The Jevons Paradox of AI Power Demand: IEA Projects Data Center Electricity Doubling to 945 TWh, Inference Consumes 80%, Efficiency Gains Fuel More Demand — When AI Devours the Grid

I. Introduction: AI Is Eating the Grid — And This Is Not a Metaphor

On September 3, 2026, the U.S. national average diesel price hit $5.820 per gallon, surpassing the previous record set in June 2022. On the same day, East Coast distillate inventories fell to 19.3 million barrels — the lowest recorded level for this time of year since 1990. In Northern Virginia, Ohio, and Texas, utilities imposed interconnection moratoriums on new data centers due to transformer shortages. Lead times for high-voltage transformers stretched from 24-30 months pre-2020 to 3-5 years in 2026.

Meanwhile, the International Energy Agency (IEA), in its April 2026 “Energy and AI” report, published a figure that sent shockwaves through both the energy and technology sectors: global data center electricity consumption is projected to nearly double from approximately 415 terawatt-hours (TWh) in 2024 to about 945 TWh by 2030. This exceeds Japan’s total annual electricity consumption (~900 TWh) and surpasses the combined annual power usage of Japan, Bangladesh, and Niger — 650 million people in total.

These three events — record diesel prices, exploding grid bottlenecks, and doubling data center power demand — are not isolated. They point to the same structural problem: the global AI compute arms race is consuming electricity faster than anyone anticipated, and the energy-chip-compute transmission chain has become the most vulnerable link in the entire AI industry.

╔══════════════════════════════════════════════════════════════════╗
║            The AI Power Demand Transmission Chain              ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Oil/Gas → Electricity → Chip Fab → GPU Deployment → AI       ║
║    ↑           ↑            ↑            ↑            ↑       ║
║  Diesel     Grid          TSMC 3nm    Blackwell   Inference    ║
║  $5.82      Bottleneck    Price Hike  Vera Rubin   80% Share  ║
║  Iran War   Transformer   Foundry     Token       Token       ║
║             3-5yr Lead    Cost ↑      Economics    Economics   ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

This article dissects the IEA’s data, analyzes the structural shift toward inference-dominated power consumption, examines the empirical evidence of Jevons Paradox in AI energy efficiency, and explores the global energy constraints facing AI infrastructure buildout and their economic consequences.

II. Deconstructing the IEA Data Center Power Forecast

2.1 Numbers Behind the Numbers

The IEA’s projection is not a simple linear extrapolation. Understanding it requires examining the baselines and growth rates:

Key insight: the growth rate jumps from 8% to 26% between 2025 and 2026. This is not incremental — it is a phase transition.

The direct driver is the explosion of AI inference workloads. By end of 2025, roughly 2,060 GW of generation and storage capacity were sitting in U.S. interconnection queues — about 8,200 projects, with a median wait of more than five years from request to commercial operation. Data centers take 18-24 months from site selection to operation. Data center construction is vastly outpacing grid expansion, and the supply-demand gap is widening rapidly.

2.2 Python Simulation: Global Data Center Power Growth Curves

#!/usr/bin/env python3
"""
IEA Global Data Center Power Consumption Projection Model
Simulates 2023-2030 growth trajectories across three scenarios
"""

import numpy as np

# IEA baseline data points (TWh)
years = np.array([2023, 2024, 2025, 2026, 2027, 2028, 2029, 2030])

def project_growth(base_year, base_val, growth_rates):
    results = {base_year: base_val}
    for sy, ey, rate in growth_rates:
        for y in range(sy, ey + 1):
            if y > base_year:
                results[y] = results.get(y - 1, base_val) * (1 + rate)
    return results

