Why This Matters
AI infrastructure spending has entered a phase with no historical precedent. The four hyperscalers — Microsoft, Google, Amazon, and Meta — have committed over $300 billion in combined capital expenditure for 2026, the vast majority directed at GPU clusters, data center construction, and power procurement. To put this in perspective, the entire US commercial construction industry — every office building, retail space, warehouse, and factory built in a year — totals roughly $500 billion. AI data center construction alone is approaching that figure.
This is not just a story about money. It is a story about physical constraints. Building a modern AI data center requires securing land, water, electrical power, fiber connectivity, construction labor, and millions of pounds of specialized equipment — all in locations where these resources are available simultaneously. The companies that solve these logistical challenges fastest will train the next generation of frontier models. The ones that cannot will fall behind regardless of how talented their researchers are.
Infrastructure is also the dimension of AI competition that is hardest to replicate. A competitor can hire researchers, license architectures, and raise funding — but building a multi-gigawatt data center campus takes 18-36 months even at unprecedented speed. The infrastructure investments being made today will determine the competitive landscape for the next decade.
For the energy and power side of this story — gigawatts by country, nuclear deals, and the US vs China power race — see the AI Energy Tracker.
Current Landscape
The AI infrastructure buildout in mid-2026 is defined by three dynamics: hyperscaler megaprojects, GPU supply chain constraints, and the emergence of independent GPU cloud providers as a significant infrastructure layer.
Hyperscaler capital expenditure has doubled in two years. Microsoft’s capital expenditure for fiscal year 2026 is projected at $80 billion or more, up from $44 billion in FY2024. Amazon has committed $100 billion in AI-related infrastructure spending for 2025-2026. Google has pledged $75 billion. Meta has committed $60-65 billion for 2025 alone. These figures exceed the GDP of most nations and represent the largest private infrastructure investment program in history.
Over 100 major data center facilities are planned or under construction. The pipeline spans the United States (with concentrations in Texas, Virginia, Georgia, and the Midwest), Northern Europe (Sweden, Finland, Norway), the Middle East (UAE, Saudi Arabia), and Southeast Asia (Malaysia, Indonesia). Each facility is measured in hundreds of megawatts to multiple gigawatts of power capacity — scales that were associated with entire cities a decade ago.
GPU supply remains the binding constraint. NVIDIA shipped approximately 3.5 million data center GPUs in 2025 and is projected to ship over 4 million in 2026, generating over $100 billion in data center revenue. Despite this volume, demand consistently exceeds supply. The H100 and its successors (H200, B100, B200, GB200) are allocated quarters in advance, and major customers negotiate directly with NVIDIA CEO Jensen Huang for priority access. The GPU supply chain — from TSMC fabrication in Taiwan to CoWoS advanced packaging to HBM memory from SK Hynix and Samsung — has become one of the most strategically important industrial pipelines in the world.
Cooling technology is becoming a differentiator. AI chips generate dramatically more heat per rack than traditional server hardware. A single NVIDIA GB200 NVL72 rack draws over 120 kW — roughly 10x a typical enterprise server rack. This has driven rapid adoption of liquid cooling systems, with companies like Vertiv, Schneider Electric, and CoolIT Systems struggling to meet demand. Direct-to-chip liquid cooling is now standard for new AI data centers, and some facilities are exploring immersion cooling where entire servers are submerged in dielectric fluid. The shift from air to liquid cooling represents the most significant change in data center thermal management in decades.
Key Players
Microsoft is the largest single spender on AI infrastructure, driven by its need to support both Azure AI cloud services and its exclusive compute partnership with OpenAI. Microsoft’s $80 billion+ annual capex funds data center construction across the US, Europe, and Asia. The company has signed the most prominent nuclear power deal in the AI industry — a 20-year power purchase agreement with Constellation Energy to restart Unit 1 of the Three Mile Island nuclear plant in Pennsylvania, providing 837 MW of carbon-free power. Microsoft’s infrastructure strategy is inseparable from OpenAI’s training requirements; the Stargate project in Abilene, Texas — a joint venture with OpenAI, SoftBank, Oracle, and MGX — is the largest announced AI infrastructure project at $500 billion over four years.
