AI Data Center Power Demand Forecast Raised Again; Reliability Value Chain Remains Positively Viewed

Overview

The report addresses three core questions: how long AI and data center power demand can continue growing, whether the AI innovation cycle is shifting from investment expansion toward an efficiency phase, and whether AI's social value can offset its emissions, resource and governance costs. Goldman Sachs raises its global data center power consumption forecast and believes community permitting, grids, equipment, labor and environmental conditions will affect the pace of construction, but these constraints also reinforce long-term investment demand for reliability infrastructure.

Core views

The report first uses the “7Ps” to organize the drivers and constraints of data center power demand: penetration, productivity, price, policy, parts, people and physical environment. Based on data center capacity expansion, the TMT team's higher forecasts for AI server shipments and the U.S. power team's updated regional demand outlook, Goldman Sachs raises its forecast for global data center power consumption growth from 2025 to 2030 from 117% to 170%, with the United States contributing more than 60%. Relative to 2023, cumulative incremental power consumption through 2030 is approximately 1,100 TWh, exceeding the 2023 annual power consumption of Japan, the world's fifth-largest electricity consumer. PPA contracting volumes related to renewable energy and nuclear power technologies increased 17% year over year in 2025, and the report therefore continues to assume a power supply mix comprising natural gas, renewable energy, energy storage and nuclear power. Capital expenditure remains a direct pillar of demand expansion. Alphabet raised its 2026 capital expenditure range by $15 billion and expects capital expenditure to continue growing significantly year over year in 2027; Amazon raised its 2026 capital expenditure target by $20 billion; Meta also raised the lower end of its 2026 forecast range, which the report previously listed as $125 billion to $130 billion. Goldman Sachs analysts expect the combined capital expenditure and R&D spending of hyperscalers including Alphabet, Microsoft, Meta, Amazon, Oracle, Baidu, Tencent and Alibaba to exceed $1 trillion in 2026. However, long-term sustainability ultimately depends on whether AI can deliver tangible results that improve returns; the report expects company-level returns at some hyperscalers to decline, with reinvestment rates potentially exceeding 100% in 2027. Efficiency improvements in chips, servers and models do not necessarily reduce total power consumption. New-generation servers have higher maximum power per unit, but compute capacity is growing faster, so energy consumption per unit of compute declines. The report believes the market has not yet entered a demand-constrained phase because companies and governments remain concerned that cutting budgets could damage their competitive positions, while the scale of tokens and compute they require has not yet been fully determined. Only when demand becomes clearer and productivity gains begin to drive budget reductions will the AI cycle move from the “assessment/hopes and dreams” phase to the “execution/efficiency” phase. Industry discussions in 2Q26 indicate that customers have begun monitoring who within their organizations is using compute and spending per token, but it is not yet possible to confirm whether such optimization will offset the incremental demand induced by efficiency improvements. The report uses scenario analysis to illustrate the importance of demand elasticity: assume a customer originally planned to purchase 10 AI servers, while a new-generation product offers 10 times the compute capacity, 6 times the maximum power and 5 times the price. In the budget-constrained scenario, the customer purchases 2 units, leaving the budget unchanged while increasing compute capacity by 100% and maximum power by 20%; in the demand-constrained scenario, the customer purchases only 1 unit, leaving compute capacity unchanged while reducing the budget by 50% and maximum power by 40%; in the unconstrained scenario, the customer still purchases 10 units, increasing the budget by 400%, compute capacity by 900% and maximum power by 500%. This shows that whether technological efficiency reduces total energy consumption depends not on per-unit efficiency, but on whether customers use the efficiency dividend to cut budgets or expand compute consumption. Electricity prices are still not a major constraint for hyperscalers themselves. Goldman Sachs believes their strong balance sheets enable them to absorb higher electricity costs and sign take-or-pay contracts. Around-the-clock low-carbon reliable power supply in the United States carries a “green reliability premium” relative to combined-cycle natural gas, but this remains moderate compared with hyperscalers' EBITDA and corporate returns and is insufficient to prevent the combined use of natural gas, renewable energy, energy storage and nuclear power. Low U.S. natural gas prices make combined-cycle generation more advantageous, while Europe, Japan, South Korea and regions dependent on imported LNG lack comparable conditions; if the capital cost of combined-cycle natural gas projects exceeds the base-case assumption, this premium would narrow accordingly. The expected timeline for the power supply mix is greater reliance on renewable energy, storage and simple-cycle natural gas in the short term, increased combined-cycle natural gas in the medium term, and the introduction of nuclear power over the long term. Policy and community permitting may slow construction more directly than electricity prices. Texas accounts for 13% of expected incremental U.S. data center capacity through 2030, and its governor issued an executive order on August 3 requiring certain projects to undergo audits and data disclosure before receiving grid-connection approval; Virginia accounts for 17% of expected incremental capacity, and bipartisan calls have emerged locally for a construction moratorium to confirm the adequacy of power supply and grid-connection capacity. Community concerns include the reliability of power and water supplies, consumer electricity bills, noise, water consumption and localized warming. Potential solutions include allowing data centers to curtail loads when the grid is strained, adding backup power supplies or backup data centers, using take-or-pay contracts to limit cost spillovers, deploying closed-loop liquid cooling, installing soundproofing facilities and site buffer zones, and utilizing waste heat where economically viable. However, the report emphasizes that technological solutions may not fully eliminate community concerns regarding AI's social benefits, project transparency and NIMBY effects. Equipment, personnel and grid access jointly constrain execution speed and drive behind-the-meter power supply. GE Vernova stated that, by the end of 2026, it expects more than 50% of its 2031 capacity to be contractually committed. Goldman Sachs expects total U.S. power demand to grow at a 3.5% CAGR through 2030, with 0.8 percentage points coming from behind-the-meter power supply; due to time-to-market requirements and grid-access timelines, predominantly natural-gas-based behind-the-meter solutions are expected to contribute approximately 30% of data center growth. Even if projects are ultimately connected to the grid, some behind-the-meter facilities may be retained as backup power. To meet overall U.S. power demand growth, the report estimates that more than 500,000 additional jobs will be required, or more than 700,000 including Europe; more than 200,000 U.S. jobs are related to transmission and distribution connections, and even if all energy-related apprentices were directed toward transmission and distribution, the base case would still have a labor shortfall of 78,000 by 2030, with electrician shortages especially pronounced. The physical environment will erode some efficiency improvements. Goldman Sachs estimates that 56% of new data centers globally and 55% in the United States are located in areas with elevated risks from high temperatures, high humidity or drought. An estimated 43% of new capacity will use direct-to-chip liquid cooling or immersion cooling inside server rooms, while 56% will use chillers, adiabatic systems or mechanical backup cooling to dissipate heat externally. The United States places greater emphasis on reducing water consumption, even if this increases electricity consumption; China uses minimum efficiency requirements to constrain PUE, which may result in greater water consumption. The report therefore incorporates environmental factors into its baseline power forecast, equivalent to increasing PUE by 5 percentage points globally and 11 percentage points in the United States, partially offsetting efficiency improvements in server and cooling technologies. On sustainability, the report advocates assessing AI based on net impact: the potential social value created by using AI to advance sustainable development goals, less the social costs arising from emissions, resource footprints, labor and other risks. Investors also need to focus on AI development, risk management and compliance, governance and implementation mechanisms. Although AI's ultimate net social value still requires ongoing measurement, Goldman Sachs maintains a positive investment outlook on the reliability theme and the data center power ecosystem, believing that growth in AI power consumption and the buildout of global redundancy capacity to guard against power, water, supply chain, network and operational disruptions will jointly support demand for power generation equipment, transmission utilities, electrical components and cooling solutions.

