# AI Datacenter

> Source: https://aiwiki.ai/wiki/ai_datacenter
> Updated: 2026-06-27
> Categories: AI Hardware, Artificial Intelligence
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

An **AI datacenter** (also written as AI data center) is a facility purpose-built for [artificial intelligence](/wiki/artificial_intelligence) workloads: it is filled with dense clusters of [GPUs](/wiki/gpu) or specialized AI accelerators, wired together by high-bandwidth networking, and cooled by liquid systems that can remove far more heat per rack than a conventional [data center](/wiki/data_center). These facilities exist to train and serve [large language models](/wiki/large_language_model) and other AI systems, and they consume electricity at a scale that is reshaping power grids: the International Energy Agency projects that global data center electricity use will roughly double to about 945 terawatt-hours (TWh) by 2030, with AI as the single most important driver [11].

Unlike traditional data centers built around general-purpose CPUs for web hosting, databases, and enterprise applications, AI datacenters are architected around accelerators connected by high-bandwidth fabric and cooled by advanced thermal management capable of dissipating far greater heat densities. The scale of investment has reached extraordinary levels: combined hyperscaler capital expenditure (Amazon, Google/Alphabet, Microsoft, Meta, and Oracle) is forecast to exceed $600 billion in 2026, a 36% increase over 2025, with roughly 75% of that spending (about $450 billion) tied directly to AI infrastructure [1].

## What is an AI datacenter?

An AI datacenter is distinguished from a conventional facility by three things: extreme compute density, specialized high-bandwidth networking, and liquid cooling. Where a standard enterprise server rack draws 5 to 15 kW, a modern AI GPU rack can draw 40 to 130+ kW, requiring direct liquid cooling rather than air. The workloads are also different: AI training synchronizes thousands of accelerators that must exchange data continuously, so the network fabric (not just the chips) determines how fast a cluster can train a model.

The IEA captures the scale vividly: "A typical AI-focused data centre consumes as much electricity as 100 000 households, but the largest ones under construction today will consume 20 times as much." [11] In 2024, data centers worldwide consumed about 415 TWh of electricity, roughly 1.5% of global demand; by 2030 that figure is projected to reach about 945 TWh, just under 3% of global electricity, with electricity use at AI-optimized data centers poised to more than triple over the period [11].

## What is the AI datacenter buildout?

The AI datacenter buildout underway in 2025 and 2026 represents one of the largest infrastructure investment cycles in history, comparable in scope to the buildout of the electrical grid, the highway system, or the original internet backbone.

### Hyperscaler capital expenditure

The following table shows projected 2026 capital expenditure for major AI infrastructure investors:

| Company | Projected 2026 Capex | Primary AI Focus | Notable Commitments |
|---|---|---|---|
| [Amazon](/wiki/amazon) (AWS) | ~$200 billion | Cloud AI services (AWS Bedrock, SageMaker, Trainium chips) | Largest single-company capex; includes custom Trainium accelerators |
| [Google](/wiki/google) / Alphabet | $175-185 billion | [Gemini](/wiki/gemini) training, Cloud TPU/GPU services, AI search | TPU v6 deployment; Trillium chips |
| [Meta](/wiki/meta) | $115-135 billion | [Llama](/wiki/llama) model training, AI-powered recommendation, AR/VR | 1 GW Ohio datacenter; 5 GW Louisiana site planned |
| [Microsoft](/wiki/microsoft) | ~$120 billion | [Azure](/wiki/azure) AI, [OpenAI](/wiki/openai) partnership, Copilot services | $80 billion Azure backlog constrained by power availability |
| Oracle | ~$50 billion | OCI cloud GPU infrastructure, Stargate partnership | Key infrastructure partner for OpenAI's Stargate project |
| CoreWeave | Growing rapidly (post-IPO) | GPU cloud for AI startups and enterprises | $22.4 billion OpenAI commitment; 32 datacenters, 250,000+ GPUs [2] |

The aggregate spending figures are staggering. Big tech companies invested an aggregate of roughly $400 billion in 2025, with plans to increase further in 2026, and roughly 75% of the 2026 spend (about $450 billion) is tied directly to AI infrastructure rather than traditional cloud [1]. This spending is driven by the conviction that AI capabilities will generate returns that justify the investment, though some analysts have raised concerns about the pace of spending relative to near-term revenue generation [1].

## How is an AI datacenter built? (Hardware architecture)

AI datacenters differ fundamentally from traditional facilities in their compute, networking, and cooling architecture.

