What had to exist before machine intelligence became possible, practical, and available for use?
The systems behind the capability
An artificial intelligence system can produce an answer in seconds, generate an image from a description, or help carry out a task that once required several separate tools. The interaction may feel self-contained. A person supplies an instruction and receives a result, with little indication of the physical equipment, accumulated research, and operational systems involved.
That apparent simplicity is part of what makes modern technology useful. It also makes the conditions behind a capability easy to overlook. Electricity arrives at a socket without exposing the network that supplies it. A digital photograph appears on a screen without revealing the manufacturing processes, software, storage, and communications systems involved in its creation and delivery.
Machine intelligence presents the same problem of visibility. The model is the component people encounter, but its development and use depend on a much larger arrangement of scientific knowledge, computation, semiconductor manufacturing, data, software, energy, communications, capital, and organizational capacity. These elements emerged through different histories, and their relationships were not established in advance.
This raises a question that is more revealing than asking what a model can do: what had to exist before that capability became possible, practical, and available for use?
The question changes the subject of investigation. Instead of looking only at the invention, it directs attention to the conditions that enabled it, the work required to make it dependable, and the systems that allowed it to operate beyond an isolated demonstration.
When a machine becomes useful
The history of electronic computing shows why the distinction matters.
ENIAC, completed in 1946 at the University of Pennsylvania, demonstrated the potential of high-speed electronic computation. Yet programming the machine involved setting switches and changing plug connections. Its ability to perform calculations did not mean that instructions could be changed as conveniently as they would be in later computers.
Stored-program computing changed that relationship by allowing instructions to be held in memory. The Standards Electronic Automatic Computer, known as SEAC, began useful computation in May 1950 at the United States National Bureau of Standards. Its program could reside in internal memory and be modified during successive stages of computation. Although originally intended as an interim machine, SEAC remained productive for fourteen years.
Russell A. Kirsch's historical account, Computer Development at the National Bureau of Standards, documents SEAC's development and use. The example matters because it shows a change in the practical organization of computation. Instructions were no longer only a matter of configuring the machine from outside; they could become part of the machine's internal operation.
This did not make earlier work irrelevant, nor did stored-program computing emerge from a single invention in isolation. Useful computers required electronics, memory, mathematical methods, engineering, maintenance, and institutions able to support expensive equipment. Government agencies, universities, and companies contributed to the development of computing in different ways, as documented in the National Academies' historical study Funding a Revolution: Government Support for Computing Research.
The broader lesson is not that every technology follows the same path as early computers. It is that a technical achievement and a useful capability are different things. The achievement establishes what can be done under particular conditions. A dependable capability requires those conditions to be understood, reproduced, supported, and adapted to practical needs.
The distinction is important because the work that turns an invention into a usable system is often distributed across many people and organizations. It may receive less attention than the breakthrough itself, yet it can determine whether the breakthrough has lasting consequences.
The distance between discovery and production
A scientific principle can be understood long before it can be used in a reliable product. A device can work in a laboratory before it can be manufactured consistently. A machine can perform a task before it can be maintained economically or integrated into a larger operation.
Semiconductor technology makes these differences particularly clear. Understanding how a device behaves is one part of the problem. Producing components with controlled properties, repeatable quality, and sufficiently low defect rates requires specialized materials, equipment, fabrication processes, measurement, and quality control. Integrated circuits add further challenges involving design, production, packaging, and the coordination of many components. The Computer History Museum's Silicon Engine history documents how the semiconductor industry developed through successive advances in devices, materials, and manufacturing processes.
The industrial capability is therefore more than the scientific principle or the device itself. It includes the accumulated knowledge and machinery required to reproduce the device under controlled conditions.
This distinction helps explain why technical progress cannot be measured by inventions alone. The practical value of a technology also depends on whether it can be produced at a suitable scale, at an acceptable cost, and with the reliability required by its intended use. These conditions vary by technology and application. They can improve over time, but they do not improve automatically simply because a new design has been discovered.
Computing's development involved similar relationships among electronic components, memory, programming methods, software, and communications. Research institutions, public funding, commercial demand, and industrial production each contributed to different parts of that development. Their contributions were neither identical nor perfectly coordinated.
Technological history is consequently not just a record of discoveries. It is also a history of how societies develop the means to reproduce, connect, finance, and use what they discover.
The physical conditions of intelligence
Contemporary AI makes these relationships visible again, particularly as the resources required to train and operate large models have expanded.
Modern computing depends on semiconductor devices produced through specialized industrial processes. Large-scale AI workloads also require computing accelerators, memory, storage, networking, and facilities that can supply and manage electricity and cooling. Those facilities depend on construction, equipment supply chains, financing, and access to supporting infrastructure.
The International Energy Agency's 2026 report, Key Questions on Energy and AI, examines the relationship between AI development, data-centre investment, electricity demand, grid capacity, chip manufacturing, and financing. Its analysis also highlights uncertainty about future demand, which will depend on model development, improvements in efficiency, adoption, and the kinds of applications people use.