Analysis framework

Goldman Sachs first uses the “7Ps” framework to identify demand growth and execution constraints, and then combines AI server shipments, data center capacity, regional power forecasts and hyperscaler capital expenditure to raise its long-term power consumption forecast. It subsequently uses three budget and demand scenarios for server technology iteration to determine whether efficiency improvements will reduce or increase total energy consumption, and then uses the shale innovation cycle as a reference to assess whether AI is shifting from investment-led exploration toward execution efficiency. On the supply side, it compares the levelized costs of different technologies, natural gas price sensitivity and the green reliability premium, while incorporating policy, community, equipment, labor, cooling and geographical conditions into its analysis of project execution. Finally, it evaluates AI sustainability using a net-impact approach that subtracts social costs from potential social value.

Methodology notes

• Industry/Sector Analysis Framework

  Data Center Power Demand “7Ps” Framework

  The report simultaneously assesses the drivers of power demand growth and construction constraints across seven dimensions: penetration, productivity, price, policy, parts, people and physical environment.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Budget-Constrained, Demand-Constrained and Unconstrained Scenario Analysis

  The report varies the number of servers purchased and budget responses to compare the different effects of technological advances on compute capacity, spending and maximum power, thereby assessing whether improved unit efficiency translates into lower total power consumption.

• Cycle and Business Conditions Framework

  Analogy Between AI and Shale Innovation Cycles

  The report draws on phase-based indicators from the shale innovation cycle to assess whether AI is shifting from a capital-intensive assessment phase toward an execution phase emphasizing budget discipline and capital efficiency.