### GPU clusters

The core compute unit of an AI datacenter is the GPU cluster. [NVIDIA](/wiki/nvidia) dominates this market, with its H100, H200, and Blackwell-generation (B100, B200, GB200, B300) GPUs deployed at massive scale. A single GPU rack in a modern AI datacenter might contain 8 to 72 GPUs, consuming 10 to 130+ kW of power.

The NVIDIA DGX and HGX platforms package 8 GPUs per node with high-bandwidth interconnects. The GB200 NVL72 system packages 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled rack, delivering approximately 720 PFLOPS of FP8 training compute [3]. A fully loaded GB200 NVL72 rack draws on the order of 120 to 132 kW, far beyond the practical ceiling of air cooling, which is why NVIDIA specifies the system as liquid-cooled [3].

Alternative AI accelerators are gaining traction:
- **Google TPU** (Tensor Processing Unit): Custom ASIC designed for [TensorFlow](/wiki/tensorflow) and [JAX](/wiki/jax) workloads; TPU v5p and v6 (Trillium) deployed in Google's datacenters
- **AMD Instinct MI300X/MI325X**: Competing GPU architecture with large HBM capacity
- **AWS Trainium**: Amazon's custom AI training chip, with Trainium2 deploying in 2025-2026
- **Intel Gaudi**: [AI accelerator](/wiki/ai_chip) acquired from Habana Labs

### How are AI datacenters networked? (NVLink and InfiniBand)

AI training workloads require moving massive amounts of data between GPUs, both within a server and across servers. Two interconnect technologies dominate:

**NVLink** is NVIDIA's proprietary GPU-to-GPU interconnect for communication within a node or rack. The latest NVLink generation (NVLink 5, used in Blackwell) provides up to 1.8 TB/s of bidirectional bandwidth between GPUs. NVLink allows GPUs within a node to share memory and operate as a unified compute fabric [4].

**InfiniBand** has traditionally dominated inter-node networking in AI clusters. NVIDIA's Quantum InfiniBand switches (acquired through the Mellanox acquisition) provide 400 Gb/s to 800 Gb/s per port with ultra-low latency. InfiniBand's advantage lies in its RDMA (Remote Direct Memory Access) capability, which allows GPUs to read and write directly to each other's memory without CPU involvement.

However, a significant shift is underway. By mid-2025, Ethernet has taken the lead in new AI backend network deployments, driven by the maturation of the Ultra Ethernet Consortium specifications and hyperscaler validation of RoCE (RDMA over Converged Ethernet) at scale. NVIDIA's Spectrum-X Ethernet platform, Broadcom's Tomahawk 6, and AMD's Pensando all incorporate adaptive routing and hardware-level congestion management optimized for AI traffic patterns [4].

| Interconnect | Scope | Bandwidth (2025) | Latency | Market Trend |
|---|---|---|---|---|
| NVLink 5 | Intra-node / intra-rack | 1.8 TB/s bidirectional | Sub-microsecond | Dominant for GPU-to-GPU within a node |
| InfiniBand NDR/XDR | Inter-node (cluster fabric) | 400-800 Gb/s per port | ~1 microsecond | Declining market share but still strong in HPC |
| AI-optimized Ethernet (Spectrum-X, etc.) | Inter-node (cluster fabric) | 400-800 Gb/s per port | ~1-2 microseconds | Growing rapidly; now majority of new deployments |

### How are AI datacenters cooled?

AI hardware generates far more heat per rack than traditional server equipment. A standard enterprise server rack dissipates 5 to 15 kW; an AI GPU rack can dissipate 40 to 130+ kW. This density requires advanced cooling approaches:

**Air cooling** remains in use for lower-density deployments but is reaching its practical limits for AI workloads. Even optimized hot-aisle/cold-aisle configurations with rear-door heat exchangers struggle above 30-40 kW per rack.

**Direct liquid cooling (DLC)** pipes coolant directly to cold plates mounted on GPUs and CPUs. This is the standard for Blackwell-generation deployments; NVIDIA's GB200 NVL72 requires liquid cooling. Liquid cooling penetration in AI datacenters is projected to rise from roughly 14% in 2024 to about 33% in 2025 as Blackwell racks ship in volume [5]. The global liquid cooling market for datacenters is projected to grow from $5.65 billion in 2024 to over $48 billion by 2034 [5].

**Immersion cooling** submerges entire servers in dielectric fluid. While offering excellent thermal performance, it raises concerns about the use of fluorinated fluids (including PFAS, or "forever chemicals") that persist in the environment.

