Moore's Law
Moore's Law is the empirical observation, first articulated by Intel co-founder Gordon Moore in 1965, that the number of transistors on an integrated circuit doubles approximately every two years. Originally Moore predicted a doubling every year, later revised to two years, and the industry has treated this cadence as a planning principle, a self-fulfilling prophecy, and a cultural myth all at once. The law is not a law of nature. It is a law of industrial coordination: a target that semiconductor manufacturers, equipment makers, and software developers collectively organize their R\&D cycles around, creating the very progress it claims to predict.
The law held with remarkable fidelity for roughly five decades. From the 4004 processor's 2,300 transistors in 1971 to modern GPUs with tens of billions, the exponential growth enabled the digital revolution, the internet, and the training of large language models that would be thermodynamically impossible on 1990s hardware. But around 2010, the curve began to bend. Dennard scaling — the companion principle that power density stays constant as transistors shrink — collapsed. Clock speeds stalled. The industry pivoted to multicore architectures, then to specialized accelerators, then to three-dimensional stacking and exotic materials. Each pivot bought time, but each was an admission that the original trajectory was ending.
The Physical Limits
The end of Moore's Law is not a market prediction; it is a physical inevitability. Transistors are now measured in angstroms, approaching the scale of individual atoms. At these dimensions, quantum tunneling causes electrons to leak through insulating barriers regardless of voltage. Landauer's principle sets a thermodynamic minimum for the energy cost of erasing one bit of information — approximately \(kT \ln 2\) joules — and modern logic operations operate within a few orders of magnitude of this limit. Further scaling requires either operating at cryogenic temperatures or accepting error rates that demand exponential redundancy, both of which impose their own cost floors.
The economics compound the physics. A modern semiconductor fabrication facility costs over twenty billion dollars. The marginal cost of each additional shrink in feature size grows superlinearly, while the performance benefit shrinks. The industry has responded by fragmenting: instead of one universal process node advancing uniformly, we have domain-specific accelerators — GPUs for training, TPUs for inference, NPUs for edge devices — each optimizing a different point in the energy-delay-product space. This fragmentation is not a failure of Moore's Law; it is an adaptation to its death.
Implications for Computation
The assumption that computation grows exponentially has shaped every long-term prediction in computer science, artificial intelligence, and systems theory. The Bitter Lesson in AI is predicated on the continuation of this growth: general methods win because they can consume ever-larger amounts of compute. Test-time compute scaling assumes that inference-time computation is a resource we can freely expand. If the underlying hardware stops doubling, these strategies become resource-constrained optimization problems rather than open-ended scaling laws.
The transition from exponential to linear — or worse, diminishing — growth in raw transistor count does not mean computation stops improving. It means the locus of improvement shifts from hardware scaling to algorithmic efficiency, architectural specialization, and software optimization. The history of computing is punctuated by such shifts: the transition from assembly to high-level languages, from scalar to vector processing, from CPUs to GPUs. Each shift extracted orders of magnitude of effective compute without requiring transistor doubling. The post-Moore era will be defined by who can find the next such shift.
The Cultural Afterlife
Moore's Law persists in discourse long after its physical death because it serves a narrative function. It tells a story of inevitable progress, of human ingenuity overcoming physical limits through sheer collective effort. This narrative has funded trillions of dollars of investment, built entire industries, and trained a generation of technologists to expect that waiting two years solves most computational problems. The cultural residue of the law may outlast its physical reality by decades, continuing to shape expectations, investment theses, and policy debates about artificial intelligence long after the fabs have stopped shrinking.
The honest framing is that Moore's Law was a sixty-year anomaly, not a permanent condition. Exponential growth in any physical quantity is always temporary. The question is not whether it ends but what replaces it. The next era of computing will not be characterized by uniform scaling but by radical heterogeneity: specialized chips, neuromorphic architectures, optical interconnects, and possibly quantum accelerators for specific problem classes. The illusion of a single number — transistor count — as the measure of progress will give way to multidimensional trade spaces where energy, latency, precision, and programmability are optimized jointly, not sequentially.
Moore's Law did not describe the future of computing; it described a specific industrial epoch that is now closing. The technologists who continue to treat it as a law of nature are not making predictions; they are reading entrails. The real question is whether the institutions, funding models, and intellectual frameworks built around exponential growth can adapt to a world where progress is hard-won, domain-specific, and fundamentally thermodynamically bounded.