that once developed independently have become increasingly useful to one another. Once those connections became obvious, powerful companies began positioning themselves around them.

The story stopped being “Who planned the future?” and became “Which layers of the future are becoming difficult to live without?”

The Stack Beneath the Magic

Imagine peeling an AI system apart layer by layer.

At the top are the things people see: chatbots, copilots, search tools, personal assistants, creative software, business applications, and perhaps eventually autonomous agents. Beneath them are the foundation models. Beneath the models are cloud platforms and enormous computing clusters. Beneath those are accelerators, memory, networking equipment, semiconductor fabrication, and the specialized machines needed to manufacture advanced chips.

Then the stack keeps going downward into the older physical world: electricity, transmission, cooling, water, land, concrete, steel, copper, construction, finance, and regulation.

And it also reaches outward. AI connects to phones, glasses, vehicles, factories, warehouses, satellites, drones, industrial machines, and robots.

This changes how power should be understood. The most famous company is not necessarily the most important company. A brilliant AI laboratory can still depend on somebody else’s chips. A chip designer can depend on a foundry. A data center can be ready while waiting years for an electrical connection. A robot can be physically impressive yet depend on a software platform, cloud service, or specialized hardware it does not control.

The crucial layer may be the one that is hardest to replace.

That is why the investigation gradually abandoned a simple picture of “vertical integration.” Some companies truly do own multiple stages of a supply chain. But elsewhere the connection is contractual, financial, technical, or platform-based. Dependency does not require ownership. A company can become extraordinarily influential simply because thousands of other companies build around its architecture.

When Software Became Heavy

The internet trained us to think of software as light. Build the code once, copy it endlessly. A service could reach millions of people without constructing a refinery or a railroad.

Frontier AI bends that logic. The final product can be almost weightless while the machinery producing it becomes enormous.

This is one of the central paradoxes of the entire investigation: the factory gets heavier while the product gets lighter.

Advanced AI increasingly requires industrial-scale investment in computing hardware and data centers. Those facilities need enormous electrical connections, cooling systems, land, construction, and financing. Technology companies that once seemed far removed from utilities now negotiate for power, study grid constraints, and sign long-term energy agreements.

Yet the intelligence emerging from those facilities can become cheaper by the month. Capabilities that once required a team, specialized software, or expensive expertise can be placed in the hands of an ordinary person through a browser.

That combination—expensive production and cheap access—is the key to understanding why simple stories about AI power can mislead us.

The Invisible Race for Electricity

The energy story sounds simple until you look closely. AI uses a lot of electricity. Therefore electricity becomes the bottleneck. But infrastructure rarely behaves so cleanly.

The real problem is often not whether a country has enough electricity in total. It is whether a particular site can receive enough reliable electricity at the moment a data center needs it. A project can have land, financing, servers, and customers and still be delayed by a transformer, a substation, a transmission line, a permit, or an interconnection queue.

This helps explain the sudden collision between Silicon Valley and energy policy. Technology companies are exploring or contracting for renewables, storage, natural gas, geothermal power, existing nuclear plants, and proposed advanced reactors. But the timelines differ. A data center can rise faster than a new power plant or major transmission project.

The important lesson is not that electricity will always be the great constraint. It is that bottlenecks move.

At one moment the scarce thing may be advanced chips. Later it may be deliverable power. Then perhaps reliable autonomous agents. Then identity, authorization, cybersecurity, or something we have not yet recognized. The future may be shaped less by one permanent choke point than by a succession of them.

The Other Infrastructure Nobody Sees

There is another layer beneath the machines: money.

Gigawatt-scale data centers, semiconductor fabs, fiber networks, power systems, and robotic factories require staggering amounts of capital. That pulls banks, private-credit funds, asset managers, sovereign wealth funds, insurers, utilities, and governments into a story that began as software.

This was another reason the original “five billionaires build the future” framing failed. Founders can set agendas. They can make unusual bets. They can commit enormous resources. But they cannot construct this system alone.

The emerging infrastructure is being co-produced by technology companies, chip suppliers, utilities, investors, governments, customers, workers, regulators, and local communities.

