Source Report
Research Question
Research projections for AI training and inference compute needs through 2030-2035, including token processing requirements, model scaling laws, and data center capacity forecasts. Pull from OpenAI, Anthropic, Microsoft, Google reports and semiconductor industry analyses. Quantify the bull case for compute growth.
AI Compute Scaling Through 2035: The Infrastructure Race
Capital Requirements: $7 Trillion by 2030
McKinsey's analysis reveals the staggering financial commitment required to sustain AI scaling: $6.7 trillion in total data center investment by 2030, with $5.2 trillion specifically dedicated to AI-capable infrastructure[1]. This represents capital expenditures for 156 gigawatts of AI-related data center capacity, with 125 GW of incremental capacity added between 2025 and 2030[1]. The remaining $1.5 trillion funds traditional IT applications, underscoring how AI has become the dominant driver of infrastructure spending[1].
- AI models require massive compute for both training and inference workloads, with inference expected to dominate by 2030[1]
- Enterprise AI deployment across automotive, financial services, and other sectors is the primary driver behind this capital acceleration[1]
- This $7 trillion figure reflects the magnitude of the infrastructure challenge; anything less leaves enterprise demand unmet
Power Demand Trajectories: 165% Growth by 2030
Global data center power demand is forecast to rise 165% by 2030 compared to 2023 levels, according to Goldman Sachs[4]. This translates to roughly 92 gigawatts by 2027 (50% growth from current levels), with a compound annual growth rate of 17% between 2025 and 2028[4]. In bullish scenarios where GPU power requirements exceed expectations or AI adoption accelerates, the CAGR could reach 20%[4].
For the US specifically, the projections diverge by analytical source—reflecting genuine uncertainty about AI adoption velocity:
| Forecasting Body | 2024 US Baseline | 2030-2035 Target | Multiplier | Source |
|---|---|---|---|---|
| BloombergNEF | 35 GW | 78 GW (2035) | 2.2x by 2035 | [5] |
| Deloitte | 4 GW (AI only) | 123 GW (2035, AI only) | 30x by 2035 | [6] |
The Deloitte projection is notably aggressive—it isolates AI data center power demand (4 GW in 2024) and projects it to consume 123 GW by 2035. This 30x expansion reflects the "bull case" assumptions: rapid large-model training scale-up, high inference loads from deployed models, and limited efficiency gains offsetting growth[6].
BloombergNEF's more moderate trajectory (2.2x overall) assumes efficiency innovations like DeepSeek V3's "Mixture of Experts" architecture will provide meaningful checks on power demand escalation[5].
Training Compute Scale: 2e29 to 5e30 FLOP Range by 2030
Epoch AI's analysis provides the most granular breakdown of training compute trajectories. Under on-trend extrapolation of the current 4x/year compute growth rate, training runs by 2030 will reach approximately 2e29 FLOP (roughly 5,000 times larger than GPT-4)[2]. This represents a median estimate and would require almost 20 million H100-equivalent GPUs for a single training run[2].
More aggressive scenarios—accounting for sustained efficiency improvements and extended training durations—project 6 GW of power demand for 2030 training facilities, implying training runs at the 2e29 FLOP scale[2]. The upper bound is far more dramatic: gigawatt-scale data centers (2-5 GW capacity) are feasible by 2030 according to utility company assessments, which would enable training runs between 1e29 and 5e30 FLOP—up to 250,000 times larger than GPT-4[2].
- The 5,000x scaling factor assumes 24x power efficiency gains from hardware improvements (4x from accelerators), lower-precision training (2x FP8 gains), and extended training durations (3x)[2]
- Computational demand specifically is projected at 2.5 × 1031 FLOPs by 2030 across all training workloads[3]
- Northern Virginia data center capacity is expected to grow from 5 GW to 10 GW, demonstrating regional infrastructure readiness[2]
GPU Production Bottleneck: 100 Million H100-Equivalents Needed
The semiconductor manufacturing constraint emerges as a critical limiting factor. Epoch AI's median projection estimates 100 million H100-equivalent GPUs will be manufactured by 2030, sufficient to power the 9e29 FLOP training runs referenced above[2]. However, this estimate carries significant uncertainty—the range spans 20 million to 400 million units, enabling training runs between 1e29 and 5e30 FLOP[2].
