AI-Infrastructure Report
Top-bottom analysis of the mega-trend.
September 2025
Ally Whitney, Jordan Yip for the America 2030 sector research.
Executive Summary
The United States dominates global AI infrastructure, accounting for about 44% of worldwide data center power in 2024 (~53.7 GW of 122 GW). This lead is underpinned by massive investments from hyperscale cloud providers (AWS, Microsoft, Google, Meta) and semiconductor firms, creating unmatched AI computing power. Since 2020, the market has expanded rapidly, driven by generative AI demand, enterprise migration to cloud, and government policy. Spending on AI infrastructure in 2025, including data centers, semiconductors, networks, and energy, is at record highs and expected to maintain strong growth through 2030. McKinsey projects nearly $7 trillion in global data center investment by 2030, with the U.S. capturing more than 40%. U.S. capacity demand is growing 20–25% annually, with Northern Virginia alone representing ~13% of global capacity. Strengths include leadership in cloud hyperscalers, semiconductor design, venture-backed startups, capital availability, and supportive policy (CHIPS Act). Weaknesses are concentrated in power-grid constraints, long interconnection delays, reliance on overseas fabs for advanced chips, shortages of GPUs and high-bandwidth memory, talent gaps, and environmental opposition to water and energy intensive data centers. The boom in AI infrastructure isn’t just about building bigger and more robust data centers. It’s running into very real limits with power, water, and land. Getting connected to the grid can take years, key electrical equipment is hard to source, and water use in dry regions is already controversial. If AI demand keeps doubling, data centers could end up using as much power as an entire country like Japan. Even though new cooling systems and chips are getting more efficient, history shows that efficiency usually just leads to more consumption, not less. Outlook is positive under a base case, with AI workloads projected to account for over one-quarter of data center demand by 2027 and drive a 165% increase in U.S. data center power use by 2030.
Market Overview
Between 2020 and 2025, AI infrastructure in the United States expanded rapidly as companies moved workloads to the cloud and generative AI emerged as a major driver of demand. By 2025, about one-third of all global data center capacity was already dedicated to AI workloads. Looking ahead, that share could approach 70% by 2030 as AI-native applications spread across industries. McKinsey estimates nearly 6.7 trillion dollars will be spent on global data centers by 2030, with more than 2.7 trillion of that in the U.S. Domestic demand is expected to keep rising 20 to 25% each year, putting continuous pressure on chip supply, the power grid, and available land. The main drivers of this growth include the massive computing needs of training large language models, the scaling of inference workloads as applications move into production, the broader digitalization of enterprises, and federal and state incentives for semiconductors and energy infrastructure.
U.S. vs. Global Peers
The U.S. and China together account for ~70% of worldwide AI infrastructure capacity. China has launched aggressive state-led programs and investments from Baidu, Alibaba, and Tencent, but U.S. export controls on advanced GPUs and semiconductor manufacturing tools hinder access to state-of-the-art hardware. Domestic Chinese alternatives exist but trail in performance. Europe had ~11.9 GW of data center capacity in 2024. The region is expanding through the EU Chips Act and AI Act, but is likely to remain secondary to the U.S. and China by 2030. Energy price volatility, regulation, and fragmented markets slow development. The Middle East, especially Saudi Arabia and the UAE, is rapidly emerging as a third AI hub. Sovereign funds and national strategies are enabling vast purchases of GPUs, with reports that UAE’s G42 contracted for up to 500,000 top-tier GPUs annually. These countries are seeking to diversify economies and build digital infrastructure. Japan and South Korea, with ~6.6 GW combined, leverage strong electronics industries and government support. India (~3.6 GW) is scaling data center capacity in line with its digitalization push and regulatory requirements for data localization. Collectively, U.S. -aligned allies (EU, Japan, South Korea, India) exceed China’s capacity, creating an advantage for U.S. -centered coalitions.