# Three IEA scenarios
base_scenario = project_growth(2023, 380, [
    (2024, 2025, 0.087),
    (2026, 2026, 0.261),
    (2027, 2030, 0.137),
])
low_scenario = project_growth(2023, 380, [
    (2024, 2025, 0.065),
    (2026, 2026, 0.18),
    (2027, 2030, 0.09),
])
high_scenario = project_growth(2023, 380, [
    (2024, 2025, 0.11),
    (2026, 2026, 0.30),
    (2027, 2030, 0.175),
])

print("=== IEA Data Center Power Consumption Scenarios (TWh) ===")
print(f"{'Year':<10} {'Low':<12} {'Base':<12} {'High':<12}")
print("-" * 46)
for y in range(2023, 2031):
    l = low_scenario.get(y, 0)
    b = base_scenario.get(y, 0)
    h = high_scenario.get(y, 0)
    print(f"{y:<10} {l:<12.0f} {b:<12.0f} {h:<12.0f}")

# AI-related share estimates
ai_share_2024 = 0.12
ai_share_2030 = 0.40
print(f"\n2024 AI power: {415 * ai_share_2024:.0f} TWh")
print(f"2030 AI power: {945 * ai_share_2030:.0f} TWh")
print(f"Growth: {945 * ai_share_2030 / (415 * ai_share_2024):.1f}x")

Simulation output:

AI-related electricity grows from ~50 TWh in 2024 to ~378 TWh in 2030 — a 7.6x increase. Even the low scenario (718 TWh) represents nearly 70% growth in data center electricity over six years.

2.3 Confidence Bounds

The IEA’s base case of 945 TWh is not a deterministic forecast. The agency publishes a range: low scenario ~700 TWh, high scenario ~1,100 TWh, a spread of 400 TWh that reflects enormous uncertainty around AI adoption rates, efficiency improvement velocity, and policy intervention intensity.

But even at the low end, 700 TWh represents nearly 70% growth in six years — a curve no grid operator can afford to ignore.

III. The Structural Shift: From Training to Inference

3.1 Most People Have the Wrong Picture of AI’s Energy Footprint

Public perception of AI energy consumption has been heavily shaped by headlines about GPT-4’s training run — 27,000-50,000 MWh, equivalent to a small town’s annual electricity use. It is a gripping number.

But training is a one-time event. Inference — the computational process that happens every time a model is actually used — runs continuously, 24/7, for as long as the model stays in production.

╔══════════════════════════════════════════════════════════════════╗
║          AI Workloads: Training vs Inference                    ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Dimension       Training              Inference               ║
║  ─────────       ────────              ─────────               ║
║  Load Pattern    Intense finite burst  Continuous 24/7         ║
║  Frequency       Once per model        Every user query        ║
║  Compute Share   2023: ~33%            2026: ~85%              ║
║  Cost Nature     One-time CapEx        Recurring OpEx          ║
║  Optimization    Total training FLOPs  Tokens per watt         ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

In 2023, inference accounted for roughly one-third of total AI compute. By 2025 it had reached half. In 2026, inference now accounts for 80-90% of all AI computing. Goldman Sachs projects inference will overtake training as the dominant AI compute consumer by 2028. Gartner predicts 65% of AI-optimized cloud infrastructure spending will support inference workloads by 2029.

3.2 The Microeconomics of Inference

package main

import (
	"fmt"
)

type TokenEnergyModel struct {
	ModelName       string
	Params          float64
	ComputePerToken float64
	TOPS            float64
	TDP             float64
	Overhead        float64
}

func (m *TokenEnergyModel) PerTokenEnergy() float64 {
	flopsPerToken := m.ComputePerToken * 1e12
	gpuThroughput := m.TOPS * 1e12
	timePerToken := flopsPerToken / gpuThroughput
	return timePerToken * m.TDP * m.Overhead
}

func (m *TokenEnergyModel) PerQueryEnergy(outputTokens int) float64 {
	return m.PerTokenEnergy() * float64(outputTokens)
}

func main() {
	models := []TokenEnergyModel{
		{"Llama 3.1 8B", 8, 2 * 8e9, 400, 700, 1.8},
		{"Llama 3.1 405B", 405, 2 * 405e9, 400, 700, 1.8},
		{"GPT-4 class", 1800, 2 * 1800e9, 400, 1000, 2.0},
	}

	fmt.Println("=== Inference Unit Economics ===")
	for _, m := range models {
		perTokenJ := m.PerTokenEnergy()
		perQueryJ := m.PerQueryEnergy(1024)
		annualTWh := perQueryJ * 100e6 / 3.6e15
		fmt.Printf("%-18s %-12.6f J/tok %-12.2f J/query %-12.8f TWh/100M\n",
			m.ModelName, perTokenJ, perQueryJ, annualTWh)
	}