Google invests $75 billion+ annually in infrastructure with a distinctive advantage: custom silicon. Google’s Tensor Processing Units (TPUs), now in their sixth generation (Trillium), provide an alternative to NVIDIA GPUs for both training and inference. This vertical integration — designing its own AI chips, manufacturing through partnerships with Samsung and TSMC, and deploying in its own data centers — gives Google cost advantages and supply chain independence that no other hyperscaler matches. Google’s data center footprint spans 40+ facilities globally, with major expansions underway in Oregon, Oklahoma, Wisconsin, and multiple international sites. The company’s nuclear power strategy includes a deal with Kairos Power for 7 small modular reactors (SMRs) by 2035.
Amazon / AWS has committed the largest absolute dollar figure to AI infrastructure — over $100 billion for 2025-2026 — reflecting its position as the world’s largest cloud provider. AWS supplements NVIDIA GPUs with its custom Trainium AI training chips (now in second generation) and Inferentia inference chips. Like Google, Amazon’s custom silicon strategy reduces dependence on NVIDIA and provides cost advantages for customers willing to port their workloads to AWS-specific hardware. Amazon’s data center expansion is concentrated in Northern Virginia (the world’s densest data center market), Oregon, Ohio, and Mississippi, with international expansions across Europe, Asia, and the Middle East.
Meta differs from the other hyperscalers in a critical way: it does not sell cloud services. Meta’s $60-65 billion infrastructure investment is entirely for internal use — training Llama models and running AI inference across Facebook, Instagram, WhatsApp, and Threads for over 3 billion monthly users. This makes Meta’s infrastructure the largest “private cloud” in the AI industry. Meta’s most ambitious project is a 4 GW data center campus in Richland Parish, Louisiana, which would be the largest single data center site in the world if fully built. The company is also building a major campus in Temple, Texas, and expanding facilities across its existing fleet.
NVIDIA does not build data centers, but it is the essential supplier for every company that does. NVIDIA’s data center revenue exceeded $100 billion in 2025, driven by the H100, H200, and B-series GPU families. The company’s CUDA software ecosystem — the programming framework that virtually all AI training code is written for — creates a lock-in effect that competing chip makers have struggled to overcome. NVIDIA’s NVLink and NVSwitch interconnect technologies, which allow thousands of GPUs to work together as a single system, are as strategically important as the chips themselves. Jensen Huang’s relationships with the CEOs of every major AI company give NVIDIA unparalleled visibility into demand and influence over allocation.
CoreWeave is the most prominent of the independent GPU cloud providers — companies that rent NVIDIA GPU clusters without the bundled cloud services of AWS, Azure, or Google Cloud. CoreWeave went public in early 2025 at a valuation of approximately $25 billion, backed by a $12 billion debt facility secured by its GPU fleet. The company has signed major contracts with Microsoft and AI labs that need raw GPU compute without the overhead of a full cloud platform. CoreWeave’s business model is essentially GPU arbitrage — buying NVIDIA chips in bulk, housing them in leased data centers, and renting them by the hour at a premium. The model works as long as demand exceeds supply, but is vulnerable to any easing of the GPU shortage.
Infrastructure Spending Comparison
| Company | 2024 Capex | 2025 Capex (est.) | 2026 Capex (est.) | Primary Purpose |
|---|---|---|---|---|
| Microsoft | $44B | $80B+ | $80B+ | Azure AI, OpenAI partnership |
| Amazon/AWS | $59B | $100B+ | $100B+ | AWS cloud, Trainium/NVIDIA |
| $32B | $75B+ | $75B+ | Cloud AI, TPU, Gemini training | |
| Meta | $37B | $60-65B | $65B+ | Llama training, internal inference |
| Total Big 4 | $172B | $315B+ | $320B+ | |
| CoreWeave | ~$3B | ~$8B | ~$12B | GPU cloud rental |
| Oracle | ~$12B | ~$20B | ~$25B | Cloud, Stargate partnership |
| xAI | ~$5B | ~$10B | ~$15B | Colossus, training Grok |
GPU Supply Chain
The AI GPU supply chain is one of the most concentrated and strategically important industrial pipelines in the world.