TSMC forecasts AI server demand will grow at 50% annually over the next five years, with actual GPU volume growth estimated at 35% per year (accounting for margin expansion and pricing dynamics)[2]. Reaching 100 million H100-equivalents by 2030 would require a vast expansion of GPU production far exceeding current levels[2].
- A single major AI lab can realistically access only ~20% of total GPU production capacity globally, necessitating the 100M+ unit production target to support competition among OpenAI, Google, Anthropic, and others[2]
- Advanced packaging and high-bandwidth memory expansion remain key production constraints; these bottlenecks dominate uncertainty in GPU supply projections[2]
- NVIDIA's Blackwell Ultra GB300 provides 50% improvement in dense FP4 compute over predecessors, enabling larger training runs per GPU[3]
Market Size: $231.8 Billion AI Hardware by 2035
The AI hardware market (GPUs, TPUs, custom accelerators, memory) was valued at $47.5 billion in 2024 and is projected to reach $231.8 billion by 2035 at a 23.2% CAGR[3]. Processors—particularly GPUs—are expected to dominate this market through 2030 and beyond[3]. Consumer electronics (smartphones, AR/VR devices with edge AI chips) represents the largest end-user segment and will grow at 23.9% CAGR[3].
Strategic capital commitments underscore industry commitment to scaling: Amazon alone allocated $30 billion to new US data centers to support AI model training and deployment[3].
Inference's Rise as Dominant Workload
A non-obvious implication emerges from McKinsey's analysis: inference is projected to become the dominant workload by 2030, overtaking training[1]. This matters because inference demands differ fundamentally—lower latency requirements, sustained continuous load (rather than episodic training runs), and distributed edge deployment alongside centralized cloud compute. The $5.2 trillion AI infrastructure investment must accommodate this shift, requiring both massive centralized inference clusters and distributed edge infrastructure.
Implications for Competitive Scaling
The gap between the 2x (BloombergNEF) and 30x (Deloitte) US AI data center power growth scenarios reveals the core strategic uncertainty: efficiency gains and model architecture innovations will determine whether compute demand scales linearly with capability advances or exponentially. Organizations betting on the bull case (Deloitte-aligned) are investing in gigawatt-scale infrastructure now. Those assuming moderate efficiency gains may face capacity constraints by 2027-2028. The GPU manufacturing ceiling (20M-400M H100-equivalents) is the hardest physical constraint—exceeding it requires winning allocation in a competitive TSMC/Samsung supply environment or building proprietary silicon at scale.
Sources:
- [1] https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers
- [2] https://epoch.ai/blog/can-ai-scaling-continue-through-2030
- [3] https://www.meticulousresearch.com/product/ai-hardware-market-6222
- [4] https://www.goldmansachs.com/insights/articles/how-ai-is-transforming-data-centers-and-ramping-up-power-demand
- [5] https://about.bnef.com/insights/commodities/power-for-ai-easier-said-than-built/
- [6] https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html
- [7] https://www.duperrin.com/english/2025/07/22/2035-ai-no-more-jobs/
Recent Findings Supplement (February 2026)
Big Five's $600B AI Infrastructure Spend in 2026
Major cloud providers (Microsoft, Google, Amazon, Meta, others) plan over $600 billion in capex for 2026, up 36% from 2025, with $450 billion targeted at AI infrastructure like GPUs and data centers; this triples GPU demand over the decade as hyperscalers secure supply amid chip/power constraints, shifting AI from software to infrastructure race.[1]
- NVIDIA captured 90% of AI accelerator spend; Q1 2025 data center revenue hit $35.6 billion.[1]
- Data center GPU market: $21.6 billion in 2025 to $265.5 billion by 2035 (28.5% CAGR).[1]
For competitors: Prioritize power-secured sites now, as $450B spend locks in NVIDIA supply, sidelining late entrants without grid access.
