14 · The Frontier

This chapter is not a floor. It is the scaffolding on the roof and the cracks in the foundation — the places where the concrete is still wet, and where a reader who wants to build something has room to stand.

Which walls of the tower are still wet concrete?

In May 1981, at MIT's Endicott House, Richard Feynman raised a problem computing had no answer for. Physicists wanted to simulate quantum systems, and the cost on an ordinary computer grows exponentially with the number of particles: a few dozen interacting particles exhaust any machine that will ever be built. His conclusion, published the following year, was not that the simulation should be made cleverer but that the computer was the wrong kind of object — nature is not classical, so a machine that simulates nature should be built out of quantum mechanics rather than in spite of it (International Journal of Theoretical Physics 21, 1982).

Forty-five years on it is the most over-promised item on this list. Start with what it buys.

Quantum computers are specialists, and the specialism is narrow. Peter Shor showed in 1994 that one could factor integers and compute discrete logarithms efficiently — an exponential speedup, and the reason chapter 11 has a deadline (Shor). Lov Grover showed in 1996 that unstructured search gets faster too, but only quadratically: √N queries where a classical machine needs N, and that bound is provably the best possible. The difference between exponential and quadratic is the whole story. A quadratic speedup can be erased by a decade of ordinary hardware progress; an exponential one cannot be erased at all. And the class of problems most people assume is the target — NP-complete problems like scheduling, routing and packing — is not believed to fall. There is no proof either way, but there is an oracle relative to which they do not, and no candidate algorithm (Aaronson). A quantum computer is not a faster computer. It is a different instrument, good at simulating quantum systems and at breaking two specific pieces of mathematics we happened to build civilization's trust on.

The engineering problem is not qubits; it is error. A qubit holds its state for microseconds before the environment ruins it, so useful machines must correct errors faster than they accumulate — spreading one logical qubit across many physical ones. The threshold theorem says this works only if the physical error rate falls below a critical value; above it, adding qubits makes things worse. Google's Willow processor demonstrated the crossing in a result published online in December 2024. Each increase of the surface-code distance by two suppressed the logical error rate by a factor of 2.14 ± 0.02, and a distance-7 code on 101 qubits reached 0.143% ± 0.003% error per cycle — below threshold, meaning bigger now means better (Nature 638: 920–926). That does not mean a useful machine exists. The best current public estimate for breaking RSA-2048 is under a million noisy physical qubits running for about a week at a 0.1% gate error rate — an architectural calculation on paper, not a device anyone has built (Gidney, 2025). Read every quantum announcement against that gap.

Simulation became science's third instrument, and its leaks are epistemic. In June 2005 the US President's Information Technology Advisory Committee put it on the record: "Computational science now constitutes what many call the third pillar of the scientific enterprise, a peer alongside theory and physical experimentation" (PITAC). Climate, protein folding, drug binding, galaxy formation, crash safety — these are now studied primarily inside machines. This is an abstraction like any other in this book, with one uncomfortable difference. When a normal abstraction leaks, something breaks and you descend the stairs to find out why. When a simulation leaks, the output is still a plausible number, and nothing tells you from the inside whether a disagreement with the world is the world's fault or the model's. The discipline that grew up to handle this — verification (did we solve the equations right?) and validation (did we write the right equations?) — is the most transferable idea in this chapter and the least famous.

The binding constraint is now energy. For thirty years, Dennard scaling meant smaller transistors used proportionally less power, so every generation was free. It broke down around 2005–2007, when leakage current made heat the limit, and chapter 04 has lived with the consequences ever since. Efficiency still improves, but slower: the doubling time of computations per joule has stretched from roughly 1.6 years before 2000 to nearer 2.5 years since (Koomey's law). There is a floor underneath all of it. Rolf Landauer showed in 1961 that erasing one bit must dissipate at least kT ln 2 of energy — about 2.8 × 10⁻²¹ joules at room temperature — the point in this book where information stops being a metaphor and becomes thermodynamics, the other half of chapter 03's story. Real processors sit between a million and a billion times above that floor, depending how you count, so the physics is not the problem yet. The bill is. The International Energy Agency reported on 16 April 2026 that data-center electricity demand grew 17% during 2025, with AI-focused facilities climbing faster still, and projects total data-center consumption roughly doubling between 2025 and 2030 (IEA). For the first time in the tower's history, the question limiting what gets computed is not what is computable, or what is fast, but what can be powered.

Several walls are unfinished, and the invitations are open. Whether P equals NP carries a million-dollar prize offered by the Clay Mathematics Institute on 24 May 2000, and remains unclaimed (Clay) — not a puzzle for specialists but the question of whether checking an answer is fundamentally easier than finding one, which is to say whether creativity is real or merely an unoptimized search. Whether one-way functions exist is chapter 11's foundation and equally unproven. Neuromorphic computing — machines organized like nervous systems rather than clocked pipelines — has produced Intel's Hala Point, 1.15 billion neurons at Sandia National Laboratories in April 2024 (Intel), with no confirmed larger system since. Reversible computing, which attacks the Landauer floor by not erasing bits in the first place, has moved from theory to silicon in a 2025 prototype recovering roughly half its switching energy (IEEE Spectrum). And room-temperature superconductivity, which would change every number above, has not arrived: the 2023 LK-99 signal was traced to a copper-sulfide impurity undergoing a structural phase transition, and no verified ambient-pressure room-temperature superconductor exists (phys.org).

None of these are conclusions. They are the parts of the tower where the plans are still argued over, which is the only part of any building where a newcomer's hands are genuinely needed. The floors below took roughly eighty years to pour. Nothing about them suggests the pouring is finished.

Connections

Below: chapter 01 owns P vs NP as a question about what is computable at what cost; this chapter only reports that it is still open. Chapter 03 supplies the bit, which Landauer's principle reveals to have a price in joules. Chapter 04 is where the energy wall is felt first, in the turn to specialized hardware. Chapter 11 is where quantum computing stops being interesting and becomes urgent, and chapter 12 is the largest current consumer of everything this chapter says is scarce. Above: nothing yet — that is what a frontier is.

Open questions

Will a cryptographically relevant quantum computer arrive before the world finishes migrating away from the mathematics it would break, and how would anyone know in advance? If simulation is the third pillar of science, what is the equivalent of peer review for a model nobody can independently rebuild? Can the energy curve of computation bend without a physical breakthrough, or does the industry's growth now depend on one? And, quietly the largest: is P vs NP hard because we lack the technique, or because the techniques we have are provably incapable of settling it?

Go deeper

• Michael Nielsen, Quantum Computing for the Very Curious, 2019 — a free, unusually honest introduction that assumes no physics, at quantum.country/qcvc.
• Clay Mathematics Institute, The Millennium Prize Problems — the standing list, including P vs NP, at claymath.org.
• Rajeev Acharya et al., "Quantum error correction below the surface code threshold," Nature 638 (2024): 920–926 — nature.com.