TSMC (Taiwan) fabricates virtually all high-end AI chips, including NVIDIA’s GPUs, Google’s TPUs, Amazon’s Trainium, and AMD’s MI300 series. TSMC’s advanced packaging technology, CoWoS (Chip-on-Wafer-on-Substrate), is the bottleneck — it allows multiple chip dies and HBM memory stacks to be assembled into a single package. TSMC has invested over $30 billion in expanding CoWoS capacity, but demand continues to outstrip supply.
SK Hynix and Samsung supply High Bandwidth Memory (HBM), the ultra-fast memory stacked directly on top of or beside AI chips. HBM is critical for AI performance because transformer-based models are memory-bandwidth limited for most inference tasks. SK Hynix has a dominant position in HBM3E (the current generation) and is investing heavily in HBM4 for 2025-2026 delivery. The HBM market has grown from $2 billion in 2022 to a projected $25 billion+ in 2026.
NVIDIA’s GPU roadmap drives the entire infrastructure cycle. The company releases new GPU architectures on a roughly annual cadence: Hopper (H100, 2023) → Blackwell (B100/B200, 2024-2025) → Vera Rubin (2026). Each generation delivers approximately 2-3x improvement in AI training performance and 3-5x improvement in inference efficiency per watt. This performance curve is what makes the continuous infrastructure investment rational — a data center built with current-generation GPUs will be significantly more cost-effective per unit of AI compute than one built with previous-generation hardware.
What We’re Tracking
JustSaid monitors AI infrastructure through a systematic multi-source approach.
SEC filings and earnings calls. Capital expenditure guidance, property and equipment additions, and management commentary from quarterly earnings calls provide the most authoritative data on spending levels and investment plans. The pipeline captures every capex-related statement from the four hyperscalers plus CoreWeave, Oracle, and xAI.
FERC dockets and EIA data. Federal Energy Regulatory Commission filings reveal power purchase agreements, transmission line requests, and interconnection applications — all of which signal where new data centers are being built before they are publicly announced. Energy Information Administration data provides context on regional grid capacity and generation sources.
Construction and permitting records. County-level building permits, environmental impact statements, and zoning applications reveal data center projects months before press announcements. The tracker monitors permit activity in the top 20 US data center markets.
Supply chain intelligence. TSMC’s earnings calls, SK Hynix’s production guidance, and NVIDIA’s channel commentary provide visibility into the component supply chain. Quarterly HBM shipment volumes, CoWoS capacity utilization, and GPU allocation patterns all feed into the infrastructure picture.
Recent Developments
May 2026: Stargate Phase 2 construction begins. The second major building at the Stargate campus in Abilene, Texas, is under construction, with Oracle providing the cloud infrastructure and SoftBank providing additional capital. The first building became operational in June 2025 and is already at capacity running OpenAI training workloads.
April 2026: NVIDIA announces Vera Rubin architecture. NVIDIA’s next-generation GPU architecture, code-named Vera Rubin, promises 3x the AI training performance of Blackwell per chip. Production is targeted for late 2026, with major customer shipments beginning in Q1 2027. The announcement triggered a new wave of data center planning as customers design facilities around the new chip’s power and cooling requirements.
Q1 2026: CoreWeave’s post-IPO expansion. CoreWeave accelerated its data center buildout following its March 2025 IPO, signing leases for 8 new facilities across New Jersey, Texas, and the UK. The company’s GPU fleet is projected to exceed 500,000 NVIDIA GPUs by end of 2026, making it the fifth-largest GPU operator after the four hyperscalers.
