The AI Compute and Power Infrastructure Boom (2020–2025)

Historical Baseline (2020–2024)

From 2020 through 2024, global compute capacity and data center energy use were already on steep upward trends. Major supercomputers (e.g. Frontier, Fugaku, Aurora) and cloud GPU clusters (Nvidia A100/H100, Google TPUs) came online, pushing total AI/HPC compute to multi-exaflops. Data centers grew rapidly – U.S. data center power usage rose by ~12% per year, reaching about 415 TWh in 2024 (≈1.5% of global electricity). With AI training accelerators demand growing ~30% per year, global data center demand is projected to roughly double again by 2030. Major hyperscalers (Microsoft, Google, Amazon, Meta, Alibaba) announced multi-billion-dollar cloud and datacenter investments throughout 2022–2024 to meet AI workloads. Governments responded with “chips” and AI strategies (e.g. US CHIPS and Science Act of 2022, EU Chips Act), aiming to spur domestic chip fabs and HPC centers. Corporate strategy pivoted: Intel scaled up AI chip efforts, AMD regained competitiveness with MI300X GPUs, and private equity became active in AI infrastructure (e.g. Blackstone, BlackRock). Power infrastructure lagged behind: many regions saw tight grid capacity and DC vacancy falling to record lows (Northern Virginia <1%, Singapore ≈1%). By 2024, it was clear that AI compute growth was outpacing traditional infrastructure, prompting new investment plans in chips, datacenters, and power.

Major AI Deals in 2025

Throughout 2025, a flurry of high-profile deals – mergers, acquisitions, investments, and partnerships – underscored the AI infrastructure build-out. Notable transactions included:

• HPE acquires Juniper Networks ($16B): In 2025 HPE agreed to buy networking vendor Juniper for about $16 billion, strengthening HPE’s AI data center and networking portfolio.
• TPG takes private software firm Proficy ($600M): Private equity firm TPG bought AI-enabled software supplier Proficy for ~$600 million, aiming to integrate its generative AI tools into larger enterprises.
• CoreWeave buys Core Scientific ($9B): GPU cloud provider CoreWeave acquired bitcoin miner Core Scientific for roughly $9.0 billion. This deal sharply expanded CoreWeave’s supply of low-cost GPU hardware and bitcoin-mining power for AI workloads.
• BlackRock/GIP acquires Aligned Data Centers ($40B): A joint bid by BlackRock and Global Infrastructure Partners (GIP) for U.S. data center operator Aligned DC (~$40 billion) was announced, potentially closing by 2026. It’s the largest AI-era private equity deal yet, reflecting investor demand for AI-ready data centers.
• AMD–OpenAI compute partnership: In October 2025 AMD and OpenAI announced a multi-year strategic partnership. AMD will supply roughly 6 gigawatts worth of MI450X/MI350X/MI300X GPU power (generally starting in late 2026) to OpenAI, unlocking “tens of billions” in expected revenue for AMD.
• Nvidia-led U.S. manufacturing pledge ($500B): Nvidia, TSMC, Samsung, Intel and others announced an initiative to invest up to $500 billion over several years in U.S. chip and server manufacturing. This includes new fabs (TSMC, Samsung, Intel expansions) and server factories (Nvidia-AIMasters). It’s touted as the largest AI hardware push to date.
• Public-private AI infrastructure partnerships: Governments joined corporates in major deals. France, Italy, UK, Spain and other European states are co-investing with NVIDIA to build national AI “clusters” using Nvidia’s Blackwell GPUs (targeting thousands of exaflops of compute). Saudi Arabia’s Public Investment Fund (PIF) partnered with Google Cloud on a $10B AI/cloud hub to serve MENA (generating ~11,000 US tech jobs and \$35B US economic benefit). In the UAE, the Abu Dhabi “Stargate” AI Campus – a 5-gigawatt AI center – was announced, backed by UAE tech firm G42 and partners (OpenAI, Nvidia, Oracle, SoftBank). The first 200 MW of that campus is slated for 2026 operation.
• Hyperscaler cloud commitments: Among cloud-specific deals, OpenAI announced a 7-year, $38B commitment to Amazon Web Services for custom AI GPU infrastructure, securing hundreds of thousands of NVIDIA GB200/GB300 chips from AWS. Meta (Facebook) signed a $14.2B multi-year deal with GPU cloud provider CoreWeave for AI compute capacity. AMD also secured a GPU compute deal with OpenAI, and Oracle expanded its cloud contract with generative AI software firms (including a pledge to deploy over 130,000 AMD MI355X GPUs in 2025).
• Venture funding for AI startups: AI-focused companies continued raising record rounds. For example, EU’s data-center-infrastructure developer Nscale closed a €1.1B Series B (largest ever in Europe) to build high-performance AI data centers. U.S. startups like Crusoe Energy (gas-to-compute) raised \$1.4B (Series D), and robotaxis AI firm Avride got \$375M. In biotech, Fractal AI and Treeline Biosciences raised \$600M and \$600M respectively for AI-driven R&D. NVIDIA’s venture arm also backed AI solution firms: e.g. battery recycling company Redwood Materials raised \$350M (Series E) with NVIDIA Ventures as lead, and conversational AI firm Uniphore raised \$260M from investors including Nvidia and AMD.