And there is an uncomfortable historical possibility. A technology can transform civilization while many of the people who financed its first great build-out lose money. Railroads can be overbuilt and still leave railroads behind. Fiber can be overbuilt and still make communications cheaper. AI infrastructure could prove technologically revolutionary and financially excessive at the same time.

If that happens, bankruptcies and write-downs would not necessarily mean the technology failed. They could leave behind cheaper infrastructure for whoever comes next.

The Great AI Paradox

Now we arrive at the contradiction that survived almost every attempt to simplify the investigation.

At the top of the supply chain, advanced AI can be extraordinarily concentrated. The most sophisticated chipmaking equipment, leading-edge fabrication, specialized accelerators, hyperscale cloud infrastructure, and frontier training systems involve huge barriers to entry.

At the bottom, intelligence can become astonishingly cheap.

A student, a small business, an independent researcher, or a person sitting at a kitchen table can gain access to capabilities that would have required a large organization not long ago. The infrastructure can centralize while the capability decentralizes.

Both things can be true at once.

That means the debate cannot be reduced to “AI oligarchy” versus “AI democratization.” A society could have concentrated infrastructure ownership and broad access. It could have broad access but restrictive platform rules. It could have public investment in infrastructure while private companies capture much of the profit. Or consumers might receive enormous benefits even while a small number of firms own the machinery underneath.

Three questions therefore have to be kept separate: Who owns the infrastructure? Who controls access and the rules of use? Who captures the gains?

The answers may be completely different.

From Machines That Answer to Machines That Act

A chatbot that gives advice is one thing. A machine that acts on the advice is another.

Imagine an AI that can read your email, schedule appointments, purchase products, negotiate with software systems, write and deploy code, move money, sign up for services, or eventually direct physical machines. At each step, the system crosses a boundary from producing information to exercising authority.

This may be a more important near-term threshold than the philosophical argument over whether a machine has reached “general intelligence.”

But capability is racing ahead of permission. Before autonomous agents become ordinary economic actors, society has to solve less glamorous problems: identity, authentication, fraud, cybersecurity, authorization, revocation, auditing, liability, and responsibility.

Who does the machine represent? What is it allowed to do? How does another machine know the answer? What happens when it is tricked? Who pays when it is wrong?

This led to one of the investigation’s most important late discoveries: trust may itself become infrastructure.

When producing words, images, recommendations, and actions becomes cheap, verifying them can become more valuable. The cheaper generation becomes, the more society may pay for provenance, reputation, identity, authorization, and proof.

Robots, Without the Science-Fiction Filter

Humanoid robots are irresistible. They look like the future, which makes them dangerous to analysis.

A machine does not need arms, legs, and a face to transform physical work. Autonomous tractors, mining trucks, warehouse systems, factory manipulators, drones, vehicles, inspection machines, and specialized robots may matter more than humanoids for a long time.

The useful question is not “How many humanoid robots exist?” It is “How much economically valuable physical work can machines perform reliably and cheaply?”

That answer depends on boring measurements: productive hours, maintenance, safety, intervention rates, task switching, supervision, lifetime cost, and depreciation. A robot that performs a dazzling demonstration but constantly needs a human to rescue it is not autonomous labor.

The investigation therefore ended with strong confidence that AI will expand industrial automation, less confidence about general-purpose physical intelligence, and much less confidence in confident forecasts of a near-term humanoid world.

The Missing Middle of Work

AI can make a person faster without making an organization better.

That distinction may decide whether the current boom becomes a true productivity revolution.

A worker can save an hour writing a report, but if the report still moves through the same approvals, meetings, incentives, legacy systems, compliance checks, and management habits, the organization may capture little of that saved time. Technology can change in months while institutions change in years.

This is the missing middle between impressive AI capability and economy-wide transformation.

The labor story is equally complicated. AI can replace tasks, complement workers, lower the cost of expertise, reduce hiring, create new demand, increase returns to capital, or do several of those things simultaneously.

One subtle risk emerged late in the investigation. Many professions train experts by giving beginners routine work. Junior programmers write simpler code. Young lawyers review documents. Analysts build models. Research assistants gather and summarize material. If AI removes the beginner work while leaving senior work intact, fewer people may get the practice that creates tomorrow’s experts.