U.S. vs. Global Peers The U.S. and China together account for ~70% of worldwide AI infrastructure capacity. China has launched aggressive state-led programs and investments from Baidu, Alibaba, and Tencent, but U.S. export controls on advanced GPUs and semiconductor manufacturing tools hinder access to state-of-the-art hardware. Domestic Chinese alternatives exist but trail in performance. Europe had ~11.9 GW of data center capacity in 2024. The region is expanding through the EU Chips Act and AI Act, but is likely to remain secondary to the U.S. and China by 2030. Energy price volatility, regulation, and fragmented markets slow development. The Middle East, especially Saudi Arabia and the UAE, is rapidly emerging as a third AI hub. Sovereign funds and national strategies are enabling vast purchases of GPUs, with reports that UAE’s G42 contracted for up to 500,000 top-tier GPUs annually. These countries are seeking to diversify economies and build digital infrastructure. Japan and South Korea, with ~6.6 GW combined, leverage strong electronics industries and government support. India (~3.6 GW) is scaling data center capacity in line with its digitalization push and regulatory requirements for data localization. Collectively, U.S. -aligned allies (EU, Japan, South Korea, India) exceed China’s capacity, creating an advantage for U.S. -centered coalitions.
Core Infrastructure
Components Compute
(Chips and Accelerators) GPUs are still the backbone of AI. NVIDIA dominates with nearly 98% of the data center GPU market, thanks to its CUDA (Compute Unified Device Architecture) software and high-speed networking. Its A100, H100, and upcoming Blackwell chips are central to U.S. leadership. AMD (MI300 series) and Intel (Gaudi, Falcon Shores) are pushing alternatives, while hyperscalers are building custom chips like Google’s TPU v6, AWS’s Trainium and Inferentia, Microsoft’s Maia, and Meta’s in-house designs. Startups such as Cerebras, SambaNova, Groq, and Graphcore are experimenting with new architectures, but most struggle against NVIDIA’s ecosystem. High-bandwidth memory (HBM) is another bottleneck. The U.S. still depends heavily on Samsung and SK Hynix, while Micron is working to build more domestic supply.
Data Centers (Facilities, Power, Cooling)
The U.S. leads in hyperscale capacity with ~53.7 GW in 2024. Northern Virginia is the largest hub, with other centers in Silicon Valley, Dallas–Fort Worth, Chicago, Phoenix, and the Pacific Northwest. Growth is also shifting to the Midwest where land and power are more available. Operators are investing in mega-campuses, while placement providers like Digital Realty, Equinix, QTS, CyrusOne, and Switch remain important for fast deployment. AI clusters pack GPUs so densely that traditional air cooling isn’t enough. This is driving liquid cooling systems, cold-plate, immersion, rear-door, which consume extra power and water. In dry states, community pushback is rising. To adapt, companies are turning to recycled water, direct-to-chip cooling, and even on-site power generation using fuel cells, gas turbines, or future SMRs.
Networking
Networking is as critical as the chips themselves. NVIDIA’s InfiniBand and RoCE-based Ethernet dominate AI training clusters, while hyperscalers design custom networking chips for speed and efficiency. U.S. cloud providers also own vast private fiber networks across regions and continents, which keep their AI services connected. On the mobile side, 5G is widespread in U.S. cities and early work on 6G is already underway.
Storage
AI systems need enormous amounts of storage because training models involve handling datasets measured in petabytes, or millions of gigabytes. Inside data center racks, GPUs use high-bandwidth memory (HBM) for quick access, but the bulk of training data sits in massive cloud storage services like Amazon S3 or Google Cloud Storage. For high-performance tasks, companies are shifting toward flash-based systems, which are much faster than older hard drives. Firms like Dell, NetApp, and IBM provide many of these enterprise solutions. To manage performance and costs, data centers use technologies like NVMe (Non-Volatile Memory Express). This is a newer way for computers to connect to solid-state drives. Unlike older systems designed for slow spinning hard drives, NVMe was built for SSDs and allows data to move much faster, like upgrading from a single-lane road to a multi-lane highway. When workloads are spread across multiple machines, NVMe-over-Fabrics extends that same speed across a network. Privacy laws also shape how data is stored. The GDPR, or General Data Protection Regulation in Europe, and the CCPA, or California Consumer Privacy Act, require that personal data be stored and processed under strict rules, sometimes within specific regions. This often means companies must duplicate data across multiple sites, raising costs and complexity. To meet these requirements, data centers increasingly rely on encryption, which locks data so it cannot be read without permission, and newer approaches like federated learning, which trains AI models without moving raw data, and homomorphic encryption, which performs computations on encrypted data without ever exposing it.
