Technology Report

AI Data Center Boom: Can We Build It If They Come?

Meeting AI compute demand will require system-level solutions, not the site-by-site workarounds plugging gaps today.

By Tatum Quinn, Peter Hanbury, David Crawford, Aaron Denman, Padraic Brick, and Alessandro Cannarsi

First published on septiembre 29, 2026
min read


Executive Summary Data centers have become critical infrastructure; Bain projects $5 trillion to $6.5 trillion of buildout, adding about 150 gigawatts or more by 2030. That would nearly triple global capacity over five years, but power, chips, skilled labor, and permitting are all constrained at once. Decisions made today set capacity several years from now, and the shortage won't resolve on its own. Closing the gap will take structural moves (new power and technology, public-private coordination, resource sharing), reshaping the industry. This article is part of Bain’s Technology Report 2026 Explore the report Data centers powering AI have rapidly become a new layer of critical infrastructure worldwide that are central to technology innovation, economic growth, and national sovereignty. US capital investment in data centers as a share of GDP now rivals what it spent building its railroads, telecom networks, and electrical grid: generational projects. The data center boom is moving faster. Five years ago, a 50-megawatt (MW) facility was considered large. Now, hyperscalers are planning to build data center campuses 100 times bigger: 5 gigawatts (GW) or more, costing $150 billion to $200 billion each. AI compute demand is already outrunning supply, and the shortage is intensifying. To overcome it, the tech industry is planning investments at unprecedented scale and speed. In a base case, Bain’s Data Center Model projects approximately $5 trillion to $6.5 trillion of spending to build nearly 150 GW of new compute capacity by 2030 (see Figure 1), almost tripling global capacity in just five years. Most of the growth will come from the US, but compute capacity will grow by double digits across regions. Figure 1 Data center critical IT capacity is expected to grow by 20% to 30% over the next five years Note: Reflects net growth in installed base; excludes GPU refresh, retrofit, and replacement of retired capacity Source: Bain Data Center Model, July 2026 What many business leaders are missing is that the AI compute shortage is unlikely to resolve on its own. The usual assumption that supply chains respond to demand and that they’ll eventually balance doesn’t hold here. Power, hardware, and skilled labor are simultaneously constrained (see Figure 2), and public approvals are increasingly difficult to secure. Lead times are long enough that decisions made today determine capacity several years from now. Figure 2 Data center expansion will accelerate demand for resources across the supply chain Note: Cooling market data excludes China Sources: Bain Data Center Model, July 2026; Bain AI Chips Model, July 2026; TD Cowen; Bain analysis Power: Grid connection is delaying and freezing projects. Adding grid capacity takes four-plus years, and critical equipment inventory (e.g., transformers, etc.) is scarce. Onsite, behind-the-meter power relieves some of the bottleneck, but it adds regulatory exposure, longer equipment lead times, and more project complexity. Semiconductors: GPUs, memory, optical networking, and other components are being locked in years ahead. Operators without long-term supply agreements are effectively out of the queue. Labor: Skilled trades are the human version of the same problem, with too few electrical workers, mechanical tradespeople, and cooling specialists. Meeting demand means recruiting, training, and retaining talent far above historical rates. Regulations and public opinion: Complex permitting rules across local, state, and national levels add time and uncertainty, even when primary inputs are secured. Citizens and politicians are increasingly pushing back regarding water use, energy prices, and noise. Local opposition blocked or delayed at least 75 projects worth $130 billion in the first quarter of this year alone, nearly matching the $156 billion worth of projects disrupted in all of 2025. In July, New York became the first US state to pause construction of the largest data centers; 14 others have introduced similar legislation. Countries including Ireland, the Netherlands, and Denmark have restricted or paused new grid connections, at least temporarily. Structural solutions are required Individual projects are finding workarounds, but they can’t be replicated at scale, and some are controversial. Colossus (from xAI, now part of SpaceXAI) was built in under a year by repurposing a factory and leasing dozens of gas turbines without federal clean air permits, drawing a lawsuit. Constructing a 1 GW data center typically takes one to four years, even on sites with existing power (see Figure 3), and it can take longer when new generation and transmission infrastructure is needed. Figure 3 Building a 1 GW data center typically takes one to four years, even on sites with existing power Source: Epoch AI Meeting the global ambition instead will require system-level fixes outside the usual free-market playbook. A grand challenge on this scale demands grand solutions, and executives planning on business as usual will get blindsided. The forces below are deliberately provocative and mostly long term, and some will create new challenges, even as they mitigate others. But overcoming the shortage will take at least one of the following forces: consolidation, new power at scale, orbital data centers, government investment, and social license. Consolidation: Consolidation could occur as capital requirements grow so large that only a handful of companies survive. Hyperscalers’ data center capex could top $1 trillion by 2028. Semiconductors provide a useful corollary. Staying at the leading edge grew so costly (tens of billions of dollars per generation) that missing one generation made catching up effectively impossible. Companies exited, were acquired, or retreated to older nodes (see Figure 4). Figure 4 Most semiconductor manufacturers have exited the market for the most advanced logic node, leaving TSMC, Samsung, Intel, and Rapidus Note: Does not include companies that entered the node after the logic node leadership period Source: Bain analysis Facing unsustainable costs, data center providers may also turn to once-unthinkable solutions: shared power procurement, colocated servers, and jointly funded grid upgrades. New power at scale: Power is one of the most challenging, multidimensional bottlenecks, and no single source solves it. The realistic path combines near-term options that could come online this decade with long-term ones that are scheduled to pay off in the 2030s. Near term, behind-the-meter generation (mostly natural gas today) is one of the few ways to add large blocks of power on data center timelines. Pairing it with carbon capture, utilization, and storage could make it cleaner, but that’s unproven at scale. Tech companies are also turning to renewables, storage, and virtual power plants (VPPs) that use software to aggregate and redirect power in real time from distributed sources such as batteries and smart thermostats. Google, for instance, is funding 1.6 GW of renewables (mostly wind) and 300 MW of storage alongside a new Minnesota data center. North America’s VPP market grew by 13.7% last year, to 37.5 GW, per Wood Mackenzie. Novel VPP agreements are emerging, such as the Google-Voltus deal that would aggregate up to 100 MW annually over three years in the PJM grid. Nuclear is another clean source that could eventually help, but new, large-scale capacity realistically won’t arrive until the mid-2030s. Recommissioning existing facilities (as at the Palisades and Three Mile Island power plants in the US) and building new traditional nuclear reactors would require national coordination of schedules, supply chains, the workforce, regulations, and financing. Small modular reactors are a credible piece of the puzzle, but likely can’t move fast enough or scale alone. Google (via Kairos Power) and Amazon (via X-energy) are among those betting on this technology. Orbital data centers: Closing the gap may require technologies that don’t fully exist yet. Orbital data centers could theoretically address land, cooling, and power constraints. Similar to nuclear, they’re long-horizon bets, not near-term relief. SpaceXAI’s post-IPO valuation reflects investor conviction that space-based compute will eventually matter, but timelines and economics remain unproven. Government investment: Sovereign entities are already investing directly in AI infrastructure and chips, including in the US, South Korea, UAE, Saudi Arabia, and EU. Governments may move beyond subsidies toward a direct role, such as coordinator, financier, or (in some countries) even owner. But government interventions can be double edged. A coordinated, Manhattan Project–style push might accelerate builds. For example, China’s government is investing $295 billion in AI data center construction and $740 billion in grid upgrades to pipe renewable power from its resource-rich west to data centers. But the same interventions that can unlock power and land can introduce control, oversight, and delays, and companies and investors will need to plan around that risk. Social license: This will be as decisive as build capability. Communities and governments will increasingly demand binding agreements on data use, environmental impact, and governance before granting permits. Frameworks are emerging. In the US, the city of Lancaster, Pennsylvania, signed a contract with Chirisa Technology Parks and Machine Investment Group pursuant to an agreement to codevelop a data center for CoreWeave that caps its water use and sets clean energy, noise, emissions, and local hiring requirements. Implications for executives Enterprise technology leaders: Treat compute access as both a supply chain risk and a strategic asset. Those who wait for the market to sort itself out will get locked out. AI providers, hyperscalers, and data center developers: Because social license and power access are critical to each project’s success, building a larger portfolio of carefully chosen and managed sites is now imperative. Regulators: The window to set coherent rules is open, but it won't stay that way for long. Absent clear frameworks, government involvement won’t shrink; it would likely grow but on a crisis timeline, when leverage is gone. Investors: Capital needs for data infrastructure will remain high, and with a growing field of firms building and operating these assets, assessing counterparty risk and end-customer exposure is essential. Bottlenecks in power, semiconductors, and other inputs carry large capital needs of their own, opening additional entry points for investors. And as sovereign infrastructure becomes a bigger part of national strategies, partnerships offer both a way in and geographic diversification. Power providers (utilities and independent power producers): Partnering earlier with AI providers, hyperscalers, and developers is critical to both shape power requirements and jointly craft creative solutions to meet system, community, and affordability needs. Read the Next Section The $100-Billion SaaS Opportunity Hiding in Cross-System Labor More from the report New Innovation Is Required to Fund AI’s $6 Trillion Buildout Software Investing in the Age of AI and Slower Growth AI Data Center Boom: Can We Build It If They Come? La oportunidad SaaS de 100 mil millones de dólares que está escondida en el trabajo de coordinación entre sistemas Hardware Strikes Back in the AI Era Tech Services: Cracking the Code on AI-led Growth Cybersecurity’s New AI Imperative: Attacking the Backlog of Vulnerability Alerts The AI-Native Enterprise: Absorption Is the New Advantage The Half-Finished Redesign: How AI Reshapes Software Organizations Managing Token Spending without Choking Off Opportunity The Missing Architecture for Agentic Software Development How AI Is Changing Data Monetization Read our Technology Report 2026 EXPLORE THE FULL REPORT DOWNLOAD THE PDF

