Moore's Law
Also known as: Moores Law, transistor doubling, semiconductor scaling, computing advancement, chip progress, Dennard scaling
Moore's Law is the empirical observation -- made by Gordon Moore (Intel co-founder) in 1965 -- that the number of transistors on an integrated circuit doubles approximately every two years.
For most of the 20th century, this doubling translated into roughly equivalent improvements in computing performance and decreases in cost per computation. Moore's Law became a self-fulfilling prophecy: the semiconductor industry set roadmaps and invested to ensure the doubling continued. Since approximately 2015, the law has significantly slowed -- transistor miniaturization has become physically harder and more expensive at the smallest scales, and the era of automatic doubling-per-two-years has ended.
Moore's Law, explained at three levels
At home
Moore's Law describes why computers have gotten so much more powerful over decades without proportionally increasing in cost. In 1971, a processor had 2,300 transistors. Today, a processor has over 100 billion -- a 40-million-fold increase in about 50 years. This consistent doubling is why a smartphone today has vastly more computing power than supercomputers from 30 years ago, and why your laptop from 5 years ago feels noticeably slower than today's. The law explains the rhythm of technology improvement that most people experience as the constant cycle of new, faster devices becoming available every few years.
At work
Moore's Law created the technology industry's planning assumption: invest now in technology and it will be significantly more capable at the same cost in 2 years, so current limitations are temporary. This expectation shaped software development (write for today's hardware; tomorrow's hardware will make it fast enough), data center planning (last year's servers will be less capable than next year's), and product roadmaps across tech. Understanding that Moore's Law has slowed means the assumption of automatic, regular improvement no longer holds as reliably -- technology investment decisions need to account for smaller performance gains at higher costs.
At home
Moore's Law vs. Dennard scaling: Dennard scaling (1974) described a related trend -- as transistors shrank, power consumption per transistor decreased proportionally, so performance could increase without increasing total chip power. Dennard scaling effectively ended around 2005-2006 when transistors became so small that leakage current became unmanageable -- shrinking a transistor no longer reduced its power proportionally. The end of Dennard scaling caused the 'power wall': CPU clock speeds stopped increasing (most consumer CPUs still run at 3-5 GHz, roughly where they were in 2005) and the industry shifted to adding more cores rather than faster individual cores.
At work
Software implications of the post-Moore's Law era: when Moore's Law reliably delivered 2x performance every two years, software developers could afford to write inefficient code -- next year's hardware would run it faster. As performance gains slow, software efficiency matters more. This has driven: renewed interest in low-level programming languages (C, C++, Rust) for performance-critical applications; hardware-specific optimization (code paths tuned for specific instruction sets, GPU offloading); the rise of specialized hardware accelerators (AI training on GPUs, TPUs); and greater attention to algorithmic efficiency rather than relying on hardware improvements to compensate.
At home
Moore's Law as an economic phenomenon: the law describes not just transistor density but the entire ecosystem of semiconductor economics. As transistor density doubled, cost per transistor fell roughly proportionally -- enabling computing to become exponentially more affordable. The physical slowing of transistor scaling is matched by an economic challenge: each new process node requires exponentially more investment to develop and build. TSMC's 2nm fab will cost $30-40 billion. As a result, R&D costs are amortized over fewer, larger-volume chips -- increasing industry concentration (fewer companies can afford leading-edge fabs) and increasing chip prices relative to the historical trend.
At work
Post-Moore's Law computing paradigms: the industry is pursuing multiple paths to continue improving computing capability beyond transistor scaling. Domain-specific architectures (GPUs, NPUs, TPUs) do specific computations far more efficiently than general-purpose CPUs. Heterogeneous computing integrates different processor types (CPU + GPU + NPU) in a single package (as in Apple's M-series SoCs). Near-memory and in-memory computing reduce the memory bandwidth bottleneck that limits many modern workloads. Quantum computing offers theoretical exponential speedup for specific problem classes (optimization, simulation) but remains pre-commercial for most applications. Each path addresses different bottlenecks rather than continuing the uniform performance-per-cost doubling that defined the Moore's Law era.
Real-world examples
At home
Fatima's grandmother showed her a 1995 family computer that cost $2,000 and ran at 66 MHz with 8MB of RAM. Fatima's $800 smartphone has a 3 GHz processor (45x faster clock) with 8GB of RAM (1,000x more) and performs billions of times more computations per second. Moore's Law explains this gap: over 29 years of roughly two-year doubling cycles, computing capability grew by roughly 2^14 = 16,000x while costs remained similar or fell. The smartphone in her pocket would have been the world's most powerful supercomputer as recently as the mid-1990s.
At work
Jerome was building a business case for infrastructure investment. His CFO expected servers to be significantly cheaper and more capable in two years -- a planning assumption based on decades of Moore's Law expectations. Jerome's analysis of recent trends showed that per-core performance improvements had slowed to roughly 10-15% per year (vs. the historical doubling every 2 years), while chip prices were rising due to higher fabrication costs. He adjusted the infrastructure refresh model to assume more modest performance improvement per generation and longer server refresh cycles, producing a more realistic long-term capex projection.
Common confusions
- Many people believe Moore's Law predicts that computers double in speed every two years — Moore's Law specifically predicts that transistor count doubles approximately every two years -- a statement about chip density, not directly about speed. Transistor count increases have historically correlated with speed improvements, but the relationship is indirect and dependent on how additional transistors are used (more cores vs. faster single core vs. larger cache vs. specialized units). Since approximately 2005, single-core CPU speed has plateaued (from the end of Dennard scaling) even as transistor count continued to grow. Today, more transistors primarily yield more cores, specialized accelerators, and efficiency improvements rather than proportionally faster single-threaded performance.
- Some assume Moore's Law is a physical law of nature rather than an observation about industrial trends — Moore's Law is an empirical observation about the pace of human engineering achievement in semiconductor manufacturing -- not a law of physics. It has held for roughly 50 years because the industry organized itself (through investment, standardization via ITRS roadmaps, and shared R&D) to achieve the predicted doubling. It could in principle have ended much earlier if the engineering challenges had proven insurmountable. The current slowing is not a surprise to physicists -- the approaching limits of silicon transistor scaling have been predicted for decades. What's remarkable is how long human ingenuity continued the trend against physical challenges.
For analysts
Moore's Law is a useful framework for discussing technology planning assumptions with clients. The key advisory point is that the era of reliable, automatic performance doubling every two years has ended for general-purpose computing -- and business planning that assumes otherwise will produce over-optimistic IT economics projections. However, Moore's Law-like progress continues in specialized domains (AI accelerators, memory bandwidth, storage density) and through architectural innovation. Technology refresh cycles, software optimization investment, and hardware selection decisions all need to account for the more differentiated, workload-specific improvement landscape of the post-Moore's Law era.
History & origin
Gordon Moore published his observation in a 1965 paper titled 'Cramming more components onto integrated circuits.' His original formulation described doubling every year; he later revised it to every two years. Intel executive David House popularized the associated claim that chip performance doubled every 18 months (combining transistor density doubling with faster clock speeds). Moore's Law became the organizing principle for the entire semiconductor industry's roadmap -- companies planned product releases around it; customers planned technology upgrades around it. The observation held remarkably well from 1965 to approximately 2015. Since then, physical limits (quantum effects, heat density, manufacturing cost) have slowed the doubling. The industry continues advancing through architectural innovation (chiplets, 3D integration), specialized accelerators (GPUs, TPUs), and advanced packaging -- but the simple 'same cost, twice the transistors every 2 years' dynamic that characterized the first 50 years has ended.