Quick read · plain language

What Comes After Silicon Computers

A plain-language version of “Computing After Silicon: A History-Constrained Forecast of Computing Machine Evolution, 2026-2050”

3 min readNo jargon, no mathematicsAssessments

The short version

This paper looks at how computers have changed many times before. It guesses what computers will look like from now until 2050. The main guess is that no single new kind of computer will replace silicon. Instead, many special kinds will work together.

How computers have always changed

Every era of computers hit a hard wall. Mechanical gears could not get more precise. Relays broke too often. Vacuum tubes used too much power and burned out. Transistors solved the heat problem, but then they too hit limits.

Each time, the same thing happened. First, the old way ran into a physical limit. Then, a new way offered much more room to grow. Making more of the new thing made it cheaper. Finally, everyone switched to the new way. This same pattern shows up in physics, biology, and how societies adopt new tools.

Seven new kinds of computers

The paper looks at seven possible next steps. The authors rank them by how likely they are and how big an impact they would make.

The most likely near-term winner is special chips built just for AI work. These are already taking over from general-purpose chips. The main risk is whether software can keep up with the new hardware.

The biggest possible impact would come from quantum computers. But these face the hardest path. The problem is not how many quantum bits you have. The problem is fixing errors. The paper compares this to early classical computers. It took about two decades to make error correction work well there.

Processing inside memory chips is the safest bet. Moving data back and forth costs the most energy in modern computers. Doing work right where data lives avoids this waste.

Other ideas include brain-like chips, light-based computing, reversible computing, and DNA computing. Each has physics merit. None is close to market like the top three.

What the forecast rests on

The ranking holds up even if you change the assumptions. The only close call is whether quantum or memory-processing comes second. That depends on whether quantum error correction stalls.

The paper also tests itself against five hard challenges. One challenge says silicon could continue with 3D stacking. The authors accept this partly. It might buy five years, but not stop the larger shift.

Paths not taken

The paper imagines four past choices that could have changed everything.

If quantum error correction had gotten massive funding in the 2000s, useful quantum computers might exist already. The field focused on making quantum bits instead of fixing their errors.

If a 1961 physics limit had guided design from the start, we might have extremely efficient computers today. The gap between that limit and current practice is huge.

If thermodynamics and information theory had joined a century earlier, computers might have been designed as physical processes from day one.

A non-Turing model of computing would change everything. No such model is known to be physically possible. This serves as a boundary check.

Twelve bets for the future

The paper makes twelve dated predictions. Strong ones rest on historical patterns. Weak ones are more speculative.

Strong bets include: by 2030, AI chips will make up most datacenter spending. By 2035, the old way of making silicon chips will effectively stop improving. By 2040, computers will do far more work per unit of energy.

Weak bets include: by 2032, AI training tools will work across any hardware. By 2040, reversible computing will show a big efficiency demo.

Each prediction can be proven wrong. The paper says exactly how.

What this does not show

The paper does not prove that any one new substrate will win. It argues the opposite. The general-purpose computer fading does not mean one replacement takes over.

The probability ranges are guesses anchored to history, not hard numbers. The authors admit this. The forecast fails if CMOS scaling does not end by 2035, or if AI chips do not dominate datacenter spending by 2030.

The paper does not address who will pay for these shifts. It argues that the mixed path needs no single massive project.

The quantum timeline remains highly uncertain. Error correction could stall. The classical analogy might not hold.

Words used here

substrate: the physical material or base that a computer is built from, like gears or silicon chips
heterogeneous: made of many different kinds mixed together
falsifiable: able to be proven wrong by future events
backcast: imagining how the past could have gone differently to learn about the future

Learn more in plain language

Moore's lawSimple English Wikipedia
Quantum computerSimple English Wikipedia
Integrated circuitSimple English Wikipedia
Error detection and correctionSimple English Wikipedia
Homogeneity and heterogeneityWikipedia (English, more technical)
FalsifiabilityWikipedia (English, more technical)

This plain-language version was written by an AI model from the text of the paper and checked against plain-language rules: everyday words, no mathematics, no unexplained names. It leaves details out and can still be wrong; the paper is the source of record. To report a problem, open the paper, select the passage and choose Report.