These deals highlight the strategic trend: massive investment flows into AI compute and infrastructure, with deal values often in the tens of billions, involving hyperscalers, chipmakers, and global investors. The build-out is accelerating.

AI Infrastructure Expansions (Major Players)

Microsoft (Azure)

Microsoft accelerated Azure’s expansion for AI in 2025. It announced roughly 2 gigawatts of new data center power capacity and unveiled what it called “the world’s most powerful AI datacenter,” claiming ~10x the performance of the fastest 2024 supercomputer. Microsoft added new Azure regions and availability zones in Asia (e.g. launched new regions in Malaysia and Indonesia in 2025, with Taiwan North coming in 2026). Overall Azure now spans 70 regions (operational or announced) globally. Internal planning papers and industry reports put Microsoft’s AI-datacenter investment (2022–25) at on the order of \$80 billion. Microsoft’s infrastructure leverages NVIDIA and AMD GPUs (e.g. Azure NDv5 with H100 GPUs), along with custom CPUs for efficient data movement. The company’s close partnership with OpenAI (Azure OpenAI Service) means Azure is one of the largest single consumers of high-end GPUs, spurring these massive hardware deployments.

Google (Cloud & DeepMind)

Google Cloud poured tens of billions into AI infrastructure in 2025. It reaffirmed plans for an \$85B capital spend on data centers over 2025–26, including \$15B toward a new AI compute hub in southern India (reflecting partnership with TSMC). In October 2025 Google Cloud announced a \$10B deal with Saudi Arabia’s PIF to build out a global AI hub (increasing Google’s Middle East cloud regions). Google’s own AI lab (DeepMind) operates large TPU-based supercomputers, and Anthropic agreed to expand its use of Google Cloud TPUs by the millions. Reports indicate Anthropic is discussing multi-year GCP deals worth “tens of billions” to accelerate training on Google TPUs. Google also continued to roll out new cloud regions (e.g. Dubai, Zurich) and to upgrade network links (e.g. new undersea cables) to support AI traffic. Overall, Google’s 2025 cloud build-out combined traditional regions and specialized AI clusters (TPU v5/v6 pods, multi-TPU racks) to meet surging demand.

Meta (Facebook AI)

Meta took a dual approach of building its own mega-datacenters while leasing extra compute. Meta’s internal AI supercluster (codenamed Prometheus) is expected to reach 1 gigawatt of power by 2026, and in 2025 it announced plans for an additional 5 GW data center campus after that. To supplement in-house GPU supply, Meta signed a \$14.2B multi-year lease agreement with cloud provider CoreWeave for GPU servers. Meta’s 2025 capital budget was in the \$66–72B range, spurred by AI, and CFOs indicated a multi-year infrastructure build to follow. (Earlier reports said Meta had earmarked \$600B through 2028 for data centers and AI infrastructure.) Meta is installing advanced liquid-cooled racks (with Nvidia GH200 superchips, etc.) in new sites. Its Arizona, Texas, and Ohio campuses are being readied for these clusters. As a result, Meta’s compute footprint (in GPUs) will roughly double or triple by late 2026, funded by the CoreWeave contract and its own build-out.