Cooling accounts for 30% to 40% of total datacenter energy consumption, making thermal efficiency a critical factor in operational costs and environmental impact [5].

## How much power does an AI datacenter use?

The power demands of AI datacenters dwarf those of traditional facilities and are reshaping energy infrastructure planning globally.

### Scale of power consumption

A single modern AI training cluster can consume 50 to 100+ MW of power. Major AI datacenter campuses are being planned at scales of 500 MW to multiple gigawatts:

| Facility / Project | Power Capacity | Operator | Status (as of early 2026) |
|---|---|---|---|
| xAI Colossus (Memphis, TN) | 150 MW (Phase 1), expanding to 2 GW | xAI (Elon Musk) | Phase 1 operational; expansion in progress |
| Stargate (Abilene, TX + 5 additional sites) | Target: 10 GW across all sites | OpenAI / SoftBank / Oracle | First Texas site operational; 5 additional sites announced [6] |
| Meta Louisiana campus | Up to 5 GW (long-term) | Meta | Planning/construction |
| Meta Ohio datacenter | 1 GW | Meta | Under construction |
| Microsoft Azure (global) | Multiple GW-scale sites | Microsoft | $80 billion backlog constrained by power [7] |
| Google (global) | Multiple sites across US, Europe, Asia | Google / Alphabet | Expanding TPU and GPU capacity |

Microsoft disclosed an $80 billion backlog of Azure orders that cannot be fulfilled due to power constraints, illustrating that power availability, not chip supply, has become the primary bottleneck for AI datacenter expansion [7].

The energy implications extend well beyond any single campus. IEA Executive Director Fatih Birol summarized the global picture this way: "Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today." [12] China and the United States together account for nearly 80% of the projected global growth in data center electricity use through 2030 [11].

## Notable Facilities

### xAI Colossus (Memphis, Tennessee)

Colossus is a supercomputer and datacenter built by [xAI](/wiki/xai) in Memphis, Tennessee. In one of the fastest datacenter builds on record, xAI transformed an abandoned Electrolux factory into an operational facility with 100,000 NVIDIA H100 GPUs in just 122 days, beginning in early 2024. The system was then doubled to 200,000 GPUs in another 92 days [8].

As of mid-2025, Colossus consists of 150,000 H100 GPUs, 50,000 H200 GPUs, and 30,000 GB200 GPUs. xAI claimed it was the largest AI training platform in the world at that time. The facility's primary purpose is training [Grok](/wiki/grok), xAI's conversational AI model. Elon Musk announced plans to expand to 2 GW with a third building at the Memphis site, aiming to have "more AI compute than everyone else combined" within five years.

### Stargate Project

The [Stargate](/wiki/stargate) Project, announced on January 21, 2025, is a joint venture between OpenAI, SoftBank, Oracle, and MGX (an Abu Dhabi investment fund) to build up to $500 billion in AI datacenter infrastructure in the United States over four years. SoftBank provides financial leadership while OpenAI holds operational responsibility [6]. OpenAI described the venture as intending "to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States," beginning with an immediate $100 billion deployment [6].

The initial $100 billion deployment is centered on two datacenters in Abilene, Texas. In September 2025, five additional sites were announced across Texas, New Mexico, Ohio, and the Midwest. The combined planned capacity approaches 7 GW and over $400 billion in committed investment, putting the project on track to reach its full $500 billion, 10 GW target ahead of schedule [6].

### CoreWeave Infrastructure

CoreWeave, which completed a $1.5 billion IPO in March 2025, has emerged as a key AI cloud infrastructure provider. With 32 datacenters and over 250,000 GPUs as of 2025, CoreWeave serves as a primary infrastructure partner for OpenAI ($22.4 billion commitment) and Meta ($14.2 billion). CoreWeave was the first cloud provider to deploy NVIDIA GB200 NVL72 (February 2025) and GB300 NVL72 (July 2025) commercially [2].

## What is the environmental impact of AI datacenters?

The rapid expansion of AI datacenters has raised significant environmental concerns across multiple dimensions. The relationship between [AI energy consumption](/wiki/ai_energy_consumption) and the power grid has become a central issue in technology and climate policy.