Executive Summary

Data centers have become critical infrastructure; Bain projects $5 trillion to $6.5 trillion of buildout, adding about 150 gigawatts or more by 2030.
That would nearly triple global capacity over five years, but power, chips, skilled labor, and permitting are all constrained at once.
Decisions made today set capacity several years from now, and the shortage won't resolve on its own.
Closing the gap will take structural moves (new power and technology, public-private coordination, resource sharing), reshaping the industry.

Data centers powering AI have rapidly become a new layer of critical infrastructure worldwide that are central to technology innovation, economic growth, and national sovereignty.

US capital investment in data centers as a share of GDP now rivals what it spent building its railroads, telecom networks, and electrical grid: generational projects. The data center boom is moving faster. Five years ago, a 50-megawatt (MW) facility was considered large. Now, hyperscalers are planning to build data center campuses 100 times bigger: 5 gigawatts (GW) or more, costing $150 billion to $200 billion each.

AI compute demand is already outrunning supply, and the shortage is intensifying. To overcome it, the tech industry is planning investments at unprecedented scale and speed. In a base case, Bain’s Data Center Model projects approximately $5 trillion to $6.5 trillion of spending to build nearly 150 GW of new compute capacity by 2030 (see Figure 1), almost tripling global capacity in just five years. Most of the growth will come from the US, but compute capacity will grow by double digits across regions.

Line chart tracking the global installed base of data center IT power from 2020 to 2030, splitting into three growth scenarios from 2025 onward. The base case rises from 79 GW in 2025 to an estimated 228 GW by 2030, with the high case reaching 262 GW and the low case 188 GW. — Data center critical IT capacity is expected to grow by 20% to 30% over the next five years

What many business leaders are missing is that the AI compute shortage is unlikely to resolve on its own. The usual assumption that supply chains respond to demand and that they’ll eventually balance doesn’t hold here.

