Source Report

1. Silicon Efficiency: ASICs and Techniques Outpace Raw Compute Growth, Shrinking Power per AI Token

Nvidia's Blackwell GPUs (B200) deliver up to 2.9x inference throughput per GPU versus H100s via FP4 precision and structured sparsity, enabling 70B-parameter models at $0.047/M tokens on spot instances—translating to 70-80% lower power per token for common workloads—while custom ASICs like Groq LPUs achieve 10-20x faster tokens/second than GPUs at sub-10ms latency by specializing matrix operations and eliminating general-purpose overhead.[1][2] This mechanism—precision reduction (FP16 to FP4), sparsity (N:M patterns skipping 50%+ zeros), and ASIC fixed-function units—collapses compute demand per task faster than AI usage scales, as inference (90%+ of future workloads) dominates training.[3]

• B200 FP4 sparsity hits 18,000 TFLOPS/GPU, 2.9x H200 on Llama 70B (MLPerf v5.1, Sep 2025).[1]
• Groq/SambaNova ASICs: 750-2000 tokens/sec on Llama 3.1 vs. Nvidia's 72-257, with 5-20x perf/watt.[2]
• Epoch AI: LLM inference costs fell 9-900x/year by task; fixed-performance inference prices drop 40x/year mid-range.[4]

Implications for entrants: Efficiency moats favor incumbents with data (e.g., Shopify's sales-derived lending), but open-source sparsity/quantization tools democratize 4x model shrinks; compete via edge ASICs for low-latency inference, viable by 2027 as inference surges to 40%+ of data center power.[5]

2. Workload Shift: Inference Efficiency Boom Overshadows Training, with No Plateau Yet but Diminishing Returns Emerging

AI compute pivots from training (one-off, bursty) to inference (continuous, 35% CAGR to 93 GW by 2030), where smaller quantized models (e.g., 4-bit Llama) run 1.5-4x faster on less power via techniques like KV-cache compression and speculative decoding—reducing effective demand as usage explodes without proportional power hikes.[5][6] Training compute grows 4-5x/year (Epoch AI), but R&D experiments consume 70-90% of total (not final runs), and post-training optimizations (test-time compute) yield gains without full retrains; no full plateau, but data exhaustion (~500T tokens indexed web) looms by 2030.[4]

• Inference to dominate: 40%+ data center power by 2030 (McKinsey), vs. training's 22% CAGR.[5]
• Efficiency: Quantization/pruning shrinks models 4x, speeds 1.3-1.7x; Google's Dynamo boosts inference 5x via memory optimizations.[6][7]
• Training trends: 4.5x/year growth continues to 2030, but open models hit 10²⁶ FLOP by late 2025.[8]

Implications for entrants: Training's high barrier (GW-scale clusters) favors hyperscalers; inference's edge/ASIC shift opens niches for flexible providers—target metro sites for low-latency by 2027, leveraging 90%+ GPU utilization via batching.[6]

3. Historical Overforecasts: Data Centers and Crypto Show Demand Surges Flatten via Efficiency and Volatility

Early 2000s data center boom exploded 90% electricity use 2000-2005 amid dot-com, but 2010-2018 global use flatlined despite soaring compute demand, as virtualization/PUE drops (from 2.0+ to 1.2) offset growth—echoing today's fears, where pre-2020 projections missed efficiency.[9] Crypto mining: 2017-18 hype forecasted massive surges, but post-bear markets and China's 2021 ban saw U.S. use stabilize at 0.6-2.3% national electricity (EIA), with operators pivoting to AI without sustained grid strain.[10]

• 2000s: Use rose linearly 1.6%/year post-peak, vs. exponential fears.[11]
• Crypto: Volatile; post-halving shutdowns cut inefficient ops, demand flexible.[12]

Implications for entrants: Utilities overbuild risks stranding assets if AI hype fades; new players secure flexible contracts (e.g., curtailment clauses) to hedge, mirroring crypto's pivot.