• Industry/Sector Analysis Framework

  Comparison of Levelized Cost of Energy and Green Reliability Premium

  The report compares the costs of power supply portfolios including combined-cycle natural gas, nuclear power, renewable energy, energy storage and peaking gas generation, and tests the effects of changes in natural gas prices and capital costs on the premium for around-the-clock low-carbon reliable power.

• Industry/Sector Analysis FrameworkUpstream, Midstream and Downstream Value-Chain Transmission

  Transmission of Data Center Load to the Power Ecosystem

  The report sequentially maps incremental data center load to demand for power generation capacity, transmission and distribution connections, behind-the-meter power supply, electrical components, cooling equipment and skilled labor.

• (Out-of-Vocabulary Method)

  Assessment of AI's Net Social Sustainability Impact

  The report uses the potential social value of AI in advancing sustainable development goals less the social costs associated with emissions, resource footprints, labor and governance risks as the basic approach for assessing AI's sustainability impact.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AI/Data Center Power Ecosystem and Reliability Theme

  Growth in AI power consumption and the buildout of redundancy for power, water, supply chains, networks and operations are the core demand drivers for the theme.

  Strengths

  The long-term data center power consumption forecast has been raised, and hyperscalers retain strong capacity to fund capital expenditure and absorb electricity costs.

  Weaknesses

  Project construction is highly dependent on approvals, grid connections, equipment delivery and skilled labor.

  Comparison

  The report believes the green reliability premium is moderate relative to hyperscalers' EBITDA and corporate returns.

  Risks

  AI returns falling short of expectations, budgets shifting toward efficiency, community opposition and infrastructure bottlenecks may slow growth.

• Power Generation Equipment and Power Supply Solutions

  Incremental data center load requires various forms of additional power supply capacity, including natural gas, renewable energy, energy storage and nuclear power.

  Strengths

  The report expects approximately 139 GW of additional capacity to be required through 2030, while long equipment lead times also improve the visibility of contracted capacity.

  Weaknesses

  Equipment such as gas turbines has long supply lead times, and rising capital costs may alter the relative economics of the technology mix.

  Comparison

  Low U.S. natural gas prices give combined-cycle generation a cost advantage over Europe, Japan, South Korea and LNG-importing regions.

  Risks

  Permitting, equipment delivery and community concerns regarding emissions and reliability may delay projects.

• Transmission-Related Utilities and Grid Infrastructure

  Growth in data center load requires additional transmission and distribution connections, upgrades to aging infrastructure and greater grid resilience.

  Strengths

  Total U.S. power demand is expected to grow at a 3.5% compound rate through 2030, and demand for more than 200,000 additional jobs is related to transmission and distribution connections.

  Weaknesses

  Grid-connection timelines are long, and there are significant shortages of skilled workers such as electricians.

  Comparison

  The grid remains hyperscalers' preferred option, but behind-the-meter power supply is expected to support approximately 30% of data center growth because it enables faster time to market.

  Risks

  Rising consumer bills, disputes over cost allocation and community concerns about reliability may affect permitting and the pace of investment.

• Cooling Solutions and Electrical Components

  Higher power density, water-saving requirements, backup power and harsh environments drive demand for liquid cooling, chillers, soundproofing, electrical connections and backup systems.

  Strengths

  An estimated 43% of new capacity will use direct-to-chip liquid cooling or immersion cooling, while 56% will require external heat-dissipation cooling solutions.

  Weaknesses

  There is a trade-off between water and electricity conservation, and some cooling solutions increase PUE and total power consumption.

  Comparison

  The United States prioritizes reducing water consumption even if electricity consumption rises; China's minimum PUE requirements may lead to greater water consumption.

  Risks

  Technology choices are constrained by local temperature, humidity, drought, noise regulations and community acceptance.