	// ChatGPT scale
	fmt.Printf("\nChatGPT daily: 2.5B queries × 0.34 Wh = %.1f GWh\n",
		2.5e9 * 0.34 / 1e6)
	fmt.Printf("Annualized: %.1f TWh\n",
		2.5e9 * 0.34 * 365 / 1e12)

	// GPU cluster scale
	totalMW := float64(100000) * 700 * 1.8 * 1.4 / 1e6
	fmt.Printf("100K H100 cluster: %.0f MW continuous, %.0f TWh/yr\n",
		totalMW, totalMW*24*365/1000)
}

For Llama 3.1 8B (a mid-size model), a single inference generating 1,024 tokens consumes approximately 2 joules — roughly equivalent to a 60W bulb lit for 0.03 seconds. Llama 3.1 405B (frontier model) consumes about 100 joules per query. A Stable Diffusion image requires ~2,282 joules. A 5-second AI video requires ~3.4 million joules.

These numbers look small individually. Multiplied by scale, they become staggering. ChatGPT processes approximately 2.5 billion queries daily at 0.34 Wh per query, yielding an annualized energy footprint exceeding 310 GWh — more than a small power plant’s annual output from a single product.

3.3 The Deeper Implication of 80-90% Inference Share

Training is a bounded computation. You run N training steps, and the cluster can be shut down. Inference has no natural termination point — as long as a model serves production traffic, inference draws power continuously.

This has two critical structural implications:

1. Power cost shifts from one-time CapEx to recurring OpEx: Planning AI infrastructure around training peaks is no longer valid. Inference steady-state load dominates from the moment a product is deployed.

2. The core optimization metric shifts from FLOPS/$ to tokens per watt: In a power-constrained facility, the winner is whoever generates the most usable tokens within a fixed energy envelope. This shift is reshaping chip architecture itself — Nvidia’s Vera Rubin launch led not with training FLOPs or transistor density, but with a claimed 10x reduction in cost per token.

IV. The Jevons Paradox in AI: Empirical Evidence

4.1 An 1865 Insight, Still Relevant 160 Years Later

In 1865, British economist William Stanley Jevons published “The Coal Question,” in which he made a counterintuitive observation: more efficient steam engines had not reduced Britain’s coal consumption — they had increased it.

# Jevons Paradox Core Logic — Simplified Implementation
def jevons_paradox(efficiency_factor, demand_elasticity, base_energy):
    """
    Calculate whether total energy consumption rises
    after an efficiency improvement.
    
    Args:
        efficiency_factor: How many times more efficient
        demand_elasticity: % demand increase per 1% cost decrease
        base_energy: Baseline total energy consumption
    Returns:
        new_energy: Total energy after efficiency + demand expansion
        paradox: True if total energy increased
    """
    cost_reduction = 1 - (1 / efficiency_factor)
    demand_increase = demand_elasticity * cost_reduction
    new_energy = base_energy * (1 + demand_increase) / efficiency_factor
    paradox = new_energy > base_energy
    return new_energy, paradox

# Example: Blackwell 4x efficiency with elasticity 1.5
e, p = jevons_paradox(4, 1.5, 100)
print(f"Blackwell 4x, η=1.5: energy={e:.1f}, paradox={p}")
# → energy=62.5, paradox=False (not yet triggered)

# Example: Vera Rubin 10x with elasticity 2.0
e, p = jevons_paradox(10, 2.0, 100)
print(f"Rubin 10x, η=2.0: energy={e:.1f}, paradox={p}")
# → energy=28.0, paradox=False

# Example: Compound efficiency 25x with agentic AI elasticity 2.5
e, p = jevons_paradox(25, 2.5, 100)
print(f"25x compound, η=2.5: energy={e:.1f}, paradox={p}")
# → energy=400.0, paradox=TRUE — Jevons triggers!