Amazon Web Services (AWS)

AWS continued to invest heavily in AI servers and networking. Most prominently, in Oct 2025 AWS announced a \$38B, 7-year cloud partnership with OpenAI – committing hundreds of thousands of NVIDIA GB200/GB300 GPU instances to power ChatGPT-scale training. Separately, AWS and Anthropic launched “Project Rainier,” an AI supercomputer with ~500,000 Amazon Trainium2 chips (for Anthropic) now online and expected to exceed 1,000,000 chips by year-end. AWS branded this the largest AI training cluster in existence. On the connectivity side, AWS led construction of the 320 Tbps FastNet subsea cable (U.S.–Ireland, online ~2028) specifically to handle surges of cloud AI traffic. AWS also expanded its own cloud regions (e.g. Israel, UAE, Spain) in 2024–25. Its infrastructure now includes custom UltraServer racks (each with 64+ Trainium2 chips), Nitro chips, and planned higher-speed interconnects (AWS Fabric Cloud). The AWS AI expansion, anchored by its deals with OpenAI and Anthropic, has made AWS one of the largest single compute providers globally.

Oracle (Cloud Infra)

Oracle’s cloud unit (OCI) emerged as a key AI infrastructure player. In mid-2025, Oracle announced it would roll out roughly 131,072 AMD Instinct MI355X GPUs in a single cloud “AI Cluster,” one of the largest single deployments of accelerators. These GPUs (already sampling in cloud regions) will enable customers to run large transformer models. Oracle also deployed thousands of NVIDIA DGX SuperPOD systems for AI workloads. Oracle opened new cloud regions (Portugal, India) in 2025 to support European AI growth. To meet density needs, Oracle’s latest data halls use extensive liquid cooling (as noted by industry observers). Overall, Oracle’s strategy is to offer high-density GPU supercluster instances on-demand, challenging the hyperscalers by selling AI cloud to enterprises and governments (for example, Oracle is a partner in UAE’s Stargate AI project). Its record sales and stock performance in 2025 reflected this pivot.

NVIDIA

NVIDIA remains at the center of compute expansion. It announced a letter of intent to invest up to \$100B with OpenAI to deploy 10 gigawatts of NVIDIA-based data centers. Internally, NVIDIA also launched major partnerships to build “AI factories”: in October 2025 it teamed with Samsung and SK Group to build two separate AI development centers in South Korea, each housing over 50,000 NVIDIA GPUs (via GH200 and later chips) by 2027. On top of that, NVIDIA initiated a U.S.-based production effort with TSMC, Foxconn, and others, pledging to produce \$500B of AI servers in the U.S. (an estimate) over four years. NVIDIA also introduced its latest chip lineup (Grace Blackwell GH200 GPUs, Grace CPU, etc.), which industry reports indicate will double performance per rack over prior generations. Finally, NVIDIA continued to expand its own data centers and “NVL” cloud service – it will purchase any unsold capacity on partner clouds and has deals to share infrastructure. Overall, NVIDIA’s deployed capacity (GH200 and H200) worldwide probably reached on the order of 10–20 GW by late 2025, given the massive factory commitments and customer clusters.

OpenAI

OpenAI’s compute footprint grew through partnerships rather than building its own mega-datacenter. The AWS \$38B deal and AMD contract lock in vast GPU/AI-accelerator resources. OpenAI is also scaling up in Microsoft Azure (still a partner) but has relinquished MS first-right-of-refusal, preparing a possible IPO. CEO Sam Altman disclosed goals like 1 exaflop training clusters and up to 30 GW of compute (30M home equivalents) in coming years. OpenAI is reportedly exploring on-shore power solutions (like building data centers in power-rich US states) to handle this load. It also joined NVIDIA to co-design future chips for AI. In short, OpenAI’s strategy is to secure huge amounts of cloud/colocation compute (AWS, Azure, GCP) and gradually diversify into building custom data centers (e.g. an announced Texas AI campus for 2026); its compute resource grew by several exaflops in 2025 thanks to these deals.