Power, hardware, and skilled labor are simultaneously constrained (see Figure 2), and public approvals are increasingly difficult to secure. Lead times are long enough that decisions made today determine capacity several years from now.

Three bar charts tracking AI data-center statistics projections from 2023 to 2030. Global electricity demand rises from 388 to 1,228 terawatt-hours, cooling-technology revenue grows from $7.9 billion to $28.4 billion, and GPU/ASIC shipments increase from 3.9 million to 38.6 million units. — Data center expansion will accelerate demand for resources across the supply chain

Power: Grid connection is delaying and freezing projects. Adding grid capacity takes four-plus years, and critical equipment inventory (e.g., transformers, etc.) is scarce. Onsite, behind-the-meter power relieves some of the bottleneck, but it adds regulatory exposure, longer equipment lead times, and more project complexity.
Semiconductors: GPUs, memory, optical networking, and other components are being locked in years ahead. Operators without long-term supply agreements are effectively out of the queue.
Labor: Skilled trades are the human version of the same problem, with too few electrical workers, mechanical tradespeople, and cooling specialists. Meeting demand means recruiting, training, and retaining talent far above historical rates.
Regulations and public opinion: Complex permitting rules across local, state, and national levels add time and uncertainty, even when primary inputs are secured. Citizens and politicians are increasingly pushing back regarding water use, energy prices, and noise. Local opposition blocked or delayed at least 75 projects worth $130 billion in the first quarter of this year alone, nearly matching the $156 billion worth of projects disrupted in all of 2025. In July, New York became the first US state to pause construction of the largest data centers; 14 others have introduced similar legislation. Countries including Ireland, the Netherlands, and Denmark have restricted or paused new grid connections, at least temporarily.

Structural solutions are required

Individual projects are finding workarounds, but they can’t be replicated at scale, and some are controversial. Colossus (from xAI, now part of SpaceXAI) was built in under a year by repurposing a factory and leasing dozens of gas turbines without federal clean air permits, drawing a lawsuit. Constructing a 1 GW data center typically takes one to four years, even on sites with existing power (see Figure 3), and it can take longer when new generation and transmission infrastructure is needed.

Gantt-style timeline tracking three build milestones — construction start, first operational date, and reaching 1 GW of power — for six named AI data center projects from 2021 to 2028. Total build time from construction start to 1 GW ranges from 1 year for SpaceXAI's Colossus 2 to 3.6 years for Microsoft's Fairwater and Amazon's Ridgeland. — Building a 1 GW data center typically takes one to four years, even on sites with existing power

Meeting the global ambition instead will require system-level fixes outside the usual free-market playbook. A grand challenge on this scale demands grand solutions, and executives planning on business as usual will get blindsided. The forces below are deliberately provocative and mostly long term, and some will create new challenges, even as they mitigate others. But overcoming the shortage will take at least one of the following forces: consolidation, new power at scale, orbital data centers, government investment, and social license.

Consolidation: Consolidation could occur as capital requirements grow so large that only a handful of companies survive. Hyperscalers’ data center capex could top $1 trillion by 2028.

Semiconductors provide a useful corollary. Staying at the leading edge grew so costly (tens of billions of dollars per generation) that missing one generation made catching up effectively impossible. Companies exited, were acquired, or retreated to older nodes (see Figure 4).

Bar chart tracking the number of semiconductor companies manufacturing the most advanced logic node over multi-year periods from 2002 to today. The count falls steadily from 25 companies in 2002–2003 to just 4 in 2020-today. — Most semiconductor manufacturers have exited the market for the most advanced logic node, leaving TSMC, Samsung, Intel, and Rapidus

Facing unsustainable costs, data center providers may also turn to once-unthinkable solutions: shared power procurement, colocated servers, and jointly funded grid upgrades.