Policy and Regulation
Government policy plays a huge role in shaping AI infrastructure in the United States. The CHIPS and Science Act directs tens of billions of dollars toward building semiconductor fabs, funding research, and securing equipment. Although workforce shortages and high construction costs make execution difficult, this law is still central to building more reliable domestic supply chains. Export controls introduced in 2022 also limit the sale of advanced GPUs and semiconductor manufacturing equipment to China. These restrictions help preserve the U.S. lead in cutting-edge technology but also encourage China to accelerate its own domestic alternatives. Other legislation like the Inflation Reduction Act and the Infrastructure Investment and Jobs Act indirectly supports AI infrastructure. These laws create incentives for clean energy, grid modernization, energy storage, and broadband expansion, all of which benefit the data center industry. Regulators such as FERC (Federal Energy Regulatory Commission) are also reforming the process for connecting new projects to the grid, though long delays remain one of the biggest bottlenecks.
Capital Flows and Investment Trends
Investment in AI infrastructure has been strong across venture capital, private equity, infrastructure funds, and public markets. Venture and growth investors have poured money into AI chip startups and infrastructure providers, with firms like SambaNova, Cerebras, and CoreWeave raising large rounds. By 2024 and 2025, funding became more selective, but companies that could differentiate themselves still attracted significant capital. Private equity and infrastructure funds have also been active, acquiring major colocation providers such as QTS, CyrusOne, Switch, and DataBank. These deals reflect the view that data centers are becoming utility-like assets with steady, long-term cash flows. Real estate investment trusts, including Digital Realty and Equinix, continue to expand their campuses, often partnering with sovereign wealth funds and pension funds to finance growth. In public markets, semiconductor firms and networking suppliers have seen valuations surge as demand for AI infrastructure has accelerated. NVIDIA, Broadcom, and Arista Networks stand out as clear winners. Above all, the largest source of capital comes from the hyperscalers themselves. Their annual spending runs into the tens of billions, and more of that money is being directed toward AI-specific infrastructure than ever before.
Key Players
Hyperscale cloud providers
AWS Leads with broad regional coverage, GPU instances and custom silicon (Trainium for training, Inferentia for inference). It also ties edge offerings (Snow, Greengrass) to AI at the edge.
Microsoft Azure Accelerated via its OpenAI partnership, scaling new GPU clusters and retooling data centers for AI-heavy workloads.
Google Cloud Deploys TPUs (progressing to v6) and a mature AI software stack; it focuses on integrating first-party models and tooling across cloud services.
Meta Pivoted data center design toward AI in 2023, re-allocating capital expenditures to high-density AI facilities and internal research. (Other large platforms continue to build, but the report centers on these leaders.)
Semiconductors and memory
NVIDIA dominates data-center GPUs for training and inference, reinforced by the CUDA software stack and high-speed interconnects. AMD (EPYC CPUs; Instinct MI series GPUs) is gaining share across supercomputing and cloud.
Intel (Xeon; Gaudi line) remains a core CPU supplier and a challenger in accelerators.
Colocation and interconnection
Digital Realty, Equinix, QTS, CyrusOne, Switch and peers provide interconnect-rich campuses, fast time-to-capacity, and neutral hubs. Hyperscalers still lease for speed and overflow while continuing to build mega-campuses.
Power, cooling, and edge
Vendors span liquid-cooling specialists and facility OEMs. Bloom Energy (fuel cells) and Vapor IO (modular/edge) appear as examples in the report’s power and edge discussions.