AI Supercomputer Market Explosive Growth to 2035
AI supercomputers, optimized for training/inference via parallel processors handling millions of parameters, reach $14.22 billion by 2035 from $1.91 billion in 2025 (22.29% CAGR), driven by government/commercial needs for deep learning and generative AI; U.S. segment alone grows from $0.56 billion to $4.15 billion.[2]
- Processors/compute hold 44% share in 2025 for neural network training/large simulations.[2]
- Cloud deployment dominates 61%, fastest-growing due to scalable AI-as-a-service.[2]
- North America leads at 41% share; Asia-Pacific fastest at 23.71% CAGR.[2]
For entrants: Target interconnects (fastest sub-segment) and healthcare apps, but face U.S./China policy hurdles on exports/security.
Power as New AI Bottleneck, Data Centers Drive 20%+ Demand Growth
By 2030, data centers fuel over 20% of advanced economy electricity growth (U.S. nearly 50%), with global demand up 10,000 TWh by 2035; AI shifts metrics from PUE to PCE, prioritizing power density over efficiency as grids strain.[1][3][4][5][6]
- Data centers: 415 TWh global in 2024 (1.5% total electricity), up 73% from 2023.[1]
- McKinsey: Capacity triples by 2030, 70% AI-driven; agentic AI pushes 75% firms to invest 2026.[3]
For scaling: Secure nuclear/renewables early—power trumps compute, per 2026 trends, enabling 2-5 GW training facilities by 2030.[8]
Bull Case Compute Projections Through 2030
Epoch AI confirms 2-5 GW AI training facilities feasible by 2030 under current trajectories, sustaining scaling laws via massive GPU/power ramps; combined with $600B spend and supercomputer surge, bull case implies 10-30x compute growth (FLOPS/training runs) by 2030-2035, limited only by energy.[1][2][8]
- UN/IEA: 30% global electricity rise by 2035, AI central.[1]
- No new OpenAI/Anthropic/Microsoft/Google reports in results, but hyperscaler capex validates prior scaling forecasts.
For bulls: Bet on Asia-Pacific (23%+ CAGR), but verify grid builds—power feasibility unlocks full Chinchilla-optimal scaling to 2035.
Sources:
- [1] https://carboncredits.com/ai-demand-to-drive-600b-from-the-big-five-for-gpu-and-data-center-boom-by-2026/
- [2] https://www.globenewswire.com/news-release/2026/01/29/3228286/0/en/AI-Supercomputer-Market-Size-to-Surpass-USD-14-22-Billion-by-2035-Research-by-SNS-Insider.html
- [3] https://www.networkworld.com/article/4118758/recent-compute-infrastructure-investments-signal-big-techs-ai-priorities-for-2026.html
- [4] https://www.digitalrealty.com/resources/blog/ai-predictions
- [5] https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html
- [6] https://www.jdsupra.com/legalnews/ai-trends-for-2026-power-not-compute-7025410/
- [7] https://epoch.ai/blog/can-ai-scaling-continue-through-2030
- [8] https://sapphireventures.com/blog/2026-outlook-10-ai-predictions-shaping-enterprise-infrastructure-the-next-wave-of-innovation/
- [9] https://maadvisor.com/maalerts/ai-tech-ma-why-decembers-100b-deal-sprint-just-defined-your-2026-opportunities/
Additional Insights from Follow-up Questions
No, building hundreds of GW per year of power generation capacity is not feasible under current global trends for supporting AI data centers, as total renewable additions reached only 585–741 GW in 2024 (mostly solar and wind), with forecasts of 793 GW in 2025—far short of hundreds of GW dedicated to AI.[1][5][6]
Current Global Power Capacity Additions