New power at scale: Power is one of the most challenging, multidimensional bottlenecks, and no single source solves it. The realistic path combines near-term options that could come online this decade with long-term ones that are scheduled to pay off in the 2030s.

Near term, behind-the-meter generation (mostly natural gas today) is one of the few ways to add large blocks of power on data center timelines. Pairing it with carbon capture, utilization, and storage could make it cleaner, but that’s unproven at scale.

Tech companies are also turning to renewables, storage, and virtual power plants (VPPs) that use software to aggregate and redirect power in real time from distributed sources such as batteries and smart thermostats. Google, for instance, is funding 1.6 GW of renewables (mostly wind) and 300 MW of storage alongside a new Minnesota data center. North America’s VPP market grew by 13.7% last year, to 37.5 GW, per Wood Mackenzie. Novel VPP agreements are emerging, such as the Google-Voltus deal that would aggregate up to 100 MW annually over three years in the PJM grid.

Nuclear is another clean source that could eventually help, but new, large-scale capacity realistically won’t arrive until the mid-2030s. Recommissioning existing facilities (as at the Palisades and Three Mile Island power plants in the US) and building new traditional nuclear reactors would require national coordination of schedules, supply chains, the workforce, regulations, and financing.

Small modular reactors are a credible piece of the puzzle, but likely can’t move fast enough or scale alone. Google (via Kairos Power) and Amazon (via X-energy) are among those betting on this technology.

Orbital data centers: Closing the gap may require technologies that don’t fully exist yet. Orbital data centers could theoretically address land, cooling, and power constraints. Similar to nuclear, they’re long-horizon bets, not near-term relief. SpaceXAI’s post-IPO valuation reflects investor conviction that space-based compute will eventually matter, but timelines and economics remain unproven.

Government investment: Sovereign entities are already investing directly in AI infrastructure and chips, including in the US, South Korea, UAE, Saudi Arabia, and EU.

Governments may move beyond subsidies toward a direct role, such as coordinator, financier, or (in some countries) even owner. But government interventions can be double edged. A coordinated, Manhattan Project–style push might accelerate builds. For example, China’s government is investing $295 billion in AI data center construction and $740 billion in grid upgrades to pipe renewable power from its resource-rich west to data centers. But the same interventions that can unlock power and land can introduce control, oversight, and delays, and companies and investors will need to plan around that risk.

Social license: This will be as decisive as build capability. Communities and governments will increasingly demand binding agreements on data use, environmental impact, and governance before granting permits.

Frameworks are emerging. In the US, the city of Lancaster, Pennsylvania, signed a contract with Chirisa Technology Parks and Machine Investment Group pursuant to an agreement to codevelop a data center for CoreWeave that caps its water use and sets clean energy, noise, emissions, and local hiring requirements.

Implications for executives

Enterprise technology leaders: Treat compute access as both a supply chain risk and a strategic asset. Those who wait for the market to sort itself out will get locked out.
AI providers, hyperscalers, and data center developers: Because social license and power access are critical to each project’s success, building a larger portfolio of carefully chosen and managed sites is now imperative.
Regulators: The window to set coherent rules is open, but it won't stay that way for long. Absent clear frameworks, government involvement won’t shrink; it would likely grow but on a crisis timeline, when leverage is gone.
Investors: Capital needs for data infrastructure will remain high, and with a growing field of firms building and operating these assets, assessing counterparty risk and end-customer exposure is essential. Bottlenecks in power, semiconductors, and other inputs carry large capital needs of their own, opening additional entry points for investors. And as sovereign infrastructure becomes a bigger part of national strategies, partnerships offer both a way in and geographic diversification.
Power providers (utilities and independent power producers): Partnering earlier with AI providers, hyperscalers, and developers is critical to both shape power requirements and jointly craft creative solutions to meet system, community, and affordability needs.