### Electricity consumption and carbon

AI datacenters are energy-intensive at a scale that is reshaping national and regional power grids. The International Energy Agency projects that global datacenter electricity consumption will more than double between 2024 and 2030, rising from about 415 TWh to about 945 TWh, with AI workloads driving most of the increase and electricity use at AI-optimized data centers set to more than triple [11]. From 2024 to 2030, data center electricity consumption is projected to grow about 15% per year, more than four times faster than total electricity demand from all other sectors [11].

The carbon impact depends heavily on the electricity source. Datacenters powered by coal or natural gas have a large carbon footprint, while those powered by renewables or nuclear energy have a much smaller one. This has created strong incentives for AI companies to secure clean energy sources.

### Water consumption

Datacenter cooling systems, particularly evaporative cooling towers, consume vast quantities of water. Google reported that its datacenters consumed approximately 5.6 billion gallons of water in 2023, a 24% increase over the previous year. Large facilities can consume 3 to 7 million gallons of water per day [9].

Annual US datacenter water consumption could double or quadruple by 2028 compared to 2023 levels, reaching 150 to 280 billion liters per year. Roughly 45% of existing datacenters are projected to face high water stress by the 2050s, particularly in regions like the southwestern United States, parts of Latin America, and Australia.

### Environmental impact summary

| Concern | Current Scale | Trend | Mitigation Strategies |
|---|---|---|---|
| Electricity use | Datacenters consume ~2-3% of US electricity (2025) | Rapidly increasing; could reach 6-8% by 2030 | Renewable PPAs, nuclear partnerships, efficiency improvements |
| Carbon emissions | Varies by grid mix; Google/Microsoft carbon goals slipping | Increasing in absolute terms | Clean energy procurement, carbon offsets, nuclear power |
| Water consumption | 5+ billion gallons/year for major hyperscalers | Increasing with expansion | Closed-loop cooling, air cooling where climate permits, waste heat recovery |
| Land use | Multi-thousand-acre campus sites | Expanding into rural areas | Brownfield redevelopment (e.g., xAI's factory conversion) |
| E-waste | GPU refresh cycles of 2-3 years | Growing concern | Recycling programs, extending hardware life |

## How are AI datacenters powered? (Nuclear and renewables)

The enormous and growing power needs of AI datacenters have driven a wave of energy partnerships, with nuclear power emerging as a particularly attractive option due to its ability to provide reliable, carbon-free baseload power.

In the United States, big tech companies signed contracts for over 10 GW of potential new nuclear capacity in 2024 alone. Notable partnerships include:

- **Microsoft** signed a 20-year power purchase agreement with Constellation Energy to restart the Three Mile Island Unit 1 nuclear reactor in Pennsylvania, providing 835 MW of carbon-free power for Microsoft datacenters
- **Google** announced partnerships with Kairos Power for small modular reactors (SMRs) to power its datacenters, with the first units targeted for deployment by 2030
- **Amazon** acquired a datacenter campus adjacent to the Susquehanna nuclear plant in Pennsylvania and invested in SMR developer X-energy
- **Meta** issued a request for proposals for nuclear power to support its AI infrastructure

Renewable energy also plays a significant role. All major hyperscalers have large-scale solar and wind power purchase agreements. However, the intermittent nature of renewables creates challenges for datacenters that require continuous power, driving interest in nuclear, geothermal, and long-duration energy storage as complementary solutions [10].

## Where are AI datacenters located? (Geographic distribution)

AI datacenter construction is concentrated in regions that offer favorable combinations of power availability, network connectivity, land costs, and regulatory environments:

**United States** remains the dominant location, with major clusters in:
- Northern Virginia ("Data Center Alley", the world's largest concentration)
- Texas (Austin, Dallas, Abilene for Stargate)
- The Midwest (Ohio, Iowa, Illinois, with cheap power and land)
- The Pacific Northwest (Oregon, Washington, with hydroelectric power)

**International expansion** is accelerating:
- **Europe**: Dublin, Amsterdam, Frankfurt, and the Nordics (cold climate, renewable energy)
- **Asia**: Singapore, Tokyo, and emerging Indian markets
- **Middle East**: The UAE (MGX fund, G42) and Saudi Arabia (NEOM) are investing heavily

The geographic distribution is heavily influenced by power availability. Microsoft's $80 billion Azure backlog is constrained primarily by the inability to secure sufficient power in desired locations, not by chip supply or construction capacity [7].

## Current State (2025-2026)

As of early 2026, the AI datacenter industry is defined by several dynamics:

**Unprecedented capital deployment.** Combined hyperscaler capex is on track to exceed $600 billion in 2026, with the majority directed toward AI infrastructure. This spending is driven by competitive pressure: companies fear that falling behind in AI infrastructure will result in losing the AI race entirely.