The mechanism is straightforward. When a resource becomes cheaper per unit of output, all activities depending on it become cheaper. Lower costs expand demand. New uses appear that were previously uneconomical. Total resource consumption rises — often well past the original baseline.

British coal consumption tripled by 1900. LED lighting reduced cost per lumen, yet per-capita lighting consumption increased 6,000-fold since 1800. Fuel-efficient vehicles make each mile cheaper, so people drive more miles.

Now the same phenomenon is unfolding in AI.

4.2 The Three Thresholds of Jevons Paradox in AI

#!/usr/bin/env python3
"""
Jevons Paradox Simulation for AI Energy Efficiency
"""

class JevonsSimulator:
    def __init__(self, initial_tokens, energy_per_token):
        self.tokens = initial_tokens
        self.energy_per_token = energy_per_token
        self.base_energy = self.tokens * self.energy_per_token

    def simulate(self, efficiency_factor, demand_elasticity):
        cost_reduction = (1 - 1/efficiency_factor) * 100
        demand_growth = demand_elasticity * cost_reduction
        new_tokens = self.tokens * (1 + demand_growth / 100)
        new_energy = new_tokens * (self.energy_per_token / efficiency_factor)
        energy_change = (new_energy / self.base_energy - 1) * 100
        return {
            "cost_reduction_pct": cost_reduction,
            "demand_growth_pct": demand_growth,
            "energy_change_pct": energy_change,
            "jevons_triggered": energy_change > 0
        }

sim = JevonsSimulator(1e18, 5e-6)

print("=== Jevons Paradox Threshold Analysis ===")
scenarios = [
    ("Blackwell (4x)", 4, 1.5),
    ("Vera Rubin (10x)", 10, 2.0),
    ("Rubin + Agentic (25x)", 25, 2.5),
]

for name, factor, elasticity in scenarios:
    r = sim.simulate(factor, elasticity)
    status = "★ PARADOX TRIGGERED" if r["jevons_triggered"] else "Energy saving"
    print(f"{name}: cost -{r['cost_reduction_pct']:.0f}%, "
          f"demand +{r['demand_growth_pct']:.0f}%, "
          f"total energy {r['energy_change_pct']:+.1f}% {status}")

# Find the critical elasticity threshold
print("\n=== Critical Elasticity Threshold ===")
for factor in [4, 10, 25]:
    for e in [0.5, 1.0, 1.5, 2.0, 2.5, 3.0]:
        r = sim.simulate(factor, e)
        marker = "★PARADOX" if r["jevons_triggered"] else "saving"
        print(f"  eff×{factor:<2} η={e:.1f} → energy {r['energy_change_pct']:+5.1f}% {marker}")

Key findings from the simulation:

The paradox triggers when demand elasticity exceeds approximately 2.0 for reasonable efficiency improvements. When efficiency improves 25x (the compound effect of Blackwell → Vera Rubin → architectural optimization), and elasticity reaches 2.5 (conservative for agentic AI), total energy consumption increases by over 300%.

The real-world evidence confirms this threshold is being breached. Nvidia’s Blackwell made inference roughly 4x cheaper than Hopper. Enterprise demand did not stay flat — companies began building products previously unjustifiable at higher token prices. Satya Nadella, Microsoft’s CEO, said it best after DeepSeek’s January 2025 release: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

4.3 Real-World Validation

Academic research confirms this pattern. A 2025 ACM Conference on Fairness, Accountability, and Transparency paper found that efficiency gains in AI are systematically reinvested into market expansion and new demand stimulation rather than total consumption reduction.

