Skip to content
Advertisement
Citigroup lifts Big Tech AI infrastructure forecast to $2.8 trillion through 2029

Citigroup lifts Big Tech AI infrastructure forecast to $2.8 trillion through 2029

Photo: panumas nikhomkhai / Pexels

The bank's raised estimate collides with a power grid that has grown far slower than data center demand

Citigroup now expects Big Tech to spend more than $2.8 trillion on AI infrastructure through 2029. That is up from the bank’s prior estimate of $2.3 trillion, a half-trillion-dollar revision in one go.

Where the trillions are going

Citi’s upgrade reflects investment arriving sooner than the bank had modeled. The spending spans data centers, graphics processing units (GPUs), networking gear and power infrastructure.

The usual suspects are writing the checks. Microsoft, Amazon and Alphabet are leading the push, with rising enterprise demand for AI tools fueling the early surge.

The near-term number moved too. Citi now estimates hyperscaler AI capital expenditures will reach $490 billion by the end of 2026, up from its earlier call of $420 billion.

Geographically, the spending tilts heavily toward home. About half of the $2.8 trillion is expected to land in the US, approximately $1.4 trillion.

Advertisement

The Brookings Institution offers an even bigger frame. Its study puts US spending on data center construction, power infrastructure and computing equipment at $10.3 trillion between 2025 and 2032.

That works out to an annual average of 3.63% of US GDP. Brookings says this outpaces historical infrastructure booms such as the railroads and the interstate highway system.

The grid problem nobody can code around

Insurance broker Howden has quantified the gap. Since 2000, US data center power demand has grown at an average of 19% a year, while grid capacity has expanded just 2% annually.

According to Howden, data centers consumed 0.1% of grid electricity in 2000, a figure estimated to reach 5% by 2025.

That squeeze is already causing significant delays for new projects.

Brookings projects 182.8 GW of additional US data center capacity will be operational by the end of 2032. Globally, AI compute demand is projected to require an additional 55 GW of power capacity by 2030, at an estimated cost of $50 billion per GW.

Behind the meter, and beyond the grid

Howden estimates that by 2030, data centers will need to source up to 58% of their power demand, or 110 GW, from behind-the-meter solutions.

“Behind the meter” means generating electricity on site or nearby, without routing it through the public utility network.

What this means for investors and the market

Citi’s upgrade signals that, at least for now, the AI spending cycle is accelerating rather than cooling. A forecast that rises by $500 billion is a vote of confidence that hyperscalers will keep pouring money into capacity.

For suppliers of GPUs, networking equipment and data center construction, that implies a longer and larger order pipeline. The revision to 2026 hyperscaler capex, from $420 billion to $490 billion, points to that demand arriving soon rather than in some distant year.

The 19% versus 2% growth gap is the number worth watching. If grid expansion doesn’t speed up, or alternative power doesn’t fill the hole fast enough, project delays could push some of that projected spending further out.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.