Anthropic

Anthropic, the AI startup behind the Claude models, also massively expanded its compute. It signed multi-year commitments with AWS (reportedly \$8B+ over time) and began using Google TPUs en masse. In October 2025, Anthropic announced it would scale up to over 1 million Google TPU v5e chips across its projects, a deal worth “tens of billions.” Anthropic’s own press noted this TPU deployment (in addition to AWS’s Trainium2) will give it well over 1 GW of dedicated AI compute by 2026. It also announced a \$50B investment plan in U.S. AI infrastructure: new data center campuses in Texas and New York that will create ~800 permanent jobs plus 2,400 construction jobs. Like OpenAI, Anthropic remains largely cloud-based but is pre-purchasing or co-designing hardware. Its massively parallel infrastructure (spread across AWS, GCP, and colos) is projected to soon rival or exceed the largest single clusters at any one company.

Tesla (Dojo)

Tesla’s AI infrastructure efforts continue on a different path. In 2025 Tesla largely completed rollout of its first Dojo supercomputer (an on-prem facility at Giga Nevada with ~90,000 GPUs by mid-2024) and shifted focus. Tesla placed a \$16.5B order with Samsung for next-generation “AI6” training chips (with production slated for 2025–26). According to earnings calls, Tesla expects to have a second-generation Dojo cluster operating at scale by 2026, roughly equivalent to 100,000 Nvidia H100 GPUs. However, they have indicated that most inference/training will run on off-the-shelf GPUs and the new Tesla chips. Tesla’s 2025 capital expenditure is expected to be around \$5B on AI infrastructure, mostly for these GPUs. So far, Tesla’s build-out remains internal (no hyperscaler partnerships), but its AI hardware (D1 chips, AI6 chips) and on-site datacenter at Gigafactories are part of the broader compute ecosystem.

Global Compute Capacity in 2025

By late 2025, the world’s installed AI compute (including HPC and cloud GPU clusters) had grown to unprecedented scale. Analysis by research firm Epoch AI estimates the United States accounts for roughly 75% of total GPU-cluster compute performance globally, with China about 15% and the entire EU around 5%. This reflects the dominance of U.S. hyperscalers (OpenAI/Microsoft, AWS, Google, Meta, etc.) and chip design centers. Top supercomputers include multiple exascale machines: in the U.S. Frontier (Oak Ridge) and El Capitan (LLNL) each surpass 1.7 exaflops (FP64) of peak; Europe’s first exascale system (JUPITER in Germany) came online September 2025 at ~1.0 exaflop; China has several exa-scale-class systems (Tianhe-3 expected). High-end clusters in Japan, Canada, and others add hundreds of petaflops. In aggregate, global on-line HPC peak performance likely exceeds 10 exaflops by 2025, with plans (e.g. Aurora, Jupiter, Leão da Montanha) to push beyond 20 exaflops by 2028.

On the GPU side, Nvidia sold millions of H100 and GH200-class accelerators in 2024–25, and AMD shipped many MI300X GPUs. Google’s TPU deployments reached the millions, and custom fabrics (like Cerebras wafers) also contribute. Industry estimates suggest tens of millions of discrete AI chips (GPUs/TPUs/ASICs) are now under contract. Data center counts continue to rise: hundreds of hyperscale sites worldwide (50+ in US, 30 in China, 20 in EU/Asia-Pacific each) carry AI workloads, with another 100+ major sites planned by 2028. For example, Google is building a \$5B data center in Belgium (300 MW), Microsoft and Oracle each have dozens of large campuses, and new sites are sprouting from North America through Latin America, Africa, and Asia. In summary, global compute capacity in 2025 is heavily U.S.-centric (the Epoch data shows nearly 75% share:) but expanded globally via cloud networks. Total on-line FP32/FP16 performance (counting all AI-focused clusters) is on the order of 20–50 exaflops of effective throughput, with plans easily doubling that by 2030. Facilities span all geographies: in the Americas (US, Canada, Brazil), Europe (Germany, Finland, UK, France), Asia (China, Japan, South Korea, India, Singapore), and the Middle East (UAE, Saudi AI sites).