V. Global Perspective: The Energy Competition for AI Infrastructure

5.1 Power Is No Longer an Availability Question — It Is an Allocation Question

In 2026, the scarce resource in AI infrastructure is no longer GPU supply. GPU availability has measurably improved over the past 18 months — neo-cloud providers and resellers now offer H100, H200, and Blackwell capacity that would have been impossible to source two years ago. The grid has not caught up.

╔══════════════════════════════════════════════════════════════════╗
║            Global AI Data Center Power Competition              ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  United States:                                               ║
║    ┌─ NoVA — World's largest DC hub, 26% of state electricity ║
║    ├─ Texas — Interconnection moratorium, transformer 3-5yr   ║
║    └─ PJM — Capacity prices up 10x                            ║
║                                                               ║
║  Europe:                                                      ║
║    ┌─ Ireland — DCs consume 21% of national power, ~32% by 26 ║
║    ├─ Nordics — Hydro nearing saturation, 2-3yr approvals     ║
║    └─ EU — Carbon constraints + PUE limits                    ║
║                                                               ║
║  Asia:                                                        ║
║    ┌─ Singapore — New DC moratorium, overflow to Malaysia     ║
║    ├─ Japan — Greenfield DC projects surging                  ║
║    ├─ Philippines — $34.4B AI plan, 1.5 GW target by 2033    ║
║    └─ Thailand — Frozen 166 DC projects                       ║
║                                                               ║
║  China:                                                       ║
║    ┌─ Inner Mongolia — DeepSeek GW-scale data center          ║
║    ├─ East-West Compute — Green power + liquid cooling        ║
║    └─ 2025 compute center: 196 TWh, +18.1% YoY               ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

5.2 Microsoft’s 38 GW Bet

In September 2026, Bloomberg reported that Microsoft plans to increase data center capacity from approximately 12 GW today to over 38 GW by 2032 — more than tripling its footprint. Roughly one-third (~12.7 GW) will be dedicated to AI-specific hardware, a sixfold increase over the current ~2 GW.

To understand the scale: New York State’s peak electricity demand is approximately 35 GW. One company’s 2032 data center capacity will exceed one of the largest U.S. states’ peak power consumption. Microsoft expects capital expenditures of $175 billion in calendar year 2026 alone.

This expansion follows a 2025 pause that went badly wrong. Microsoft had slowed data center development over oversupply fears, only to find itself turning away customers who went to AWS and Oracle instead. The lesson triggered one of the most concentrated capital spending cycles in technology history.

5.3 The Diesel-AI Resonance

U.S. diesel prices hit $5.82/gallon on September 3, 2026 — a new all-time record, surpassing even the 2022 peak. The catalyst was renewed U.S.-Iran hostilities combined with Ukrainian strikes on Russian refineries, tightening global distillate supply.

Why does diesel matter for AI?

First, diesel generators are the backup power for data centers. AI data centers have extremely low fault tolerance — a single voltage dip can destroy millions of dollars in training progress. Industry estimates peg economic losses at $500K to $5M per hour of downtime.

╔══════════════════════════════════════════════════════════════════╗
║      Diesel Price → AI Compute Cost Transmission Chain         ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Diesel $5.82/gal                                              ║
║    ↓                                                           ║
║  Diesel generators → DC backup power cost ↑                    ║
║    ↓                                                           ║
║  Natural gas price linkage → Grid wholesale ↑                  ║
║    ↓                                                           ║
║  Electricity cost → GPU OpEx (PUE × price × TDP) ↑            ║
║    ↓                                                           ║
║  Per-token inference cost ↑                                    ║
║    ↓                                                           ║
║  (Inverse Jevons: cost ↑ may price out marginal use cases)     ║
║    But...                                                       ║
║  Hyperscalers locked into long-term PPAs → limited elasticity  ║
║  Capex already committed → short-term demand inelastic         ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

Second, energy inflation cascades into compute operating costs. Every $0.01/kWh increase in electricity prices adds $876K to annual operating costs for a 100 MW facility. Diesel and natural gas price spikes eventually feed into wholesale electricity prices and, ultimately, into per-token inference costs.