Key data

• Global Data Center Power Consumption Growth170% growth in 2030 versus 2025Previous forecast was 117%
• Cumulative Incremental Power Consumption Relative to 20231,100 TWhThe incremental amount through 2030 exceeds Japan's total power consumption in 2023
• U.S. Contribution to Incremental DemandMore than 60%Share of global data center power consumption growth
• Renewable Energy and Nuclear Power PPA Contracting Volume17% year-over-year growth2025
• Hyperscaler Capital Expenditure and R&D SpendingMore than $1 trillionGoldman Sachs analysts' aggregate forecast for 2026E
• Alphabet Capital Expenditure AdjustmentRaised by $15 billion2026 capital expenditure range
• Amazon Capital Expenditure AdjustmentRaised by $20 billion2026 capital expenditure target
• Budget-Constrained Server ScenarioCompute capacity +100%, maximum power +20%, budget unchangedWhen new-generation servers offer 10 times the compute capacity, 6 times the power and 5 times the price, purchases decline from 10 units to 2 units
• Demand-Constrained Server ScenarioCompute capacity unchanged, budget -50%, maximum power -40%Purchases decline from 10 units to 1 unit
• Unconstrained Server ScenarioBudget +400%, compute capacity +900%, maximum power +500%The customer still purchases 10 new-generation servers
• Consumer Electricity Bill ForecastUnited States +3% to +6% per year; Europe +2% to +4% per yearU.S. forecast period through 2029 and Europe through 2030
• Incremental Power Generation Capacity Required for Data Center LoadApproximately 139 GWThrough 2030
• U.S. Total Power Demand Growth3.5% CAGRThrough 2030, with 0.8 percentage points from behind-the-meter power supply
• Behind-the-Meter Power Supply ContributionApproximately 30%Expected share of U.S. data center growth through 2030
• Incremental Employment DemandMore than 500,000 in the United States; more than 700,000 including EuropeRequired to meet overall power demand growth, with more than 200,000 U.S. jobs related to transmission and distribution connections
• Transmission and Distribution Labor Shortfall78,000 peopleThe shortfall that would remain under the 2030 base case even if all energy-related apprentices shifted to transmission and distribution
• New Data Centers Facing Elevated Physical Environment RisksGlobal 56%; United States 55%Risks from high temperatures, high humidity or drought
• Cooling Methods for New CapacityIn-room liquid cooling or immersion cooling 43%; external heat-dissipation cooling 56%External heat dissipation includes chillers, adiabatic systems or mechanical backup cooling
• Impact of Environmental Factors on PUEGlobal +5 percentage points; United States +11 percentage pointsUsed in the baseline power forecast to reduce the magnitude of annual efficiency improvements
• Gas Turbine Capacity ContractedMore than 50%GE Vernova expects that by the end of 2026, more than half of its 2031 capacity will already be contracted

Impact & implications

The report argues that growth in AI power consumption is not determined solely by the number of servers, but by the combined effects of capital expenditure, efficiency per unit of compute, demand elasticity and infrastructure executability. Even if technology continues to reduce energy consumption per unit of compute, backlogged demand, behind-the-meter power supply, backup capacity and water-saving cooling may still drive up total demand for power and infrastructure; meanwhile, community permitting and resource constraints may alter project locations, timelines and power supply structures. Accordingly, AI demand and the buildout of global redundancy will continue to support the reliability value chain, but the pace of actual delivery depends on project approvals, grid access, equipment delivery and labor supply.

Risks

• Opposition from U.S. communities regarding power and water reliability, electricity bills, noise, water consumption and localized warming may delay or reduce the scale of data center projects.
• If the AI cycle enters the execution and efficiency phase, productivity gains may prompt customers to reduce budgets, weakening growth in server and power demand.
• Declining company-level returns among hyperscalers and reinvestment rates exceeding 100% in 2027 may raise doubts about the sustainability of capital expenditure.
• Shortages of power generation equipment, grid connections and skilled labor may constrain project execution, with a baseline transmission and distribution labor shortfall of 78,000 people.
• High temperatures, high humidity and drought may constrain cooling technology choices and offset some energy-efficiency improvements by increasing PUE.
• Higher electricity prices, capital costs for natural gas generation and consumer bills may intensify disputes over cost allocation and community permitting.
• Emissions, resource footprints, labor and governance risks may weaken AI's net contribution to sustainable development goals.

What to watch

• Track hyperscalers' 2027 capital expenditure, reinvestment rates and whether AI-related corporate returns can support continued investment.
• Monitor when token and compute demand becomes fully defined and whether customers shift from expanding consumption to reducing budgets.
• Watch Texas audit and disclosure requirements, calls for a construction moratorium in Virginia and other changes in community permitting.
• Track changes in natural gas prices, the capital costs of combined-cycle projects and the green reliability premium.
• Monitor contracting progress for equipment such as gas turbines, grid-access timelines and the share of behind-the-meter power supply in incremental data center growth.
• Watch whether shortages of electricians and transmission and distribution workers affect U.S. power infrastructure construction.
• Track adoption rates for liquid cooling, water-saving cooling and waste heat utilization, as well as the effects of environmental conditions on PUE and total power consumption.
• Monitor investors' areas of focus regarding AI development, risk management, compliance, governance and measurement of net social value.