AI infrastructure cost: Where is America’s $1.4 trillion going?
From hungry data centres to power grid overhauls, America’s AI infrastructure bill will reach $1.4T next year. Here's a full breakdown of where the money is going.
AI is not running on magic. Behind every chatbot reply, coding assistant, and AI agent sits a very real and expensive physical world of chips, data centres, electricity, cooling systems, land and construction.
The US is expected to spend around $1.4 trillion on AI infrastructure in 2027, up from around $800 billion this year, according to comments by former White House AI official David Sacks at a G20 session on emerging technologies.
Sacks compared the scale of the build-out with America's 19th-century railway expansion, pointing to just how much capital is now flowing into the physical foundations of AI.
Why AI infrastructure is getting so expensive
The simplest reason is demand. Companies are moving beyond AI experiments and using the technology to answer customers, write code, analyse documents and automate workflows. That means AI systems need to handle more users and operate around the clock.
A modern AI data centre requires advanced GPUs or custom AI chips, high-speed networking, storage, backup power and cooling. Then there is the electricity needed to keep everything running. AI consumes power not only when models are trained but also every time someone uses them.
This second process is known as inference, which simply means the AI generating a response after receiving a prompt. A large part of America's infrastructure spending is therefore going towards data centres and semiconductors.
Nvidia's GPUs remain central to the AI boom, while major technology companies are also developing custom chips to improve efficiency and control computing costs.
The cost does not stop after the data centre is built
This is where the numbers become more interesting. Bryan Catanzaro, Nvidia's vice president of applied deep learning, said the computing costs for his team are currently far higher than the cost of its employees.
The comment highlights a key reality of today's AI economy: powerful AI can still be extremely expensive to run at scale. Currently, the industry expects those costs to fall as chips become more efficient and AI models require less computing power.
But for now, every additional AI task creates demand for more computing, electricity and infrastructure. Harvard Business Review argues that businesses may increasingly feel this cost themselves. AI providers have so far absorbed some of the high costs of GPUs, tokens and inference to encourage adoption.
As the market moves towards usage-based pricing, companies could eventually see their AI bills rise alongside usage.
AI needs a lot more than servers
All those machines need electricity, and that is becoming one of the biggest challenges for the industry.
New data centres can put pressure on local power grids and raise concerns about electricity prices.
David Sacks suggested AI companies could help address this by building their own power generation, including “behind-the-meter” systems that supply electricity directly to data centres. The infrastructure boom could therefore drive investment beyond technology, into power generation, transmission and construction.
For countries such as India, the lesson is hard to miss. Building an AI economy will require more than talented engineers and clever applications. Affordable power, data-centre capacity, stable policies and long-term infrastructure planning will matter just as much.
The AI boom may be powered by code, but building it will take a lot of steel, electricity, chips and money.