Global renewable capacity grew by a record 585 GW in 2024 (15.1% increase to 4,448 GW total), driven by solar (over 75% of additions) and wind; this pace falls short of the 16.6% annual growth needed to triple renewables to 11 TW by 2030.[1]
Forecasts predict 793 GW added in 2025 (11% rise from 2024's 717 GW), putting tripling within reach if sustained, but still not at hundreds of GW specifically for AI data centers.[5]
In the US, just 15 GW of total generation (91% renewables: 11.5 GW solar, 2.3 GW wind) was added January–May 2025, with 133 GW "high probability" additions by 2028 (84% solar/wind).[2]
Relevance to AI Data Center Needs
AI forecasts emphasize total data center power demand growing to 92 GW globally by 2027 (17–20% CAGR) or 123 GW AI-only in the US by 2035 (Deloitte bull case), requiring 125 GW incremental AI data center capacity by 2030—not annual generation builds at that scale.[1 from context][4 from context][6 from context]
- Data centers currently use ~415 TWh globally (1.5% of electricity), projected to drive 20%+ of advanced economy growth by 2030, but grid constraints prioritize power density over massive new builds.[Recent Findings Supplement]
- No sources project hundreds of GW/year in new generation; even optimistic renewable surges (e.g., G20 at 90% of 2024 additions) are broadly distributed, not AI-focused, amid regional disparities and grid bottlenecks.[1]
Bull case AI scaling (2–5 GW training facilities by 2030, 10–30x compute growth) relies on securing existing/new power (e.g., nuclear/renewables), not constructing hundreds of GW annually, which exceeds 2024's entire global renewable record by 50–100%.[Bull Case from context][7 from context]
Sources:
- [1] https://www.weforum.org/stories/2025/04/renewable-energy-transition-wind-solar-power-2024/
- [2] https://www.utilitydive.com/news/renewables-make-up-91-of-the-15-gw-of-generation-the-us-added-in-first-5-m/758603/
- [3] https://www.eia.gov/energyexplained/electricity/electricity-in-the-us-generation-capacity-and-sales.php
- [4] https://ourworldindata.org/grapher/installed-global-renewable-energy-capacity-by-technology
- [5] https://ember-energy.org/latest-insights/renewable-additions-in-2025-are-once-again-expected-to-surge-putting-tripling-within-reach/
- [6] https://www.c2es.org/content/renewable-energy/
- [7] https://www.carboncollective.co/sustainable-investing/gigawatt-gw
Additional Insights from Follow-up Questions
No, building hundreds of GW per year of power generation capacity is not feasible under current global trends for supporting AI data centers, as total renewable additions reached only 585–741 GW in 2024 (mostly solar and wind), with forecasts of 793 GW in 2025—far short of hundreds of GW dedicated to AI.[1][5][6]
Current Global Power Capacity Additions
Global renewable capacity grew by a record 585 GW in 2024 (15.1% increase to 4,448 GW total), driven by solar (over 75% of additions) and wind; this pace falls short of the 16.6% annual growth needed to triple renewables to 11 TW by 2030.[1]
Forecasts predict 793 GW added in 2025 (11% rise from 2024's 717 GW), putting tripling within reach if sustained, but still not at hundreds of GW specifically for AI data centers.[5]
In the US, just 15 GW of total generation (91% renewables: 11.5 GW solar, 2.3 GW wind) was added January–May 2025, with 133 GW "high probability" additions by 2028 (84% solar/wind).[2]
Relevance to AI Data Center Needs
