AI data centers: the GPU campuses, chip supply, and power grids that decide the AI race

What it is

The "AI data centers" beat covers the physical infrastructure layer of AI: the purpose-built facilities, accelerator hardware, and power and cooling supply that enable frontier model training and inference. A modern AI data center is not a conventional server farm. It is a high-density campus housing thousands of GPU or AI accelerator cards, networked at ultra-low latency via InfiniBand or NVLink, powered at hundreds of megawatts, and cooled by liquid systems that consume millions of liters of water per day. Training clusters run large batch jobs for weeks; inference infrastructure serves real-time user requests and is the dominant cost center at scale. Facility capacity is measured in megawatts (MW) or gigawatts (GW) of installed power, a direct proxy for accelerator count.

History

Google's Tensor Processing Units, first deployed in its data centers in 2016, were the first large-scale custom AI silicon in production. Nvidia's H100 GPU, launched in 2022, became the de facto unit of AI compute. The late-2022 ChatGPT moment converted AI infrastructure from an R&D cost into a strategic imperative. Combined capital expenditure by Google, Amazon, Microsoft, and Meta reached approximately US$410 billion in 2025, nearly all of it AI infrastructure. In January 2025, OpenAI, SoftBank, Oracle, and MGX announced Stargate, a US$500 billion joint venture for a national US AI data center network. IEA data show AI-focused facilities tripled in capacity in the 18 months to early 2026, while their electricity consumption grew 50 percent in 2025 alone.

Current state

The four largest hyperscalers plan combined capex of approximately US$725 billion in 2026, up 77 percent from 2025. Amazon leads at US$200 billion; Microsoft guides to US$190 billion; Alphabet to US$175-185 billion; Meta to US$115-135 billion. Nearly all of it is AI: GPU clusters, custom silicon, and the facilities and power contracts to run them. The IEA projects global data center electricity consumption will reach approximately 945 TWh by 2030, nearly double 2024 levels, with AI-focused demand growing 30 percent per year. Around 20 percent of planned data center projects globally risk delay from grid connection constraints. On chips, Nvidia holds an estimated 80-85 percent of the AI accelerator market by revenue; AMD's MI300X and MI325X hold 5-10 percent, primarily in inference at Microsoft and Meta, with the MI400 series expected in H2 2026.

Relationships

The tracked subjects interlock across the supply chain. Hyperscaler capex is the spending signal: quarterly guidance from Amazon, Alphabet, Microsoft, and Meta is the most reliable leading indicator of GPU order volumes and construction pipelines. Nvidia's GB200 and GB300 Blackwell systems anchor most large cluster orders; AMD is the primary credible alternative. Stargate, anchored by OpenAI and SoftBank, is the most visible dedicated infrastructure vehicle, with seven US sites and more than 9 GW of planned capacity. Power demand is the acute constraint: IEA data show a 17 percent surge in data center electricity demand in 2025, and grid interconnection queues are lengthening across the US, Europe, and the Gulf. The HVDC transmission buildout and small modular reactor orders are direct infrastructure responses. Water (data-center-water) is the second physical limit: liquid-cooled racks consume millions of liters per day per campus, and siting now turns on water-table access as much as land cost or fiber. OpenAI's Jalapeño custom Broadcom chip illustrates how hyperscalers hedge chip concentration by commissioning ASICs outside Nvidia's supply chain. Micron's record HBM quarter shows how high-bandwidth memory supply, not GPU count, can become the binding constraint.

What to watch

Power connection queues: whether US and European grid operators clear the data center interconnection backlog or force multi-year delays. AMD's MI400 series in H2 2026 and whether it takes inference share from Nvidia at Microsoft and Meta. Stargate financing: Abilene, Texas, is the only fully operational site as of mid-2026; bond issuance for the remaining six campuses will test the US$500 billion commitment. Water regulation: Virginia, Texas, and Arizona are all considering limits on data center water withdrawals. Custom silicon: Google's TPUs and Meta's MTIA compete with Nvidia GB300 in training; their production performance will determine whether Nvidia's 80-85 percent share holds through 2027. The Cerebras-OpenAI 750MW deal is the clearest signal of whether wafer-scale inference chips can sustainably displace GPU clusters.