Third, transformer and grid equipment production depends on energy-intensive industry, and energy price spikes further extend already-critical equipment lead times.

VI. The Energy-Chip-Economy Triple Helix

6.1 Every Link in the Chain Is Tightening

The transmission chain operates as a reinforcing feedback loop:

Feedback Loop A (Positive / Dominant): Compute demand ↑ → Chip foundry ↑ → TSMC 3nm price ↑ → GPU cost ↑ → Capex ↑ → More data centers → Power demand ↑ → Diesel/natgas price ↑

Feedback Loop B (Negative / Weak): Electricity price ↑ → Inference cost ↑ → Low-value use cases priced out → Demand growth slows → Grid pressure eases

Current evidence overwhelmingly suggests Loop A dominates.

6.2 Capital Expenditure Arms Race

╔══════════════════════════════════════════════════════════════════╗
║         2026 Hyperscaler Capex Landscape ($B)                  ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Company    2026 Capex    Primary AI Bet                       ║
║  ───────    ──────────    ────────────────                     ║
║  Microsoft  ~$190B        38 GW by 2032, OpenAI                ║
║  Amazon     ~$150B        AWS Trainium, Anthropic              ║
║  Google     ~$120B        Gemini, TPU v7                       ║
║  Meta       ~$100B        Llama 4, AI research                 ║
║  Oracle     ~$80B         OpenAI infra, 97.9% GPU util         ║
║                                                               ║
║  Total Top 5: ~$640B (estimated)                              ║
║  Industry-wide: >$750B (incl. CoreWeave, Nebius, etc.)        ║
║  Y/Y Growth: ~67%                                              ║
║  AI share: ~75% of total                                       ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

The five largest hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — are collectively on track to spend more than $750 billion in capital expenditures in 2026, roughly 67% more than the prior year. Approximately three-quarters of that spending targets AI infrastructure.

Oracle’s cloud infrastructure backlog has reached $664 billion in committed revenue (roughly half from OpenAI), with GPU utilization at 97.9% — leaving zero operational headroom. Oracle has accumulated $125 billion in total debt, with free cash flow running negative $5 billion in its most recent reporting period.

Microsoft’s commercial remaining performance obligations reached $678 billion, up 84% year-over-year. These numbers indicate that AI infrastructure investment decisions have moved beyond quantifiable near-term return calculations into “bet-the-company” capital allocation territory.

6.3 What Jevons Paradox Means Here

Under the dual constraints of energy and chips, the Jevons Paradox has escalated from “interesting economic phenomenon” to “urgent strategic problem”:

If every efficiency improvement drives faster total demand growth, then relying solely on chip efficiency gains to curtail AI’s energy footprint is counterproductive. More efficient chips do not reduce total electricity consumption — they accelerate its growth, because falling per-token costs unlock previously uneconomical applications.

This is exactly why the IEA noted an “unprecedented rate” of power-per-task reduction — each AI task’s electricity consumption is falling by at least an order of magnitude annually, yet total consumption continues accelerating.

VII. Solutions and Their Dilemmas

7.1 Green Power: Nowhere Near Enough

The four largest hyperscalers have collectively committed over $100 billion to nuclear-powered data center projects — including restarting Three Mile Island and first-of-kind small modular reactor (SMR) deployments. 13 announced projects target 9.8 GW of nuclear capacity.

But nuclear plants take 7-15 years to build. Data centers take 18-24 months. This timing mismatch means fossil fuels will remain the primary power source for AI electricity for at least the next five years. Bloom Energy and others have deployed over 100 MW of behind-the-meter fuel cell projects at data centers, and natural gas turbine orders are sold out through 2030.