AI forecasts emphasize total data center power demand growing to 92 GW globally by 2027 (17–20% CAGR) or 123 GW AI-only in the US by 2035 (Deloitte bull case), requiring 125 GW incremental AI data center capacity by 2030—not annual generation builds at that scale.[1 from context][4 from context][6 from context]
- Data centers currently use ~415 TWh globally (1.5% of electricity), projected to drive 20%+ of advanced economy growth by 2030, but grid constraints prioritize power density over massive new builds.[Recent Findings Supplement]
- No sources project hundreds of GW/year in new generation; even optimistic renewable surges (e.g., G20 at 90% of 2024 additions) are broadly distributed, not AI-focused, amid regional disparities and grid bottlenecks.[1]
Bull case AI scaling (2–5 GW training facilities by 2030, 10–30x compute growth) relies on securing existing/new power (e.g., nuclear/renewables), not constructing hundreds of GW annually, which exceeds 2024's entire global renewable record by 50–100%.[Bull Case from context][7 from context]
Sources:
- [1] https://www.weforum.org/stories/2025/04/renewable-energy-transition-wind-solar-power-2024/
- [2] https://www.utilitydive.com/news/renewables-make-up-91-of-the-15-gw-of-generation-the-us-added-in-first-5-m/758603/
- [3] https://www.eia.gov/energyexplained/electricity/electricity-in-the-us-generation-capacity-and-sales.php
- [4] https://ourworldindata.org/grapher/installed-global-renewable-energy-capacity-by-technology
- [5] https://ember-energy.org/latest-insights/renewable-additions-in-2025-are-once-again-expected-to-surge-putting-tripling-within-reach/
- [6] https://www.c2es.org/content/renewable-energy/
- [7] https://www.carboncollective.co/sustainable-investing/gigawatt-gw
No, AI data center power demand is not projected to increase by hundreds of GW per year in the 2030s; even the most aggressive forecasts show total cumulative growth reaching only around 100-130 GW by 2030, with annual increments far below 100 GW and no sources indicating hundreds of GW/year thereafter.[5][7]
Key Projections for Data Center Power Demand
Forecasts focus on total capacity or annual increments through 2030, not per-year explosions in the 2030s. Here's a synthesis of the highest estimates:
Source
2024/2025 Baseline
2030 Projection
Annual Growth Implication
Notes
Goldman Sachs
~35 GW (implied)
92 GW global (165% rise from 2023)
~10-15 GW/year avg. to 2027
Bull case: 20% CAGR if GPUs more power-hungry.[5]
S&P Global / 451 Research
55 GW (US IT+overhead, implied)
134 GW US (nearly triple)
~15-20 GW/year avg.
Excludes enterprise data centers; Virginia up 30% YoY.[7]
Deloitte
96 GW global by 2026 (incl. non-AI)
Not specified beyond doubling electricity use
AI >40% of 96 GW by 2026
GenAI doubles global DC electricity to 4% by 2030.[1]
US-focused estimates: Data center demand could hit 145 GW peak summer by 2031 (from 85 GW in 2024), with ~32 GW from data centers/crypto—implying ~5-10 GW/year avg. growth, half from AI.[4]
Electricity consumption (not capacity): Ranges 200-1,050 TWh/year by 2030 (midpoint ~650 TWh), equating to ~70-380 GW average power (at 8760 hours/year), but growth is gradual, not hundreds GW/year.[3][6]
2030s outlook: DNV projects US/Canada data centers at 16% of electricity by 2040 (AI 12%), but global AI share <3%—no annual hundreds GW jumps; growth slows post-2030 as electrification (EVs, industry) competes.[2]
Why Not Hundreds GW/Year?