4. Grid Buildout Acceleration: FERC 1920 Mandates 20-Year Planning, Unlocking Transmission via Scenarios and GETs

FERC Order 1920 (May 2024) forces 20-year scenario planning with seven benefits metrics (e.g., reliability deferral, production savings), evaluating grid-enhancing tech (GETs like dynamic line ratings, up 40-80% capacity) and high-performance conductors before new lines—compliance filings due 2026, first projects 2028-2033, preempting AI bottlenecks.[13][14]

• Requires three diverse scenarios, state input, ATT evaluation (DLR/APFC/switching).[13]
• Builds on Order 1000; reviews start 2026.[15]

Implications for entrants: Speeds interconnections for flexible loads; co-locate with generators under PJM rules (FERC Feb 2025), but fund upgrades—viable 2027+ for AI inference.

5. Demand Flexibility: Software Shifts AI Loads, Absorbing 18-55% Peaks Without New Gen

Google's 1GW demand response (2026) with TVA/Entergy shifts ML training (deferrable) off-peak, cutting peaks 25% for hours via job queuing; Microsoft power-caps coordinate scheduling/infra; studies show 18-55% flexibility in regulation reserves/emergencies, unlocking 100GW data centers sans new plants if <50 hours/year curtailed.[16][17]

• Google: ML workloads shifted, 1GW capacity (TVA, Indiana Michigan, etc.).[16]
• Duke: 100GW via modest peaks; Emerald AI: 25% cut 3hrs.[18]

Implications for entrants: Mandate DR in leases; hyperscalers lead—partner for "flex credits" reducing interconnection queues by 2027.

6. Low-End Forecasts: EIA/NERC Project Manageable Growth at 1-2%/Year Overall, with Risks Overstated

EIA AEO 2026: Electricity demand grows 0.9-1.6%/year to 2050 (post-2.1% recent), data centers key but servers to 818 billion kWh high-case (16x 2020, ~8-12% total by 2030); NERC LTRA 2025: 224GW summer peak +10yrs (conservative per Grid Strategies), but critiques note 2/3 data center risk from chips/finance.[19][20]

• EIA: Commercial (data centers) drives, but efficiency tempers.[21]
• NERC: Data centers ~55% growth, but Grid: Overstated 90->~65GW.[20]

Implications for entrants: Low-end (1%/year) viable with flexibility; monitor EIA for policy.

Assessment of Disconfirming Factors

Most Likely to Reduce Severity (High Confidence, 2027-2030): Efficiency (1) and flexibility (5)—mechanisms proven, scaling fast; could halve effective demand growth. Medium (2028+): Inference shift (2), transmission (4). Low: History (3), plateau (no evidence), low forecasts (EIA conservative but risks noted). Timeline: Material relief by 2028 via software/hardware; full by 2030 if DR/GETs adopt. Additional research: Chip forecasts, DR pilots.

---

Recent Findings Supplement (April 2026)

GPU and Silicon Efficiency Improvements

NVIDIA's Vera Rubin platform (announced January 2026) integrates advanced NVFP4 precision and co-designed CPU-GPU architectures to deliver 10x higher inference throughput and 10x lower cost per token versus prior generations, enabling sustained efficiency gains that could outpace some task demand growth through hardware-software fusion—though Jevons paradox risks offsetting this via expanded AI usage.[1][2]
- Rubin doubles NVLink-C2C bandwidth and triples memory capacity, reducing power per inference operation; Blackwell already claims 25x inference efficiency over Hopper.[3][4]
- ArXiv papers (e.g., ZipServ, March 2026) show lossless compression yielding 1.22x end-to-end inference speedup and 30% model size reduction on GPUs like RTX5090, with fused kernels cutting memory bandwidth needs.[5]
Implications for competition/entry: New entrants gain via inference-optimized ASICs (e.g., Trainium, TPUs), but incumbents like NVIDIA hold data moats; efficiency could cap power per task at 2026 levels if sparsity/quantization scales, but aggregate demand likely rises (my inference from Jevons mentions, no direct source).

