Chapter 1

The Wall at the End of Moore's Law

On the 19th of April, 1965, Gordon Moore published a paper in Electronics Magazine that contained, almost as a footnote, an observation that would quietly govern the shape of civilisation for the next six decades.

Moore noticed that the number of transistors on a chip had roughly doubled each year since the first integrated circuit was produced. He predicted this trend would continue. He was right — for fifty years. And the entire architecture of the modern world was built on that reliability. Every smartphone in your pocket. Every cloud server processing your search. Every autonomous vehicle parsing its environment in real time. All of it is downstream of Moore's Law behaving as promised.

What nobody fully prepared for was the day it stopped.

That day did not arrive with a press release. It arrived quietly, in the semiconductor fabs of TSMC, Samsung, and Intel, as engineers discovered that moving to the next node — the next generation of miniaturisation — was consuming an exponentially growing share of research budget for an exponentially shrinking return. The gains were still real. But they were no longer the gains of a covenant. They were the gains of a civilisation reaching the bottom of a barrel it had not previously known was finite.

The Physics of an Ending

Here is the engineering reality that most technology commentary never quite captures: the transistors being built at the frontier of semiconductor fabrication are no longer reliably classical objects. At 2 nanometres — the current bleeding edge, the node at which TSMC is now shipping production chips — engineers are working at the scale of ten silicon atoms placed side by side. A human hair is roughly 80,000 nanometres wide. The features being carved into silicon today are objects that exist at the boundary between classical and quantum mechanics.

At this scale, electrons do not behave the way they do in introductory physics textbooks. Quantum tunnelling — the phenomenon in which a particle passes through a barrier it classically should not be able to penetrate — becomes not an exotic edge case but a dominant engineering constraint. Electrons begin leaking through gate oxides, through insulators, through barriers designed to control them. The transistor, which is fundamentally a switch — a controlled gate that allows or blocks current — begins to blur. The off state is not fully off. The on state is not reliably on.

The industry has responded with increasingly baroque engineering solutions. Gate-all-around transistors wrap the gate material around the channel on all sides, improving electrostatic control. Three-dimensional stacking layers chips vertically to improve memory bandwidth without requiring further planar shrinkage. Chiplets — disaggregated chip designs that package multiple dies in a single package — allow continued performance scaling without requiring every functional block to shrink simultaneously.

These are not failures of imagination. They are the products of extraordinary engineering talent working at the absolute limit of what physics permits. But they are also, unmistakably, workarounds. The elegant covenant of Moore's Law — print smaller, get faster, get cheaper — has been replaced by a complex negotiation between materials science, thermal management, packaging innovation, and economic viability. The era of free improvements is over.

The Energy Wall

There is a second constraint that the public conversation around AI performance rarely addresses adequately: power consumption. Training a large language model of the GPT-4 class requires on the order of tens of gigawatt-hours of electricity — roughly the annual consumption of a small town. The frontier models being trained in 2025 and 2026 are substantially larger. Data centre power consumption globally is growing at rates that grid infrastructure was not designed to accommodate.

This energy constraint is not merely an environmental concern, though it is that. It is also a competitive and strategic constraint. The countries and corporations that can secure low-cost, reliable, high-density power for computational infrastructure will have structural advantages in the AI race. This is already reshaping where data centres are built, which energy companies are valued most highly, and which nations are positioning themselves for computational relevance in the coming decade.

The Parallelism Detour

When individual transistors could no longer be made faster, the industry made a decision so pragmatic and so consequential that its full implications are still unfolding: instead of making chips faster, make them wider. Put more processing units side by side. Let them work simultaneously. This is parallelism, and for the better part of the last fifteen years, it has been the engine powering everything you associate with modern computing performance.

The graphics processing unit tells this story most clearly. The GPU was designed for a narrow problem: rendering the pixels of a video game frame in real time, a task that requires performing the same mathematical operation — transform, light, shade — on millions of independent pixels simultaneously. For this problem, massive parallelism is ideal. A GPU is, architecturally, a device that does many things at once rather than one thing very fast.

The insight that transformed the AI industry — the insight that was, in retrospect, the founding insight of the NVIDIA that now dominates global technology markets — was that training neural networks looks, mathematically, remarkably like rendering graphics. Both are dominated by matrix multiplications: operations that multiply enormous arrays of numbers together. The GPU, built to render Halo, turned out to be nearly ideal for training the systems that would power ChatGPT.

NVIDIA's transformation from a gaming peripherals company to the central infrastructure provider for the most strategically important technology in human history is perhaps the defining corporate story of the 2020s. In 2019, NVIDIA's market capitalisation was roughly $100 billion. By mid-2024, it had crossed $3 trillion, briefly making it the most valuable company in the world. This was not a coincidence of timing. It was the direct consequence of the parallelism strategy reaching full industrial deployment.

But parallelism, too, has limits. Energy consumption scales with core count. A data centre running ten thousand GPUs consumes ten thousand times the power of one. Heat dissipation becomes a physical engineering problem of the first order. And certain classes of problems — the problems that matter most for science, for medicine, for optimisation of truly complex systems — do not parallelise cleanly. They require something fundamentally different.

What Quantum Promises

Classical computers work with bits — discrete units of information that are either zero or one, off or on, the fundamental binary logic of all classical computation. Quantum computers work with qubits: quantum mechanical systems that can exist in superposition — effectively in multiple states simultaneously — until they are measured. The distinction is not merely technical. It is a different relationship between information and reality itself.

Consider a maze. A classical computer solves a maze the way a human does: it tries one path, hits a dead end, backtracks, tries another. Even a very fast classical computer is doing this sequentially, or in parallel — running multiple paths simultaneously on separate cores. A quantum computer, exploiting superposition and entanglement, can, for certain problem structures, explore all paths simultaneously and arrive at the solution through quantum interference — constructive interference amplifying the correct answer, destructive interference cancelling the wrong ones.

The theoretical implications are staggering. A quantum computer with 300 qubits in superposition can represent more simultaneous states than there are atoms in the observable universe. For certain classes of problems — factoring large numbers, simulating molecular interactions, optimising complex logistics networks — this offers computational advantages that no amount of classical hardware can ever replicate, regardless of how many chips you stack or how much electricity you supply.

The promise is extraordinary. The practical reality, as of 2026, is that this promise remains largely theoretical — not because the physics is wrong, but because building a quantum computer that is large enough, stable enough, and accurate enough to demonstrate meaningful advantage over classical systems is an engineering challenge of breathtaking difficulty that has consumed decades of effort from some of the most talented physicists and engineers on Earth.

That challenge is the subject of Chapter 3. But first, we need to understand the intelligence — or rather, the intelligent-seeming systems — that many believe will ultimately solve it.