Or the opposite could happen. AI could become the greatest tutor ever created, helping novices learn faster.

We do not know yet. The important measurement is not merely how well someone performs while AI is beside them. It is what they can do after the AI is taken away.

Who Actually Chooses?

The original investigation imagined a dramatic contest between private vision and public choice. Reality is messier.

Technological societies rarely hold one grand vote on their infrastructure. Founders decide what to attempt. Investors decide what to finance. Governments permit, subsidize, purchase, regulate, tax, restrict, or encourage. Utilities decide how loads connect. Standards bodies determine whether systems can communicate. Courts define liability. Customers adopt products. Workers adapt. Communities approve or resist construction. Network effects reward some winners and punish others.

Then infrastructure accumulates around yesterday’s decisions.

That is how path dependence works. Nobody needs to order society to become dependent on a technology. If schools, workplaces, commerce, customer service, medicine, finance, and government gradually assume that people will use AI, refusing it may become costly even if refusal remains technically possible.

At the same time, dependence on AI can increase independence elsewhere. A disabled person may rely more heavily on technological infrastructure while becoming less dependent on another human for daily tasks. A tiny company may become dependent on a hyperscale cloud while gaining capabilities once reserved for a corporation.

Dependency and freedom are not opposites. They can increase together in different dimensions.

The practical question is therefore not simply who “controls the future.” It is: at each important layer, who can make consequential decisions, who can say no, how easy is it to switch, and how reversible are the choices being made?

The Future Will Not Arrive All at Once

One reason predictions about AI become unreliable is that the different parts of society move at different speeds.

Computer science can advance quickly. A new model can appear in months. Electrical transmission can take years. Organizations often change more slowly than software. Law frequently reacts after deployment. Professional norms can take a generation to evolve. Buildings and power systems can remain in place for decades. Human habits can persist even longer.

So there is no single “AI revolution” moving at one speed. There are several coupled transitions—scientific, industrial, financial, organizational, legal, psychological, and political—continually getting ahead of one another.

For the next several years, the safest expectations are the least cinematic ones: more computing capacity, more data centers, more electricity demand, more AI embedded in ordinary software, and more experimentation with industrial automation. Beyond that, confidence falls. Agents may become far more autonomous. Robotics may spread dramatically. National AI strategies may deepen. But the scale and timing are uncertain.

Two decades out, almost everything becomes scenario rather than forecast. Ubiquitous humanoids, post-scarcity economies, corporate quasi-states, orbital computing, machine consciousness, or any particular company dominating civilization belong in the realm of possibilities, not findings.

The Human Premium

There is another possibility that turns the usual automation story inside out.

If machines can generate endless competent writing, images, music, advice, and service, human creation does not necessarily become worthless. It may become scarce.

A handmade object can become more valuable in an age of factories. A live performance can matter more in an age of recordings. In the same way, a verified human teacher, artist, caregiver, writer, craftsperson, performer, or advisor may acquire a new kind of premium precisely because the work is demonstrably human.

This is not established. It is a hypothesis. But it reveals a broader principle that became increasingly important as the investigation matured: abundance creates new scarcity.

Infinite content makes attention scarce. Synthetic media makes authenticity scarce. Cheap advice can make judgment scarce. Autonomous systems can make responsibility scarce. Easy generation can make verification scarce.

The success of AI may therefore increase the value of things that once seemed secondary.

The Scarcity Inversion

This is where the entire investigation finally turned over.

We began by assuming that advanced intelligence was the prize. Whoever possessed the most powerful intelligence would possess the future.

But what if AI succeeds so well that useful machine intelligence becomes abundant?

Industrialization made physical power cheap relative to human muscle. Computers made calculation cheap. The internet made distribution cheap. Each time, value migrated toward something that remained difficult to obtain.

AI may do the same to cognitive production.

If analysis is abundant, perhaps judgment becomes scarce. If writing is abundant, perhaps trustworthy authorship becomes scarce. If images are abundant, perhaps provenance becomes scarce. If advice is abundant, perhaps responsibility becomes scarce. If agents are abundant, perhaps legitimate authority becomes scarce. If content is infinite, attention becomes scarce.

