What Happens When Moore's Law Stops Working

Why Moore's Law Was a Miracle

In 1965, Intel co-founder Gordon Moore observed that the number of transistors on a semiconductor chip doubled roughly every two years while costs per transistor fell. This observation—Moore's Law—drove 60 years of exponential computing improvements. Computing power increased 2 billion-fold since 1960, with corresponding cost decreases that made computers accessible globally.

Remarkably, Moore's Law held for 6 decades despite being a prediction, not a law of physics. It became a self-fulfilling prophecy: the industry organized its roadmaps around achieving Moore's Law targets, and remarkably, usually succeeded.

Physical Limits and Quantum Physics

The Atomic Barrier: Transistors are shrinking toward atomic dimensions. Modern cutting-edge chips operate at 2-3 nanometers (2-3 billionths of a meter). For context: a silicon atom is 0.2 nanometers. We're essentially at the atomic limit.

Below 2 nanometers, several physical phenomena prevent further shrinking:

Quantum Tunneling: At sub-nanometer scales, electrons don't reliably stay where you force them. They can "tunnel" through barriers that should block them, like a ghost walking through a wall. This causes transistor gates to leak current uncontrollably, generating heat and wasting power.

Gate Control Challenges: A transistor's "gate" (the part controlling electron flow) needs to maintain precise control. As gates shrink below 2nm, maintaining this control becomes impossible—electron behavior becomes probabilistic rather than deterministic.

Current Leakage: Current flowing where it shouldn't causes exponential power waste. Power density (watts per square millimeter) increases faster than transistor density as you shrink, meaning more transistors generate more heat in the same area.

Economic Barriers: The Cost Explosion

Beyond physics, economics makes further scaling unaffordable.

Fab Costs: A modern semiconductor fabrication plant costs $20+ billion and takes 4-5 years to build. Only 3 companies globally (TSMC, Intel, Samsung) can afford this investment. Only TSMC can afford the most advanced nodes.

Machinery: Extreme ultraviolet (EUV) lithography machines used to etch 2nm patterns cost $200+ million each. A single fab requires dozens of these machines. The machinery is so expensive that some researchers argue the industry has already reached economically practical limits even if physics allowed further scaling.

Yields: Defects increase exponentially as features shrink. Maintaining acceptable yields (usable chips per wafer) becomes nearly impossible at extreme scales. Companies spend enormous engineering effort improving yields—2-3% yield improvements are celebrated.

The End of Dennard Scaling

For decades, Moore's Law was accompanied by Dennard Scaling: as transistors shrank, they operated at lower voltages and consumed proportionally less power.

This ended around 2005-2007. Transistors no longer get proportionally cooler as they shrink. Instead, power density (power per square millimeter) increases, meaning you can't actually pack more transistors without thermal runaway (overheating).

This explains why CPU clock speeds stalled around 3-4 GHz despite transistor improvements continuing. Adding more transistors no longer increases clock speed if you can't remove the heat they generate.

What Happens Now: Post-Moore Strategies

Chiplet Architecture: Instead of maximizing transistors on a single die, companies divide chips into smaller "chiplets" from different process nodes, then interconnect them. This enables flexible optimization: expensive 2nm logic mixed with cheaper 7nm memory and 10nm I/O components.

3D Stacking: Vertically stacking dies increases density without shrinking further. A 10-layer stack provides 10× the compute in similar area to a flat chip. Heat extraction becomes the limiting factor, solvable through engineering rather than physics limits.

2D Materials: Graphene and transition metal dichalcogenides (like MoS₂) may extend transistor scaling another generation. These materials have atomic thickness, enabling gate lengths below 1nm while maintaining better control than silicon.

Heterogeneous Integration: Combining logic, memory, analog circuits, and photonics on different nodes in single packages enables targeted optimization. Not all components need to be at the most advanced node.

System-Level Optimization: The post-Moore era focuses on designing entire systems (architecture, memory, network) around data flow rather than transistor density. Performance comes from intelligent data movement and specialized processors rather than brute-force transistor scaling.

The Real Implication: Slowdown, Not Stop

Moore's Law isn't completely dead—it's just slowing exponentially.

1995-2005: Transistor doublings every 1.5-2 years; clock speeds increased 10× (100MHz to 3GHz)
2005-2015: Transistor doublings every 2-3 years; clock speeds stalled (still 3-4GHz)
2015-2025: Transistor doublings every 3-5 years; density gains from 3D/chiplets, not shrinking
2025+: Estimated 10+ years between major node transitions; progress through specialized architectures

What This Means for Computing

Performance improvements will come from architecture, not transistors:

Specialized accelerators for AI, graphics, cryptography
Multi-core optimization instead of clock speed increases
Software and hardware co-optimization rather than raw compute increases

Power efficiency becomes the limiting factor:

Data movement energy exceeds computation energy
Cooling costs exceed manufacturing costs for advanced nodes
Energy becomes the binding constraint on compute density, not physics

Specialization replaces generalization:

Chips optimized for specific workloads (AI, video processing, finance) beat general-purpose processors
Companies building custom chips (Google TPUs, Amazon Trainium, Microsoft Maia) gain advantages
Heterogeneous computing (multiple specialized processors in same device) becomes standard

Economic Restructuring

TSMC's monopoly: TSMC manufactures 99% of advanced AI accelerators and 90%+ of advanced logic chips. This geopolitical dependency has prompted governments to fund alternative manufacturers.

Increased competition: As transistor scaling becomes less important, companies can compete on architecture and software rather than manufacturing prowess. AMD's recent gains reflect this—they compete on design, not process node.

Regional chip ecosystems: The US (CHIPS Act: $52B), EU ($43B), and others are funding domestic fabs. This won't match TSMC's efficiency but provides supply chain resilience.

Common Myths

Myth 1: "Moore's Law is dead; computing advancement will stop"

Reality: Moore's Law is ending as stated, but computing advancement continues through other mechanisms: specialization, 3D integration, architectural optimization, and heterogeneous systems.

Myth 2: "We've hit physics limits immediately; nothing further is possible"

Reality: Physics limits exist, but engineering solutions exist within those limits. 2D materials, 3D stacking, chiplet architectures, and other techniques provide another 10-20 years of continued optimization before fundamental barriers force radical rethinking.

Myth 3: "Moore's Law slowdown means computers will stop improving"

Reality: Slowdown ≠ stop. Performance improvements will continue at 5-10% annually—substantial over decades—but not the 50-100% annual improvements of the Moore's Law era.

Why It's Happening Now

Physical limits are hitting simultaneously with economic limits. The cost of advancing to next nodes (5nm → 3nm → 2nm) has become unjustifiable for all but the most demanding applications. Companies are rationally choosing to optimize within current nodes rather than fund $20+ billion investments in new fabs.

Long-Term Implications

Specialization wins: General-purpose processors may become commoditized while specialized accelerators command premiums.

Software becomes critical: With hardware improvements slowing, software efficiency, algorithms, and architecture matter more than raw processor speed.

Energy becomes the economy: Power consumption and cooling costs may dominate chip costs by 2030, shifting focus from performance metrics to energy metrics.

Geopolitical implications: Chip manufacturing capacity becomes strategic national infrastructure like oil reserves in the 20th century.

Conclusion

Moore's Law wasn't a law—it was a miracle of engineering and industrial coordination lasting 60 years. That miracle is ending as physics and economics converge. Computing won't stop improving, but improvement sources will shift from transistor scaling to architectural innovation, specialization, and system-level optimization—a transition that's fundamentally reshaping the semiconductor industry and its global supply chains.