Physical limits: Current global renewable additions max ~800 GW/year (all sources, not AI-dedicated); AI needs are incremental capacity (e.g., 125 GW total by 2030 per prior McKinsey).[5 from context]
No 2030s acceleration: Projections taper due to efficiency gains, on-site power (e.g., fuel cells, nuclear), and grid constraints; demand is "lumpy" but not exponential at that scale.[4][7]
Bull cases (e.g., BCG's 1,050 TWh) imply ~120 GW avg. power by 2030 total, with <50 GW/year peak growth—still orders below hundreds GW/year.[6]
These align with prior context: Bull scaling relies on securing power, not unprecedented builds.[Bull Case from context]
Sources:
- [1] https://www.deloitte.com/ro/en/about/press-room/studiu-deloitte-utilizarea-inteligentei-artificiale-generative-va-dubla-consumul-de-energie-electrica-al-centrelor-de-date-la-nivel-global-pana-2030.html
- [2] https://www.axios.com/2025/10/07/ai-power-cost-demand-future
- [3] https://www.breckinridge.com/insights/details/quantifying-power-demand-from-artificial-intelligence/
- [4] https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
- [5] https://www.goldmansachs.com/insights/articles/how-ai-is-transforming-data-centers-and-ramping-up-power-demand
- [6] https://www.wri.org/insights/us-data-centers-electricity-demand
- [7] https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/101425-data-center-grid-power-demand-to-rise-22-in-2025-nearly-triple-by-2030
- [8] https://insideclimatenews.org/news/12022026/inside-clean-energy-data-center-boom-electricity-demand/
Yes, demand for AI data centers could theoretically be very high if power and compute supply constraints are resolved, as projections are primarily driven by explosive AI workload growth (training, inference, agentic AI) rather than hard limits on end-user needs—potentially scaling to 100+ GW total capacity by 2030-2035 in bull cases.[3][7]
Drivers of Uncapped High Demand
AI adoption across enterprises, cloud, and edge creates insatiable compute hunger if infrastructure scales:
- Workload explosion: Inference dominates by 2030 (overtaking training), with AI at 27% of data center power by 2027 (up from negligible); sustained 24/7 GPU loads at 700W-1200W/chip and 30kW+ racks vs. 8kW traditional.[1][3]
- Quantified bull cases:
| Scenario | Power Demand Projection | Implied Annual Growth |
|----------|-------------------------|-----------------------|
| Goldman Sachs | 84 GW global by 2027; 165% rise by 2030 | ~10-15 GW/year[3] |
| Deloitte | 123 GW US AI-only by 2035 (30x from 4 GW) | ~10 GW/year avg.[7] |
| McKinsey/US | 12% of US electricity by 2030 | High if agentic AI surges[6] |
- Electricity analogs: 200-1,050 TWh/year by 2030 (~23-120 GW avg. power), with AI driving majority; hyperscalers/colos absorb via $80B DCPI spend (power/cooling) by 2030.[2][4]
Supply Removal Enables This
Power: Utilities need $720B grid investment by 2030; on-site gen (gas turbines, renewables), liquid cooling, and redesigns (slab layouts, DC power) bypass grid delays—operators already deploy multi-GW campuses.[2][3]
Compute: $600B hyperscaler capex (2026) triples GPU demand; if TSMC/NVIDIA ramp to 100M+ H100-eq. and Blackwell efficiency, training hits 5e30 FLOP.[Recent Findings]
Mitigations: Efficiency (176 kW/sq ft density), workload migration, and modular infra allow "spending out" of bottlenecks, shifting focus from scarcity to hyperscaler dominance.[1][2]
Without constraints, enterprise demand (auto, finance, healthcare) + AI agents could justify 10-30x growth, per scaling laws—though real-world adoption velocity (monetization, regulation) tempers extremes.[3][5]
Sources:
- [1] https://www.hanwhadatacenters.com/blog/power-requirements-for-ai-data-centers-resilient-infrastructure/
- [2] https://www.rcrwireless.com/20260210/infrastructure/ad-data-center-spend-power-crunch
- [3] https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030
- [4] https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
- [5] https://www.spglobal.com/en/research-insights/special-reports/look-forward/data-center-frontiers/global-ai-power-demand-challenges-opportunities
- [6] https://www.weforum.org/stories/2025/12/data-centres-and-energy-demand/
- [7] https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html