Networking, storage, and AI-data tooling
The stack includes high-throughput switching, smart NICs, NVMe/NVMe-oF arrays, and specialized AI data systems; the report cites innovators such as Ayar Labs (optical I/O), Lightmatter (photonic compute), and vector-database providers like Pinecone and Weaviate.
Capital providers
Hyperscalers’ own capital expenditure dwarfs external flows, but infrastructure investors, REITs, sovereigns, pensions, and venture/growth funds remain active across data centers, fiber, chips, cooling, and power solutions.
Bottlenecks and Risks
Supply chain
Advanced GPU and HBM manufacturing remains concentrated in Asia, exposing U.S. firms to geopolitical and supply risks. Construction costs for data centers and substations are rising.
Power
Interconnection delays stretch 5–7 years in some regions, slowing deployment. The growth in AI data centers is so fast that it risks overwhelming the grid. If governments can’t speed up the process of connecting new facilities to power lines, energy security could become a major roadblock. That’s why more operators are looking at generating their own electricity directly on-site. It’s a bit of a throwback to older models where big industrial projects ran on their own power plants, but it may be necessary if grids can’t keep up.
Water and cooling
Liquid cooling systems consume significant water, sparking opposition in drought-prone areas. Operators increasingly rely on wastewater or dry cooling.
Talent
Shortages persist across chip engineering, AI research, data center construction, and operations trades.
Finance and regulation
High interest rates raise financing costs. Evolving AI regulation creates uncertainty around compliance and liability.
Geopolitical & National Security
Strategic competition
AI infrastructure is no longer just a technology issue, it has become a strategic asset in global competition. The United States and China are the two main anchors of capacity growth. China is backing its domestic AI chip design efforts and building massive data centers, while the U.S. is focused on protecting its advantage in advanced GPUs, semiconductor tools, and the surrounding software ecosystem. Access to top-tier hardware is now one of the most decisive factors in the race.
Export controls and friend-shoring
Export controls are central to U.S. strategy. By restricting China’s access to advanced AI chips and chip-making equipment, the U.S. has slowed China’s ability to compete at the highest levels. At the same time, these measures push China to accelerate its own alternatives. To reduce risks, the U.S. is strengthening alliances with the EU, Japan, South Korea, and India, coordinating on fabs, materials, and standards to build more resilient supply chains.coordinate on fabs, equipment, materials, and standards to reduce single-point vulnerabilities.
Allied regions and policy moves
The EU expands via the European Chips Act and AI Act but balances growth with energy costs, privacy rules, and sustainability targets. Japan and Korea leverage strengths in memory, materials, and precision equipment. India grows domestic cloud and enforces data-localization, pushing in-country capacity.
Middle East emergence
They are also emerging as a serious AI hub. Saudi Arabia and the United Arab Emirates are deploying massive capital and energy resources to purchase GPUs and build hosting centers. For example, the UAE’s G42 has been reported to contract for hundreds of thousands of top-tier GPUs annually. If these investments play out, the region could become a third global center for AI alongside the U.S. and China.
Defense, intelligence, and secure cloud
U.S. defense and intelligence stakeholders accelerate AI adoption through secure cloud/on-prem programs and organizational moves (e.g., the consolidation of JAIC, Joint Artificial Intelligence Center, into the Chief Digital and AI Office). This drives demand for cloud services, resilient networks, and specialized data facilities.
Energy security and critical infrastructure
Power supply, grid reliability, and interconnection timelines are national-security relevant. Large AI campuses need stable, scalable power. This emphasizes the intersection of energy strategy (renewables, storage, gas, and future options like SMRs), siting, and AI-specific loads. Competing in AI isn’t only about who has the most GPUs anymore, it’s also about who can secure enough power and water to keep those GPUs running. Countries that can move quickly with permits, build reliable local grids, and plan for cooling resources will have a serious edge.
Risks and contingencies
There are risks if supply chains are disrupted. A crisis affecting leading-edge fabs could cut off access to GPUs, memory, or packaging, forcing emergency measures like stockpiling or rapid onshoring. Overregulation or fragmented rules could also slow U.S. progress, especially if rivals move ahead faster under looser regimes. The stakes are clear, whoever secures both the technology and the infrastructure to power it will hold a major advantage in global competition.