Q4 2025: Liquid cooling becomes standard. Every major data center project announced in Q4 2025 specified liquid cooling as the primary thermal management system for AI compute racks. Air cooling, which was standard for decades of data center operation, is now insufficient for the power densities required by modern AI chips. The transition has created supply shortages for liquid cooling components, with lead times extending to 6-9 months for coolant distribution units and rear-door heat exchangers.
Outlook
AI infrastructure investment will continue to accelerate through 2027, but the nature of the buildout is shifting.
From training to inference. The first wave of AI infrastructure investment was driven by training — building the massive GPU clusters needed to train frontier models. The next wave will be driven by inference — building the distributed infrastructure needed to serve these models to billions of users. Inference workloads have different hardware requirements (more memory bandwidth, less raw compute per chip) and different geographic requirements (closer to end users for latency). This shift will spread infrastructure investment beyond the current training hub concentrations.
Custom silicon will gain share. Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia, and Meta’s MTIA chips will collectively reduce NVIDIA’s dominance of AI compute from its current ~90% market share to an estimated 70-75% by 2028. This diversification is positive for the industry — it reduces single-vendor risk and creates competitive pressure on pricing — but negative for NVIDIA’s margins.
International expansion will accelerate. Sovereign AI initiatives in the UAE, Saudi Arabia, Japan, India, and multiple European countries are driving data center construction outside the US. These projects are often backed by government investment and motivated by a desire for AI sovereignty — the ability to train and run AI models on domestic infrastructure without dependence on American cloud providers.
The financing model is evolving. The traditional model of hyperscalers funding infrastructure from operating cash flow is being supplemented by new structures: CoreWeave’s asset-backed GPU lending, SoftBank’s Stargate co-investment vehicle, and sovereign wealth fund partnerships. As infrastructure costs grow, even companies with the largest balance sheets in history are seeking external capital to share the risk.
Frequently Asked Questions
Why are AI data centers so much more expensive than traditional data centers? Three factors drive the cost difference. First, power density: an AI rack draws 50-120 kW versus 5-10 kW for a traditional server rack, requiring proportionally more power infrastructure and cooling. Second, GPU costs: a single NVIDIA GB200 NVL72 rack costs over $3 million, compared to roughly $200,000 for a traditional compute rack. Third, networking: AI training requires ultra-high-bandwidth interconnects (NVLink, InfiniBand) that are far more expensive than standard Ethernet. A fully equipped AI data center costs $15-25 million per megawatt of IT power, compared to $8-12 million for traditional data centers.
Can the power grid handle this much new demand? Not without significant upgrades. The US grid added roughly 20 GW of new generation capacity in 2025, while AI data center demand alone is projected to require 15-20 GW of new capacity by 2028. Grid interconnection queues — the waiting list for new generation and load to connect to the transmission system — extend 4-7 years in many markets. This is why AI companies are pursuing behind-the-meter power solutions (on-site generation, nuclear restarts, dedicated renewable installations) rather than relying solely on grid capacity. See the AI Energy Tracker for detailed coverage of power procurement.
Is this level of spending sustainable? The companies making these investments believe AI will generate returns that justify the capex. Microsoft’s AI-related Azure revenue is growing over 60% year-over-year. Google’s Cloud AI revenue is growing similarly. The risk is not that AI infrastructure is unprofitable — it is that overbuilding could create excess capacity if demand growth slows. The historical analogy is the fiber optic buildout of the late 1990s, where massive overinvestment led to a glut and industry consolidation. Whether AI infrastructure follows the same pattern depends on whether AI application revenue growth sustains the current trajectory.
What happens to data centers when GPU generations change? Modern AI data centers are designed for GPU refresh cycles. Power and cooling infrastructure — the most expensive and longest-lead components — are sized for future GPU generations, not just current ones. When a new GPU generation arrives, the facility’s servers are swapped while the building, power, and cooling infrastructure remain. This is why data center useful life is typically 20-30 years even though the GPUs inside them may be refreshed every 2-3 years.