Huang's Law vs Wright's Law

Comparison

Huang's Law and Wright's Law are two of the most important frameworks for understanding why artificial intelligence is becoming simultaneously more powerful and less expensive. Huang's Law describes the supply-side acceleration: GPU performance for AI workloads doubling every one to two years through architectural innovation, far outpacing Moore's Law. Wright's Law describes the demand-side feedback loop: for every cumulative doubling of production, costs fall by a constant percentage — typically 20–30%. Together, they form a compounding flywheel that is driving AI inference costs down by orders of magnitude and expanding what is economically feasible in the agentic economy.

Feature Comparison

DimensionHuang's LawWright's Law
OriginJensen Huang, NVIDIA GTC 2018Theodore Wright, 1936 study of aircraft manufacturing
Core ClaimGPU AI performance doubles roughly every 1–2 yearsCosts decline ~20–30% for every cumulative doubling of production
Independent VariableTime (calendar years between GPU generations)Cumulative production volume (units built)
MechanismArchitectural innovation: tensor cores, memory bandwidth, numerical formats (FP8/FP4), software co-optimizationLearning-by-doing: yield improvements, process innovation, tooling optimization, design simplification
ScopeNarrow — GPU/accelerator performance for AI workloads specificallyBroad — applies across semiconductors, solar, batteries, genome sequencing, AI inference
Predictive ReliabilityStrong for NVIDIA's roadmap (Blackwell → Rubin → Feynman); less validated outside NVIDIASanta Fe Institute found it outperformed Moore's Law across 62 technology domains
Recent EvidenceRubin (2026) delivers 5× inference FLOPS over Blackwell; 1,000× improvement over a decadeGPT-3.5-level inference costs dropped 280× between Nov 2022 and Oct 2024; AI hardware costs decline ~37.5% per cumulative doubling
DependencyRequires continued R&D investment, TSMC process advances, and architectural breakthroughsRequires sustained production volume growth — stalled adoption breaks the curve
Failure ModePhysics limits (power walls, interconnect bottlenecks), slowing architectural gainsDemand saturation, supply chain disruptions, or geopolitical constraints on production
Strategic ImplicationInvest in GPU-native architectures; plan for next-gen hardware capabilitiesScale production aggressively to ride the cost curve down ahead of competitors
Relationship to Jevons' ParadoxEach GPU generation halves the cost of intelligence, stimulating super-elastic demandFalling production costs expand the addressable market, which drives more production and further cost declines
Primary BeneficiaryHardware vendors (NVIDIA), hyperscalers, AI infrastructure buildersHigh-volume deployers: cloud platforms, AI-native applications, platform ecosystems

Detailed Analysis

Two Sides of the Same Flywheel

Huang's Law and Wright's Law are not competing theories — they are complementary forces that reinforce each other. Huang's Law operates on the supply side: each new GPU architecture (Blackwell in 2024, Rubin in 2026, Feynman on the roadmap) delivers a step-function improvement in inference throughput per dollar. Wright's Law operates on the production side: as TSMC fabricates more advanced chips, as cloud providers deploy more inference capacity, and as AI-native applications proliferate, cumulative experience drives costs down a predictable learning curve. The flywheel works because Huang's Law creates the performance headroom that stimulates demand, and Wright's Law ensures that scaling that demand makes each unit cheaper — which stimulates even more demand via Jevons' Paradox.

The Measurement Problem: Time vs. Volume

The deepest difference between these two laws is what they measure against. Huang's Law is time-indexed: it predicts that GPU performance will double within a fixed calendar interval, driven by NVIDIA's cadence of architectural releases. Wright's Law is volume-indexed: it predicts cost declines as a function of cumulative units produced, regardless of how long that production takes. This distinction matters enormously for forecasting. If GPU demand collapses (a scenario that would break Wright's Law predictions), Huang's Law could still hold — NVIDIA might still release faster chips on schedule. Conversely, if NVIDIA's architectural innovation slows (breaking Huang's Law), Wright's Law could still drive cost reductions through manufacturing learning on existing architectures. The Santa Fe Institute's 2012 study across 62 technologies found that volume-based Wright's Law outperformed time-based laws in forecasting accuracy, suggesting that production experience is the more fundamental driver of cost improvement.

Evidence from the Current GPU Cycle

The Blackwell-to-Rubin transition provides a live test of both laws. NVIDIA's Rubin architecture (shipping Q2 2026) delivers 50 PFLOPS of FP4 inference compute — a 5× improvement over Blackwell at the chip level and 3.3× at the rack level with NVL72 configurations. Memory bandwidth jumps from 8 TB/s to 13 TB/s, and NVLink throughput doubles to 260 TB/s. This is Huang's Law in action: architectural innovation (HBM4 memory, NVLink 6, new tensor core designs) delivering compound performance gains. Simultaneously, Wright's Law is visible in the broader inference market: H100 cloud prices have dropped to $1.49–3.90/hr as cumulative deployment scales, inference optimization techniques (quantization, speculative decoding, distillation) improve with deployment experience, and the all-in cost of GPT-3.5-class inference fell 280× in just two years.

