Highlights

• Hyperscalers deploying 20-30 GW of AI data center capacity could spend $800B-$1.2T, with each gigawatt requiring ~$42B in capex, including $23.6B in chips alone.
• AI infrastructure is materials-intensive, requiring NdPr magnets, gallium nitride, silicon carbide, and specialty alloys—with supply chains heavily concentrated in Asia, creating GDP leakage.
• Three critical bubble risks emerge: unproven AI revenue models, power grid interconnection delays stranding assets, and rapid GPU obsolescence compressing depreciation windows.

---

What happens when hundreds of billions of dollars are deployed into AI data centers before long-term monetization is proven?

That is the uncomfortable question beneath the latest capital intensity estimates. According to a widely circulated institutional analysis, building 1 gigawatt of AI data center capacity requires roughly $42 billion in capex. The breakdown is striking: about $28.7B in IT equipment (including $23.6B in chips alone), $5.5B in power systems, $2.1B in cooling equipment, and $6.2B in construction.

Scale that math. If hyperscalers collectively deploy 20–30 GW over this cycle, total spending approaches $800 billion to $1.2 trillion.

The GDP Multiplier — and the Leakage Problem

The projected U.S. GDP impact of roughly $30 billion per GW assumes meaningful domestic capture. That’s partially true. High-margin GPU economics accrue largely to U.S. firms. Construction activity is local and labor-intensive.

But the supply chain tells a more complicated story. Server assembly is globally distributed. Advanced substrates, optics, and networking hardware remain heavily Asia-based. Power electronics and magnet manufacturing are still import-exposed and mostly China at least for rare earth-related magnets. The GDP multiplier is real — but it leaks.

The Materials Reality: AI Is a Physical System

At Rare Earth Exchanges™, we’ve consistently emphasized that AI is not merely software — it is materials-intensive infrastructure.

High-efficiency HVAC motors and liquid-cooling pumps depend on NdPr permanent magnets, often stabilized with dysprosium and terbium. Advanced power conversion increasingly relies on gallium nitride (GaN) and silicon carbide. Copper, aluminum, and specialty alloys scale with compute density. As racks densify, cooling becomes the gating constraint, not compute.

So Where’s the Bubble Risk?

Three pressure points stand out:

1. Revenue uncertainty: AI workloads must generate durable cash flow tojustify unprecedented capital intensity. Where’s the Apps?
2. Grid constraints: Power interconnection delays could strand billions in idle assets.
3. Obsolescence cycles: Rapid GPU evolution compresses depreciation windows.

Other factors include water, a problem in much of the western USA.

Infrastructure build-outs can be strategic — but history warns that when capital outruns cash flow, volatility follows.

The now widely shared estimate placing 1 GW at $42 billion originated in an analysis attributed to Bridgewater. (opens in a new tab) Whether that number proves conservative or aggressive, the scale alone forces a sober question:

Are we financing durable infrastructure — or accelerating into the next capex supercycle overshoot?