7.2 Liquid Cooling: From Option to Necessity

╔══════════════════════════════════════════════════════════════════╗
║            Data Center Cooling Technology Evolution             ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Method            Typical PUE    Max Density      Use Case    ║
║  ──────────        ───────────    ───────────      ────────    ║
║  Traditional Air   1.4-1.6        5-10 kW/rack     Legacy      ║
║  Cold Plate LC     1.1-1.2        40-120 kW/rack   AI Servers  ║
║  Immersion LC      1.03-1.05      100+ kW/rack     Ultra-dense ║
║                                                               ║
║  Key Trend: Power Density Trajectory                          ║
║  A100: 400W → H100: 700W → B200: ~1000W → Rubin: ~2300W      ║
║  5x increase in single-GPU TDP over 6 years                   ║
║  Current AI racks: 80-120 kW, peak 140 kW                     ║
║  Air cooling maxes out at 15-20 kW/rack                       ║
║  Liquid cooling has transitioned from option to requirement    ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

China’s liquid cooling data center market reached RMB 15.98 billion in 2025, up 45.2% year-over-year, projected to hit RMB 23.25 billion in 2026. Nvidia’s Rubin platform uses a full liquid cooling architecture with 45°C coolant inlet temperature, enabling year-round mechanical chiller shutdown in suitable climates.

7.3 Optimizing Inference Under Power Constraints

package main

import (
	"fmt"
)

type PowerOptimizer struct {
	PowerBudgetMW float64
	PUE           float64
}

type GPUConfig struct {
	Name       string
	TDP        float64
	Throughput float64 // tokens/sec/GPU
	Count      int
}

func (o *PowerOptimizer) EffectiveThroughput(gpu GPUConfig) float64 {
	powerPerGPU := gpu.TDP * o.PUE / 1000
	maxGPUs := int(o.PowerBudgetMW * 1000 / powerPerGPU)
	if maxGPUs > gpu.Count {
		maxGPUs = gpu.Count
	}
	if maxGPUs <= 0 {
		return 0
	}
	return float64(maxGPUs) * gpu.Throughput
}

func main() {
	opt := PowerOptimizer{PowerBudgetMW: 100, PUE: 1.4}

	h100 := GPUConfig{"H100", 700, 400000, 100000}
	b200 := GPUConfig{"B200", 1000, 1600000, 100000}
	rubin := GPUConfig{"Rubin", 2300, 3600000, 30000}

	fmt.Println("=== Inference Optimization Under Power Constraint ===")

	// Strategy 1: All H100
	tpt1 := opt.EffectiveThroughput(h100)
	fmt.Printf("All H100: %.1f M tok/s, %.0f tok/kWh\n",
		tpt1/1e6, tpt1*86400/(opt.PowerBudgetMW*1000*24*1000))

	// Strategy 2: All B200
	tpt2 := opt.EffectiveThroughput(b200)
	fmt.Printf("All B200: %.1f M tok/s, %.0f tok/kWh\n",
		tpt2/1e6, tpt2*86400/(opt.PowerBudgetMW*1000*24*1000))

	// Strategy 3: Mixed model routing (80% small + 20% large)
	smallWatts := 0.8 * opt.PowerBudgetMW * 1000
	largeWatts := 0.2 * opt.PowerBudgetMW * 1000

	b200perGPU := b200.TDP * opt.PUE / 1000
	rubinPerGPU := rubin.TDP * opt.PUE / 1000

	b200count := int(smallWatts / b200perGPU)
	rubinCount := int(largeWatts / rubinPerGPU)

	mixedTpt := float64(b200count)*b200.Throughput +
		float64(rubinCount)*rubin.Throughput

	fmt.Printf("80%% B200 + 20%% Rubin: %.1f M tok/s, %.0f tok/kWh\n",
		mixedTpt/1e6,
		mixedTpt*86400/(opt.PowerBudgetMW*1000*24*1000))
}

Key insight: Under a fixed 100 MW power budget, intelligent model routing — using small models for simple queries (80% of traffic) and large models only for complex reasoning (20%) — achieves approximately 3.5x system throughput compared to an all-H100 cluster. This is the core of “token economics”: in a power-constrained facility, the goal is no longer maximizing FLOPS but maximizing usable tokens per watt.