Implications for AI Platform Economics

For platform builders, the interaction of these two laws creates a specific strategic calculus. Huang's Law means you should architect systems to absorb step-function hardware improvements — design for hardware abstraction, not hardware lock-in. Wright's Law means you should scale deployment volume aggressively, even at thin margins, because cumulative production experience creates a structural cost advantage that compounds over time. This is the playbook that Tesla used for batteries, that TSMC uses for semiconductor fabrication, and that cloud AI providers are now using for inference pricing — pricing ahead of the cost curve to maximize cumulative volume. The platforms that deploy the most inference today will have the lowest costs tomorrow.

Where the Laws Diverge: Failure Scenarios

Understanding where each law might break is critical for risk assessment. Huang's Law depends on continued architectural innovation at NVIDIA (and increasingly AMD, Intel, and custom ASIC builders like Google's TPU team and Amazon's Trainium). If the architectural design space for AI accelerators reaches diminishing returns — if there are no more easy gains from numerical format tricks, sparsity exploitation, or memory hierarchy optimization — Huang's Law slows. The skeptical view from EpochAI is that GPU price-performance has historically doubled every 2.5 years, not annually, suggesting Huang's Law may overstate the pace. Wright's Law, by contrast, fails when production volume stalls. Geopolitical risk to semiconductor supply chains (particularly TSMC's concentration in Taiwan), energy constraints on data center buildout, or a collapse in AI investment could all slow cumulative production growth and flatten the learning curve.

The Compounding Effect on the Agentic Economy

When both laws hold simultaneously, the effects compound. Each new GPU generation (Huang's Law) makes AI inference cheaper per unit of compute. That cheaper inference expands the addressable market for AI agents, driving up cumulative deployment (Wright's Law), which further reduces costs through production learning. This double-exponential cost decline is why features that seem economically impossible today — real-time AI NPCs, persistent simulated worlds, autonomous agent swarms — become feasible within a few hardware generations. The Jevons' Paradox dynamic ensures that each cost reduction stimulates more than proportional demand growth, financing the next round of GPU R&D and manufacturing investment. As long as both laws continue to hold, the agentic economy's growth trajectory remains exponential — bounded not by compute economics but by the pace at which organizations can redesign their workflows around abundant, cheap intelligence.

Best For

Forecasting AI Hardware Performance

Huang's Law

For predicting what the next generation of GPUs will be capable of — throughput, memory bandwidth, inference FLOPS — Huang's Law and NVIDIA's published roadmap (Rubin 2026, Rubin Ultra 2027, Feynman beyond) are the direct guide. Wright's Law doesn't predict performance, only cost.

Projecting AI Inference Cost Trajectories

Wright's Law

Wright's Law is the superior framework for forecasting how much inference will cost in 2–5 years. ARK Invest's modeling shows AI hardware costs declining ~37.5% per cumulative production doubling, and the Santa Fe Institute validated Wright's Law across 62 technologies. Time-based laws are less reliable for cost prediction.

Platform Pricing Strategy

Wright's Law

If you're setting pricing for an AI-powered platform, Wright's Law tells you to price ahead of the cost curve — accept thin margins now to maximize cumulative volume, knowing costs will decline predictably with scale. This is the TSMC and Tesla battery playbook.

Hardware Procurement and Capacity Planning

Huang's Law

For deciding when to buy GPUs, which generation to target, and how to plan data center capacity, Huang's Law and NVIDIA's roadmap are essential. Knowing that Rubin delivers 5× Blackwell's inference throughput directly informs buy-vs-wait decisions.

Venture Capital and Technology Investment

Both Essential

Investors need both: Huang's Law to understand the performance trajectory of AI hardware, and Wright's Law to model when specific applications become economically viable as cumulative deployment drives costs down the learning curve. ARK Invest's framework explicitly combines both.

Building an AI Agent Startup

Both Essential

Agent builders benefit from Huang's Law (each GPU generation makes previously impossible architectures viable) and Wright's Law (as the ecosystem scales, tools, frameworks, and inference costs all improve with cumulative experience). The compounding effect means capabilities expand and costs shrink simultaneously.

Competitive Moat Analysis

Wright's Law

Wright's Law explains winner-take-most dynamics: the company that accumulates production experience fastest enjoys a structural cost advantage that compounds. This is why TSMC, not Intel, dominates advanced fabrication — and why early AI infrastructure scale creates durable advantages.

Modeling Jevons' Paradox in AI Demand

Both Essential

The full Jevons' Paradox flywheel requires both laws: Huang's Law provides the step-function cost reductions that trigger demand expansion, and Wright's Law ensures that the expanded production further reduces costs, creating the self-reinforcing cycle that makes AI demand super-elastic.

The Bottom Line

Huang's Law and Wright's Law are not alternatives — they are the twin engines of AI's economic transformation. Huang's Law tells you how fast AI hardware will improve: expect 3–5× inference throughput gains per GPU generation on NVIDIA's annual cadence, with Rubin (2026) already delivering 5× over Blackwell. Wright's Law tells you how predictably costs will fall as deployment scales: 20–37.5% cost reduction per cumulative production doubling, validated across decades of technology history. Use Huang's Law for hardware strategy and performance planning. Use Wright's Law for cost modeling, pricing strategy, and competitive analysis. Use both together to understand the compounding flywheel — mediated by Jevons' Paradox — that is driving AI inference from luxury to commodity and making the agentic economy inevitable.