Episode 13: AI Data Centres: The $30 Billion Price Tag
Why the Chips, Not the Building, Decide Whether the AI Build Pays Back
Summary
A one-gigawatt AI data centre in Abilene, Texas, is estimated to cost around $30 billion, and more than half of that goes on chips rather than the building. In this episode, Tim Hardwick follows the money from the concrete to the customer: how rising rates change the economics of the build, why chips lose their edge long before the building does, and what CXMT’s memory progress and Microsoft’s Maia 300 mean for pricing power. He also covers diesel diplomacy and why a late transformer can leave billions of dollars of chips sitting idle, and sets out what to watch from US inflation, ASML, TSMC and the banks.
Key points
- Of the reported $30bn cost, $17bn is chips, $3bn is power infrastructure and just $2.4bn is the building and its construction.
- On an illustrative $18bn of debt, every extra percentage point on rates adds $180m a year to the interest bill. Delays make that worse.
- The building lasts decades, while the chips start losing their edge as soon as faster ones arrive. The real question is how quickly the equipment earns back its cost.
- Building owner, cloud operator and AI customer are three separate streams of cash, and the contracts decide who carries the risk.
- CXMT’s denser memory design is a step forward, but yields, customer approval and shipment volumes have yet to be proven.
- Microsoft’s Maia 300 claims 40% more output per megawatt. That is a company claim, and it does not mean 40% off the cost of a data centre.
- Cheaper AI models could expand usage while making pricing harder to hold.
- The most expensive equipment doesn’t always decide the opening date. A late transformer or grid connection can.
Chapters
0:00 Introduction
0:09 Texas Data Centre Economics
1:14 The AI Supercycle Overview
3:19 Financing the Build
5:14 What $30 Billion Buys
8:19 Capital and Competition
10:53 Microsoft’s Custom Chips
13:36 Cheaper AI, Bigger Pressure
14:40 Resources and Geopolitics
17:41 Bringing It Together
21:10 What to Watch Next
22:53 Cash Versus Construction
23:58 Research and Wrap-Up
Read more
The Weekly Outlook and research behind this episode: https://qfmi.substack.com
Everything discussed is for information and education only. It is not investment advice or a recommendation to buy or sell any financial instrument.
Tags
AI Supercycle, AI data centres, Abilene, cost of capital, interest rates, semiconductors, memory chips, HBM, CXMT, Microsoft Maia 300, Nvidia, inference, diesel, power infrastructure, ASML, TSMC, US CPI, global macro
- (00:09) - Texas Data Centre Economics
- (01:14) - AI Supercycle Overview
- (03:19) - Financing the Build
- (05:14) - What $30 Billion Buys
- (08:19) - Capital and Competition
- (10:53) - Microsoft’s Custom Chips
- (13:36) - Cheaper AI, Bigger Pressure
- (14:40) - Resources and Geopolitics
- (17:41) - The PRISM Framework
- (21:10) - What to Watch Next
- (22:53) - Cash Versus Construction
- (23:58) - Research and Wrap-Up