Labor & Talent Infrastructure
Acute shortages across the stack
Construction trades needed to build AI campuses (electricians, HVAC, high-voltage technicians, controls specialists).
Facilities operations (data-center technicians, reliability engineers, critical-environment operators).
Semiconductor talent (process/packaging engineers, equipment techs) amid simultaneous fab construction.
AI research & engineering (foundation-model researchers, distributed-training experts, MLOps/SRE for AI).
Public-sector expertise (regulators and practitioners who understand AI, chips, and grid planning) to craft workable rules and programs.
Pipeline, training, and costs
There is a clear need for stronger training pipelines that connect universities, apprenticeships, and industry. Skills are in short supply across electrical and mechanical trades, semiconductor engineering, and AI disciplines. This shortage raises costs, slows delivery timelines, and makes top AI researchers and infrastructure engineers extremely expensive to hire.
Geography, visas, and mobility
AI infrastructure is clustered in a few U.S. regions with strong power and fiber advantages, which quickly tightens local labor pools. Immigration constraints make it harder to bring in global experts, slowing research and deployment. Local zoning restrictions and community opposition can also push operators toward friendlier states or allied countries.
Operational scale-up
As dense, liquid-cooled AI clusters spread, facilities teams must learn new skills such as managing coolant systems, leak detection, and high-voltage troubleshooting. On the software side, engineers running large-scale training need to optimize interconnects, storage pipelines, and job scheduling while keeping reliability high.
Smarter Designs for the Future
Some researchers argue that instead of endlessly building bigger data centers, we should look at designs inspired by the human brain. The brain uses only about 20 watts of power to perform tasks that today’s AI systems need thousands of GPUs to approximate. If scientists can understand how primates process information and apply those lessons to computer chips, AI could become far more efficient.
Scenario Planning and Outlook
Bull case: Successful CHIPS Act execution, rapid grid expansion (including small modular reactors), and breakthroughs in energy-efficient chips sustain U.S. dominance. A real breakthrough in neuromorphic or brain-inspired computing could change the game. Instead of AI always requiring bigger data centers, it could become vastly more energy-efficient. That would allow the U.S. to keep its lead in AI without pushing grids and water supplies to their limits.
Base case: The U.S. maintains leadership but faces periodic supply and power constraints. Growth continues steadily.
Bear case: Geopolitical shocks, overregulation, or financing stress disrupt supply chains or investment flows. U.S. leadership could erode, with implications for national security.
On the other hand, if power shortages, water conflicts, and slow permitting aren’t solved, they could become major brakes on the entire AI industry, even if chips remain available. In that scenario, the bottleneck wouldn’t be technology, but infrastructure.
Scale
By the end of the decade, AI computing could be orders of magnitude larger than today. Workloads will spread across hyperscale campuses, smaller edge nodes, and specialized facilities. Designs optimized for AI, including dense accelerators, liquid cooling, smarter networking, and storage, will become standard. The line between AI data centers and traditional cloud campuses will blur.
Power
Faster grid connections and more on site power generation will be essential. Options like fuel cells and small modular reactors are likely to grow in importance, alongside new cooling approaches. Regions with quicker permitting and stronger utility partnerships will attract the bulk of investment. Sustainability goals will also influence where and how facilities are built.
Networking and edge
Global backbone fiber, subsea cables, and AI specific interconnects will need to scale to handle petabyte level workloads. Fifth generation mobile networks are already common in U.S. cities, while planning for sixth generation networks and stronger metro backbones is underway. These advances will support real time analytics and latency sensitive inference at the edge, from autonomous vehicles to industrial systems.
Data governance and compliance
Privacy and localization rules will require more duplication of data across regions, raising costs and complexity. To adapt, companies will combine public cloud, private clusters, and regulated domain deployments. Encryption and secure data pipelines will become the default, ensuring compliance with frameworks such as the General Data Protection Regulation in Europe and the California Consumer Privacy Act.