That uncertainty is important. The resources required by AI are not fixed across all systems or tasks. More efficient hardware and algorithms can reduce the resources needed for a given workload. At the same time, lower costs and new capabilities can encourage wider adoption or make new resource-intensive applications viable. The overall effect depends on how these changes interact.
The physical requirements also have economic and institutional consequences. Computing equipment must be purchased and operated. Electricity must be supplied. Facilities must be built. Specialized components must be manufactured and delivered. Organizations differ in their access to capital, equipment, expertise, and supply chains, which can shape the systems they are able to develop and deploy.
This does not mean that only large organizations can build useful AI, or that infrastructure alone determines which approaches succeed. It means that access to knowledge and access to the means of putting that knowledge to work are different things. A model's technical availability does not guarantee that every organization can operate it at the same cost, reliability, or scale.
The distinction between capability and access is one of the reasons the infrastructure behind AI deserves historical attention. It affects not only what can be built, but also who can put it to use and under what conditions.
Why the layers interact
It is tempting to arrange the history of machine intelligence as a straightforward sequence: energy enabled machinery, machinery enabled precision manufacturing, semiconductors enabled modern computers, computers enabled machine learning, and machine learning produced contemporary AI.
The sequence offers a convenient overview, but it is not a sufficient explanation. It can make the past appear to have been moving toward the present in a single direction, as though each development existed primarily to prepare the way for the next one.
The actual relationships are more complicated. Manufacturing capabilities influence which designs are practical. New computing methods create demand for different hardware. The availability and quality of data affect which learning approaches can be used. Commercial demand can attract investment, while public research can support work whose applications remain uncertain for years. A new capability can also expose limitations in the infrastructure that supports it, creating demand for further research and investment.
These interactions mean that technological development involves feedback as well as accumulation. Later capabilities draw on earlier ones, but they can also change the value, use, and limitations of the systems that came before them.
The path is not inevitable. Different designs compete, institutions make choices, research directions receive unequal support, and constraints redirect effort. Some technologies become widespread while others remain specialized or disappear. The existence of an earlier invention does not prove that it directly caused a later breakthrough, and an enabling condition should not be confused with a sufficient cause.
This matters when evaluating AI itself. Strong performance on a benchmark does not establish reliability in every real-world setting. A successful demonstration does not establish that a system can be operated economically at scale. Broad access to a model does not guarantee that an organization has the data, skills, processes, or incentives to use it effectively.
These are separate questions because each depends on different conditions. Technical performance concerns what a system can do under specified circumstances. Reliability concerns whether it continues to do so under the conditions in which it will be used. Economic viability concerns the resources and costs involved. Effective use also depends on the surrounding organization and the task being performed.
The capability matters, but the conditions determine much of what can be done with it.
A different way to study intelligence
Beneath Intelligence begins with a question that can be applied to every major technological development: what had to exist before this became possible?
Answering it requires attention to several overlapping domains. Scientific knowledge establishes what can be understood or attempted. Materials, tools, and manufacturing processes determine what can be built. Measurement and standards make it possible to test and reproduce results. Computing and communications allow information and instructions to be processed and transferred. Institutions, markets, and public investment influence which capabilities are developed, sustained, and made available.
These domains should not be treated as a rigid checklist. Their importance changes across periods and technologies, and the connections between them need to be demonstrated rather than assumed. A historical account must distinguish direct causes from enabling conditions, identify what remained constrained, and explain how the available options changed.
It must also leave room for contingency. An enabling system can expand the range of possible outcomes without determining which outcome will occur. Research choices, institutional priorities, commercial incentives, and unexpected discoveries all influence the path from possibility to practice.
The purpose of this series is to investigate those relationships through the histories of the systems that underpin machine intelligence. It will move from physical capacity and precision production to semiconductors, computation, programmable machines, software and networks, data, learning, and the infrastructure required to operate intelligence at scale.
Each essay will examine a particular set of conditions, but the larger subject is the way those conditions interact. The aim is not to produce a list of inventions that eventually led to AI. It is to understand how scientific ideas become engineered systems, how engineered systems become industrial capabilities, and how those capabilities become available for practical use.
Beginning with energy
The first essay turns to a condition that predates electronic computing by centuries: the ability to obtain useful energy, control its conversion, and deliver it where work must be performed.
Energy alone does not explain the emergence of machine intelligence. Its history does, however, provide a starting point for examining how societies expanded the capacity to perform work, developed machinery, organized industrial production, and built physical systems on which later technologies came to depend.
The central question is how the ability to generate, control, and distribute energy changed what machinery and production could accomplish, and what new requirements emerged as those systems expanded.
From there, the series will examine precision manufacturing, semiconductors, computation, programmable machines, software and networks, data, learning, and the infrastructure required to operate intelligence at scale.
A capability that appears immediate at the interface may depend on generations of scientific work, industrial learning, institutional decisions, and physical investment. Understanding that history does not diminish the achievement of the technology. It explains the conditions under which the achievement became possible and what it takes to make it useful.
Machine intelligence is a present-day capability. The systems that made it possible have a much longer history.