For specialists: the technical digest, quoted from the paper

In the paper's words

A history-constrained forecast of computing machine evolution 2026-2050. Seven post-silicon candidates assessed; heterogeneous convergence (AI accelerators, quantum co-processors, processing-in-memory) is the central thesis; twelve dated falsifiable predictions registered for 2030-2040.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

Assumption Candidate Confidence Pillar A1: Accelerator software ecosystems mature within a decade AI-Heterogeneous 0.75 Empirical Base Rate (CUDA history) A2: Quantum error correction achieves logical qubits at scale by 2035 Quantum 0.35 Reference Class (classical ECC timeline) A3: Memory-bandwidth wall persists (no disruptive memory technology) PIM 0.70 Empirical Base Rate A4: Neuromorphic algorithms reach parity on real workloads Neuromorphic 0.30 Calibrated Subjective A5: Photonic logic integration overcomes device challenges Optical 0.20 Calibrated Subjective A6: Adiabatic control overhead falls 100× Reversible 0.15 Calibrated Subjective
from 4.1 Enabling Assumptions
Candidate Allocation Rationale AI-Specialized Heterogeneous 30% Already dominant; focus on software/programmability Quantum Utility 25% Highest impact (9/10) but high uncertainty Processing-in-Memory 20% Lowest risk, engineering-driven near-term win Neuromorphic 10% Digital spiking path avoids materials risk Optical/Photonic 5% Interconnects first; logic is 2035+ Hedge (Unknown) 10% Anti-fragility floor — unlisted paradigm will emerge
from 4.4 Research Effort Allocation
The Landauer limit is $10^{6}\times$ below current computing efficiency; even approaching it by 100× would transform mobile, edge, and datacenter computing.
from Tier 2: Reversible Computing from Inception (1960s fork)
Quantum error correction at 5× current funding — the classical analogy suggests error correction, not qubit count, is the binding constraint.
from Near-Term Fork Recommendations
AI-specialized architectures as the new processor baseline — every general-purpose CPU design from 2026 onward should assume >50% die area for domain-specific accelerators.
from Near-Term Fork Recommendations
If CMOS scaling does not effectively end by 2035, or if AI accelerators do not exceed 50% of datacenter compute spending by 2030, the central thesis is disconfirmed.
from 8. Conclusion

How it is done

No section of this paper is shaped like a method, so no method sentence is quoted.

What would falsify it

This paper synthesizes the history of computing paradigm shifts through a cross-domain lens spanning physics, information theory, economics, and biology, and produces a structured, falsifiable forecast of computing machine evolution through 2050.
from Abstract
Third, what falsifiable predictions follow, and how would each be disconfirmed?
from 1. Introduction
Section 6 registers the falsifiable predictions and practical applications.
from 1. Introduction

Where the claim stops

If CMOS scaling does not effectively end by 2035, or if AI accelerators do not exceed 50% of datacenter compute spending by 2030, the central thesis is disconfirmed.
from 8. Conclusion
We assess seven paradigm-shift candidates qualitatively, identifying AI-specialized heterogeneous computing as the highest-probability near-term trajectory and fault-tolerant quantum computing as the highest-impact but most uncertain candidate.
from Abstract
Section 4 presents the forecast, including assumption audits and sensitivity analysis.
from 1. Introduction

Where to read

1. Introduction205 words · 1 min
2. The Constraint-Saturation History of Computing324 words · 1 min
3. The Post-Silicon Candidate Landscape341 words · 1 min
4. The Forecast: Assumptions, Sensitivity, and Effort Allocation276 words · 1 min
5. Counterfactual Backcasting256 words · 1 min
6. Calibration Register: Twelve Dated, Falsifiable Predictions123 words · 1 min
7. Practical Applications48 words · 1 min
8. Conclusion102 words · 1 min
Declarations115 words · 1 min

Read from the paper itself: 30 sections, 2,117 words. Nothing in this digest is written by a model; a line that is not a sentence of the paper is not shown. Have a specific question? Ask this paper.

The full text · Ask the whole corpus