AI data centre costs to hit $10.3 trillion by 2032

AI data centre costs in the United States alone will total USD 10.3 trillion between 2025 and 2032, according to an opinion piece by Geoff Huston published on the APNIC blog on 30 September 2026. That figure reflects a known construction schedule of 183 gigawatts of AI data centre capacity being built by 2032, plus a further 118 gigawatts planned after that, and it equals almost 4% of US Gross Domestic Product. Huston argues that the 'free' consumer AI tools now being handed out are anything but free when you look at the infrastructure layer behind them.

The article starts by pointing out that 'free' has long been used to break into new markets. Google used it for search, Gmail and Docs, Meta for Facebook, and Cloudflare for web caching. The problem is that 'free' is often a misleading signal, because there are real operator costs that must be covered, and a zero price acts as a barrier to competition because rivals have to match it. Below-cost pricing, sometimes called price leading or dumping, is not new either. Huston cites Uber's USD 5 billion operating loss in a single three-month period in 2019 and its thirteen-year journey to revenue-positive status, with funding from SoftBank, the Saudi Arabian Public Investment Fund and Google. DoorDash lost USD 1.4 billion in 2022, with every delivery subsidized by its financial backers. AI, Huston writes, now follows the same playbook.

AI is expensive because of the sheer scale of computation needed to create the Large Language Models (LLMs) that power these systems. A modern AI data centre may require around 60,000 advanced Graphics Processing Units (GPUs), a type of chip designed for heavy parallel processing, deployed at a density of 72 GPUs per rack across about 3,000 to 8,000 racks. The GPUs must be mesh-connected to each other and to high-speed storage in a lossless connectivity fabric using 800G optics and high-density switches, and then there is mass storage. Each rack needs 150kW to 200kW of continuous power and liquid cooling, plus a large-scale liquid cooling plant and a large source of water to remove heat. Total power demand can reach one gigawatt, which means dedicated substations, high-voltage connections and significant grid-interconnection investment, because AI load profiles do not match the daily, weekly or seasonal patterns of homes and businesses, nor the periodic output of solar and wind generators.

A representative 200MW AI training campus currently costs around USD 8.2 billion, about two-thirds of which goes to IT equipment and one-third to real estate and power infrastructure. The known US construction schedule totals 183 gigawatts by 2032 and a further 118 gigawatts after that, leading Huston to estimate the total infrastructure spend from 2025 to 2032 at USD 10.3 trillion. That average investment would be almost 4% of US GDP, larger than previous national landmark boom investments in rail, electrification, transportation and telecommunications infrastructure.

Aggregate capital expenditures by the enterprise database and cloud services vendor, Microsoft, Amazon, Meta and Alphabet rose from about USD 97 billion in 2020 to more than USD 400 billion in 2025 and are projected to exceed USD 800 billion in 2026, surpassing their combined operating cash flow for the first time. Because this scale surpasses the companies' internal resources, they have broadened their financing channels. Data centre developers, infrastructure investment funds, private equity investors and bond markets provide equity capital, while banks, private credit funds and securitization vehicles supply debt. Vendor financing now also covers IT equipment. Huston notes that these arrangements tend to conceal rather than eliminate risk, because long-duration debt is supported by cash flows and collateral values that depend on uncertain AI demand and revenue models, rapid technological change, timely access to power and hardware, and the continued credit quality of a small number of tenants.

The resulting capital structure makes exposures more layered, correlated and difficult to observe. There is also technical obsolescence. The silicon chip industry has improved compute power, cost and energy use for about six decades, which means existing AI equipment depreciates quickly and requires constant reinvestment. But Huston warns that cheaper chips do not mean cheaper AI, because demand is growing faster than unit cost falls. Users want longer contexts, multimodality, agents, search, memory and task execution. If silicon progress stalls, any evolution in AI will require larger and more expensive assemblies of computational components, with matching demand for larger data centres and more power.

Debt is already growing. Meta closed 2025 with USD 72 billion in capital expenditure and anticipated between USD 115 billion and USD 135 billion in 2026. In February, Anthropic announced a USD 30 billion financing round at a USD 380 billion valuation, and Microsoft, NVIDIA and Anthropic formed an alliance under which Anthropic commits to buy USD 30 billion of Azure capacity and up to an additional gigawatt of compute. Google undertook a USD 32 billion global bond spree in 2026 across the US, Canada, Japan, Europe and Australia, while Alphabet indicated capital expenditures could approach roughly USD 175 billion to USD 185 billion in 2026 and issued a 100-year bond to tap pension funds and insurers.

Chip designers, chip fabricators, cloud operators, model developers, data centre operators and infrastructure investors are interlinked through long-term contracts, strategic investments and financing arrangements. A shock to one segment, such as weaker demand from model developers or the end of continuous silicon refinement, can propagate as a fall in service operator revenues, lease payments, asset values and creditor recoveries. Huston concludes that AI no longer looks like an experiment; it is a capital-intensive industry that will sooner or later demand profitability or dictate the terms of a rather impressive financial crash of global proportion.

The analysis then turns to revenues. Huston traces how Google used a classic two-sided market model, where search is free to consumers but Google assembles a profile of each consumer and sells it to advertisers. The more users use Google search and the more accurate the profiles, the more valuable those profiles are to advertisers. Economist Hal Varian, later Google's Chief Economist, argued in the late 1990s that spam was essentially a failure of information about the consumer, meaning more data would turn ads into helpful suggestions and raise the probability of a transaction. Google's advertising activity now earns USD 300 billion per year, so advertisers pay for search.

Advertisers did not simply increase their total advertising budgets. They shifted spending away from newspapers and free-to-air television. Huston writes that the newspaper business is now an impoverished shadow of its former self