Shift to Inference Workloads

AI compute is shifting from training (predictable bursts) to inference (bursty, user-driven), projected to comprise 2/3 of workloads by 2026 and 65%+ by 2029, potentially plateauing training FLOPs as models mature while inference efficiency improves via distillation/quantization—reducing overall power intensity if adoption doesn't explode via agents.[6]
- Inference market grows to $255B by 2030 (19% CAGR); 70% of data center demand from inference by 2030, with specialized chips favoring lower-power edge deployment.[6]
- Training-like baseload shifts to inference variability, enabling flexibility (e.g., non-urgent tasks rescheduled).[7]
Implications for competition/entry: Favors hyperscalers with global inference fleets (Google, AWS); smaller players enter via edge inference, but power savings hinge on no agentic explosion (low confidence, training plateau inferred from shift data).

Demand Response and Flexible Loads

Google integrated 1GW demand response into U.S. data center contracts (March 2026) by shifting ML workloads during peaks, turning inflexible loads into "ghost batteries" that unlock grid headroom—Duke University models show U.S. grids absorbing 76-100GW new flexible data center load with <1% curtailment (e.g., <50 hours/year), avoiding new generation for 5-10 years.[8][9]
- AI data centers offer 18-55% flexibility vs. average power via power capping/rescheduling; EPRI Flex MOSAIC classifies for grid ops.[10][11]
- Utilities/data centers collaborate on curtailment, saving 6-54% costs while stabilizing grids.[12]
Implications for competition/entry: Most viable near-term mitigator (2026-2028); entrants partner with utilities for DR incentives, but requires software for workload shifting—highest potential to reduce severity without capex.

Accelerated Transmission via FERC Order 1920

FERC Order 1920 compliance filings began December 2025 (PJM/CAISO), mandating 20-year planning with state input and three selection criteria (performance, cost, maximization)—early progress in PJM/SPP/MISO unlocks GW-scale transmission for data centers by 2028-2030, complemented by colocation rules (Dec 2025) favoring on-site gas/nuclear.[13][14]
- FERC directs large-load reforms by June 2026; PJM adds transmission options for data center-generator colocation.[15][16]
Implications for competition/entry: Timeline too slow for 2026-2027 crunch (years away); favors regions with early filings (PJM), but federal siting dormant—medium impact post-2028.

Grid Forecasts and Overstated Demand Risks

EIA AEO2026 (April 2026) projects U.S. electricity growth at 0.9-1.6% annually to 2050 (low end baseline), with data centers driving commercial sector but EVs/data centers only 10-25% of total demand; Enverus 2026 Outlook calls AI boom "overstated" as behind-the-meter/on-site gen mutes grid impact.[17][18][19]
- High-demand case hits 818 TWh data centers by 2050 (16x 2020), but baseline assumes efficiency offsets some growth.[18]
Implications for competition/entry: Low-end scenarios (efficiency/DR) enable entry without crisis; no grid operator "no severe constraints" quote, but Duke/EPRI headroom suggests mitigable (estimated from models).

Historical Precedents (Weak Recent Evidence)

Early 2000s data center/video fears kept demand flat via efficiency (cited Nov 2025); BTC mining warnings (2017-2021) didn't overwhelm grids (0.8% global electricity), as economics capped growth—AI differs in inflexibility, but miners pivot to flexible AI hosting.[20][4]
- No 2026-specific AI parallels; crypto stabilized predictably.[20]
Implications for competition/entry: Least compelling; AI's scale/steadiness worse than crypto, low confidence in repeat.

Strongest Disconfirmers (Likelihood/Timeline): Flexible loads/DR (high, 2026 via Google/Duke); inference shift+efficiency (medium-high, 2026-2028); transmission reforms (medium, 2028+). Combined, could halve effective constraints by 2028 (inferred, no direct source); low-end forecasts reinforce. Additional research on NERC 2026 LTRA needed for confidence.