Huang's Law vs Moore's Law

Different Engines, Different Eras

Moore's Law and Huang's Law are often presented as sequential — Moore's Law carried computing for fifty years, and now Huang's Law picks up where it left off. The reality is more nuanced. Moore's Law was an observation about manufacturing: as lithography improved, you could etch more transistors per square millimeter, and each generation delivered more compute essentially for free. It was a rising tide that lifted all boats — CPUs, GPUs, memory, embedded processors, and smartphones all rode the same curve.

Huang's Law is an observation about specialization. GPU performance gains come from designing silicon specifically for the mathematical operations that AI requires — matrix multiplications, convolutions, attention mechanisms — and then co-optimizing the software stack to exploit every available transistor for that narrow purpose. This means Huang's Law doesn't replace Moore's Law; it rides on top of it while adding architectural and software multipliers that general-purpose CPUs cannot match.

The Numbers: Claimed vs. Measured

NVIDIA's marketing claims are dramatic: 1,000× improvement in AI inference over a decade, and Jensen Huang stated at CES 2025 that the company's systems are progressing at "hyper Moore's Law" rates. The jump from A100 (2020) to H100 (2022) to Blackwell B200 (2024) did deliver roughly 30× improvement in large language model inference throughput. Looking ahead, NVIDIA's roadmap shows Rubin NVL144 (2026) offering 3.6 exaFLOPS of dense FP4 compute versus Blackwell Ultra's 1.1 exaFLOPS — a 3.3× improvement in a single generation.

Independent measurement tells a more moderate story. Research by EpochAI found that GPU price-performance (FLOPS per dollar) has doubled approximately every 2.5 years between 2006 and 2025 — faster than Moore's Law's recent pace but well short of the "doubling every year" narrative. The gap reflects the difference between measuring raw peak throughput (which benefits from format changes like FP4) versus measuring useful work per dollar across real workloads. Both framings matter: peak throughput determines what's possible; cost-performance determines what's economical.

Why the Divergence Matters for AI Economics

The practical consequence of Huang's Law outpacing Moore's Law is that the cost of AI inference is falling far faster than the cost of general-purpose compute. This creates the conditions for Jevons' Paradox to operate at full force: as each GPU generation halves or thirds the cost of running a large language model query, total demand for AI inference doesn't merely double — it explodes. Enterprise adoption accelerates, new use cases become viable, and agents that were too expensive to run continuously become economically rational.

This flywheel connects directly to Wright's Law: as NVIDIA and TSMC produce more GPUs and wafers, manufacturing learning curves drive unit costs down further, which stimulates more demand, which finances the next round of R&D. The result is a self-reinforcing cycle where Huang's Law (performance), Wright's Law (cost), and Jevons' Paradox (demand) compound each other — a dynamic that has no precedent in the Moore's Law era because CPU performance gains were never steep enough to trigger this kind of demand explosion.

The Manufacturing Foundation Still Matters

Huang's Law depends on Moore's Law continuing in some form. NVIDIA's Rubin GPUs (2026) will be manufactured on TSMC's 3nm process; future generations will move to 2nm. Each process node improvement delivers 10–15% more transistors per unit area and meaningful power efficiency gains — improvements that Huang's Law then multiplies through architectural innovation. If transistor scaling stopped entirely, Huang's Law would slow dramatically, though it wouldn't stop: architectural innovation, numerical format changes, and software optimization can still deliver meaningful gains without smaller transistors.

The economics of advanced manufacturing increasingly favor this co-evolution. TSMC's 2nm node entered mass production in late 2025 with capacity already fully booked for 2026, and Intel's 18A process is ramping at competitive yields of 55–70%. But leading-edge wafer costs now exceed $20,000, which means only products with extremely high value per transistor — like AI accelerators selling for $30,000–$40,000 per chip — can justify the investment. Moore's Law's economics now serve Huang's Law's market.

Concentration Risk vs. Distributed Innovation

One critical difference between the two laws is market structure. Moore's Law was a broadly shared phenomenon: hundreds of chip companies rode the same lithography improvements, and competition kept prices low and innovation distributed. Huang's Law, by contrast, is overwhelmingly concentrated in a single company. NVIDIA controls roughly 80–90% of the AI accelerator market, and the CUDA software ecosystem creates powerful lock-in. AMD's MI300X and Google's TPU v5 offer competitive hardware, but neither has matched NVIDIA's full-stack integration of hardware, software, networking, and developer ecosystem.

This concentration creates both opportunity and fragility. The opportunity is that a single company with massive R&D budgets ($12B+ annually) and a clear roadmap can push the performance curve faster than a fragmented market. The fragility is that any disruption to NVIDIA's execution — supply chain issues, architectural missteps, or a breakthrough from competitors — could slow the pace of improvement that the entire AI industry now depends on.

The Road Ahead: 2026–2030

NVIDIA's published roadmap extends through 2028: Blackwell Ultra (2025), Rubin (2026), Rubin Ultra (2027), and Feynman (2028), with each generation targeting 2–4× improvement in AI workload performance. If this cadence holds, total inference performance per dollar will improve roughly 100× between 2024 and 2028 — a pace that would make today's most expensive AI applications trivially cheap.

Moore's Law, meanwhile, has a clear if slower path forward. Gate-all-around (GAA) nanosheet transistors at 2nm, backside power delivery at sub-2nm nodes, and 3D chip stacking (TSMC SoIC, Intel Foveros) will continue delivering transistor density improvements through the end of the decade, albeit at 18–24 month cadences rather than the original 24-month doubling. The two laws will continue to co-evolve: Moore's Law providing the transistor budget, Huang's Law spending that budget far more efficiently than general-purpose designs ever could.