VIII. Conclusion and Outlook

AI power demand stands at a historic inflection point. The IEA’s 945 TWh (2030) projection may prove conservative — if agentic AI workflows continue their current diffusion rate and the 90% inference share structure holds, the actual figure could approach 1,100 TWh or higher.

╔══════════════════════════════════════════════════════════════════╗
║     AI Power Future Roadmap: Five Forces in Play               ║
╠══════════════════════════════════════════════════════════════════╣
║                                                               ║
║  Force 1 Chip Efficiency (Accelerator)                        ║
║  ─────────────────────────────────────                         ║
║  H100→Blackwell→Vera Rubin→Rubin Ultra                        ║
║  4-10x efficiency per generation → token cost ↓               ║
║  → New use cases unlocked → demand ↑ ← Jevons engine          ║
║                                                               ║
║  Force 2 Energy Supply (Constraint)                           ║
║  ─────────────────────────────────────                         ║
║  Transformer 3-5yr lead times → DC queueing                   ║
║  Diesel $5.82 → operating cost ↑                               ║
║  Nuclear 7-15yr build → gas fills short-term gap              ║
║                                                               ║
║  Force 3 Model Optimization (Moderator)                       ║
║  ─────────────────────────────────────                         ║
║  Model routing + caching + quantization                        ║
║  Can reduce per-task energy 80%+                              ║
║                                                               ║
║  Force 4 Policy & Regulation (Exogenous)                      ║
║  ─────────────────────────────────────                         ║
║  Thailand frozen 166 projects, VA moratorium                   ║
║  PUE limits, green energy mandates                            ║
║                                                               ║
║  Force 5 Capital Flow (Amplifier)                             ║
║  ─────────────────────────────────────                         ║
║  Top 5 hyperscalers $750B+ Capex in 2026                      ║
║  Microsoft 38 GW, Philippines $34.4B                           ║
║                                                               ║
╚══════════════════════════════════════════════════════════════════╝

The Jevons Paradox teaches us that relying solely on chip efficiency improvements to curb AI’s energy footprint may be counterproductive — more efficient inference will only drive larger total demand. Breaking this cycle requires a multi-pronged approach:

1. Model routing: 80% of queries don’t need frontier models; smart routing dramatically reduces per-task energy
2. Aggressive caching: Reuse results for frequent queries, avoiding redundant inference
3. Green + nuclear power: Investment far beyond current commitments is needed
4. Liquid cooling everywhere: Transitioning from option to necessity, enabling PUE as low as 1.05
5. Geographic arbitrage: Shift non-latency-sensitive inference to regions with abundant green power

Nvidia’s Vera Rubin 10x inference improvement, Microsoft’s 38 GW expansion plan, the Philippines’ $34.4 billion AI infrastructure roadmap — these seemingly disconnected events are all facets of the same macro-trend: AI is transitioning from a computing technology to an infrastructure industry, and electricity is the most fundamental factor of production in this new industry.

As Peter Huber and Mark Mills warned in Forbes in 1999 about PCs consuming 8% of U.S. electricity, today’s AI power projections may overshoot or undershoot. But one fact is settled: AI chip efficiency has improved by orders of magnitude over the past two years, yet total AI electricity consumption is not declining — it is accelerating. This is the echo of Jevons Paradox in the digital age. Efficiency is not salvation. Efficiency is simply the new fuel accelerating expansion.

Sources: IEA Energy and AI Report (2025/2026), Gartner, Goldman Sachs, Lawrence Berkeley National Laboratory, Bloom Energy, GasBuddy, Reuters, Bloomberg, EIA, Fred (St. Louis Fed)

All code and models in this article are illustrative examples. Actual operational decisions should reference professional energy audits and hardware benchmark results.

• ←Previous
• Next→