And beneath all of that remain stubbornly physical scarcities: energy, land, minerals, housing, factories, skilled trades, healthcare capacity, construction, transportation, and the ability to make something happen in the real world.

This may be the most important question produced by all 456 pages of research:

What remains scarce when intelligence no longer is?

Perhaps the answer will be energy. Perhaps trust. Perhaps human attention. Perhaps authenticity, relationships, judgment, responsibility, capital, authority, or access to physical resources. Most likely it will be some shifting combination.

That possibility changes how we should think about power. The most durable power may not belong to whoever builds the smartest model. It may belong to whoever controls the scarce complements that intelligence cannot cheaply reproduce.

And it changes how we should think about ourselves.

For most of history, intelligence has been tangled up with scarcity. Expertise was scarce because experts required years of training. Advice was scarce because a person had to provide it. Creation was scarce because somebody had to do the creating. If machines make portions of those things abundant, human beings may have to discover which parts of intelligence were truly valuable and which parts were merely expensive.

Goals may matter more than answers. Taste may matter more than generation. Responsibility may matter more than recommendation. Relationships may matter more than communication. Wisdom may matter more than information.

None of that is guaranteed. The investigation deliberately resisted turning possibilities into prophecies. AI may produce extraordinary productivity gains, or organizations may discover that intelligence was never their main bottleneck. Infrastructure investors may earn fortunes, or build too much too quickly. Agents may become trusted economic actors, or security and liability may keep humans firmly in the loop. Robotics may transform physical labor, or remain specialized far longer than enthusiasts expect.

The future is still open.

So, Who Builds the Future?

Not one founder. Not one company. Not one government. Not one technology.

The future emerging around AI looks less like a blueprint and more like an evolving negotiation among machines, markets, infrastructure, institutions, laws, communities, and human behavior.

The famous founders matter. Their decisions can redirect billions of dollars and accelerate entire industries. Large corporations matter because they own infrastructure and platforms that others depend upon. Governments matter because they control law, permitting, procurement, trade, energy policy, competition rules, and public investment. Investors matter because visions without capital remain visions. Consumers and workers matter because technologies that nobody adopts do not become infrastructure.

No actor controls every conversion point.

An AI laboratory can build a remarkable model but cannot command a transmission line into existence. A cloud company can build a data center but cannot abolish local permitting. A government can regulate a technology but cannot legislate a scientific breakthrough into being. A robot company can build a machine but cannot force the economics to work. A consumer can enthusiastically adopt an AI service without owning any of the infrastructure beneath it.

That may be the most reassuring and the most unsettling conclusion at the same time.

There is no single architect of the future. But there are moments when choices harden into infrastructure, standards become defaults, convenience becomes dependence, and temporary advantages become difficult to reverse.

Those are the moments worth watching.

The investigation began by looking upward at a handful of powerful people and asking what they were building for us. It ended by looking downward—beneath the software, beneath the companies, beneath the personalities—at the layers of dependency being assembled across society.

And then it looked inward.

If intelligence becomes cheap, what will we still value? What will we still trust? What will we refuse to delegate? What will we insist on doing ourselves?

Those questions cannot be answered by a benchmark, a data center, a billionaire, or an algorithm.

They belong to everyone who will live in the world these systems help create.

A Note on Certainty

This Reader Edition is designed to tell the story clearly without flattening the evidence. The investigation found strong support for rapidly improving AI capability, massive infrastructure investment, growing physical dependence on chips, data centers and electricity, concentration in important upstream layers, falling costs for many AI capabilities, and increasing government involvement.

Confidence becomes lower as the story moves toward economy-wide productivity, long-term labor restructuring, autonomous agents, general-purpose robotics, durable platform lock-in, future wealth distribution, cognitive deskilling, and political dependency. Claims about specific conditions decades from now remain scenarios rather than established forecasts.

The point is not to make the future sound inevitable. It is to make the choices, dependencies, and unanswered questions easier to see.

THE INVESTIGATION ARCHIVE