**Power as the binding constraint.** GPU supply, while still tight for the newest Blackwell chips, has improved significantly. Power availability has replaced chip supply as the primary bottleneck. Projects in power-constrained regions face delays of 2 to 4 years to secure sufficient electricity [7].

**Liquid cooling as the new standard.** The transition from air to liquid cooling is effectively complete for new AI deployments. All Blackwell-generation systems require liquid cooling, and datacenter operators like CoreWeave have committed to liquid cooling across all new facilities [2].

**[Inference](/wiki/inference) demand growing.** While training dominated the early wave of AI datacenter demand, inference workloads (serving AI models to end users) are growing rapidly. Inference has different requirements: lower latency, higher throughput per watt, and wider geographic distribution closer to users.

**Financial sustainability questions.** Some analysts question whether the current pace of spending can be sustained. Amazon faces projected negative free cash flow of $17 to $28 billion in 2026 from its datacenter investments. The industry's bet is that AI revenues will eventually justify these investments, but the timeline remains uncertain [1].

The AI datacenter buildout is likely to continue accelerating through 2026 and beyond, reshaping energy markets, real estate patterns, and the physical infrastructure of the global economy in the process.

## ELI5 (Explain Like I'm 5)

An AI datacenter is like a giant kitchen built just for cooking up AI. A normal computer building is full of regular computers that do everyday jobs like running websites. An AI datacenter is packed instead with thousands of super-powerful chips called GPUs, all bolted together so they can think about one huge problem at the same time, like teaching a chatbot to talk. These chips get extremely hot, so instead of just blowing air on them, special tubes carry cool liquid right onto the chips, like running cold water over a hot pan. They also gulp down enormous amounts of electricity, so much that companies are building their own power plants, including nuclear ones, just to keep them running.

## See also

- [Data center](/wiki/data_center)
- [GPU](/wiki/gpu)
- [NVIDIA](/wiki/nvidia)
- [Stargate](/wiki/stargate)
- [AI energy consumption](/wiki/ai_energy_consumption)
- [Inference](/wiki/inference)
- [Large language model](/wiki/large_language_model)

## References

1. "Hyperscaler capex > $600 bn in 2026, a 36% increase over 2025." *IEEE ComSoc Technology Blog* (2025). https://techblog.comsoc.org/2025/12/22/hyperscaler-capex-600-bn-in-2026-a-36-increase-over-2025-while-global-spending-on-cloud-infrastructure-services-skyrockets/
2. "CoreWeave Deep Dive: How a Former Crypto Miner Became AI's Essential Cloud." *Introl Blog* (2025). https://introl.com/blog/coreweave-gpu-cloud-ai-infrastructure-deep-dive-2025
3. "Gearing Up for the Gigawatt Data Center Age." *NVIDIA Blog* (2025). https://blogs.nvidia.com/blog/networking-matters-more-than-ever/
4. "InfiniBand vs Ethernet for AI Clusters: GPU Networks 2025." *Vitex LLC* (2025). https://www.vitextech.com/blogs/blog/infiniband-vs-ethernet-for-ai-clusters-effective-gpu-networks-in-2025
5. "AI's Cooling Problem: How Data Centers Are Transforming Water Use." *Environmental Law Institute* (2025). https://www.eli.org/vibrant-environment-blog/ais-cooling-problem-how-data-centers-are-transforming-water-use
6. "Announcing The Stargate Project." *OpenAI* (2025). https://openai.com/index/announcing-the-stargate-project/
7. "Tech AI spending approaches $700 billion in 2026, cash taking big hit." *CNBC* (2026). https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
8. "Colossus (supercomputer)." *Wikipedia*. https://en.wikipedia.org/wiki/Colossus_(supercomputer)
9. "How AI Growth Is Intensifying Data Center Water Consumption." *Net Zero Insights* (2025). https://netzeroinsights.com/resources/how-ai-intensifying-data-center-water-consumption/
10. "The hidden impacts of AI data centres on water, climate and future power costs." *Daily Maverick* (2026). https://www.dailymaverick.co.za/article/2026-02-10-the-hidden-impacts-of-ai-data-centres-on-water-climate-and-future-power-costs/
11. "Energy and AI: Executive summary." *International Energy Agency* (2025). https://www.iea.org/reports/energy-and-ai/executive-summary
12. "AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works." *International Energy Agency* (2025). https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works