Why Building AI Data Centres Isn’t Working Anymore

Overview

The episode argues that the rapid AI data-center buildout—fueling recent AI expansion—is increasingly failing on real-world constraints. Rather than representing permanent, sustained growth, it may signal a coming “AI bubble” reset.

Key Claims and Evidence

Massive investment, but weak execution

• Four large tech companies were expected to spend $650B in 2026 on new data centers, with plans to scale to $9T by 2030.
• However:
  • Nearly half of US projects announced for the year were delayed or canceled.
  • Only about one-third appear to be actually under construction.
• The episode cites satellite imagery and reported delays, including claims that Microsoft facilities announced since 2023 haven’t been completed.

Communities are paying the cost

The show emphasizes public backlash, including:

• Noise from cooling systems
• Polluted/contaminated water concerns
• Higher local energy prices

It argues that tax incentives largely benefit companies, while local communities absorb costs through:

• Lost property-tax revenue
• Rate hikes

Examples cited

• Major utility rate increases tied to data-center load (e.g., Georgia Power rate hikes; Oregon consumers seeing large increases)
• Reports of sediment/water issues and persistent low-frequency hum

Buildout stalling due to bottlenecks beyond demand

The episode claims hyperscalers are encountering shortages in:

• Power infrastructure
  • Power draw described as comparable to large cities
  • Many projects not disclosing how they’ll power themselves
  • Transformer/switchgear supply constraints
  • Reliance on imported components from China
• Skilled labor
  • Difficulty hiring installation and fiber technicians quickly enough
• Financing stress and investor caution
  • Hyperscalers are spending at a scale beyond available cash
  • Investors are pulling back due to uncertain profitability

Supply-chain and hardware economics problems

• AI infrastructure is vulnerable to supply delays, where one delayed component can halt an entire project.
• It also notes memory shortages, claiming AI data centers absorb a large share of DRAM supply, contributing to sharp price increases for consumers.

Backlash is growing, and political tolerance may be shrinking

• Cancellations due to local opposition allegedly quadrupled in 2025.
• It describes community-level resistance, including:
  • State bans (or consideration of bans)
  • Lawsuits
  • The idea that anti-AI protests may be treated as potential security concerns

Possible “overbuilding” and demand uncertainty

The episode argues the strategy depends on assumptions that don’t necessarily hold, including:

• Infinite AI demand
• Unlimited grid capacity
• Reliable supply chains
• Sustained community tolerance

It also suggests:

• Open-source models and cheaper alternatives could reduce willingness to pay for large centralized compute.

Financing signals risk similar to past crises

• The episode cites data-center bond issuance rates that are high for “A”-rated paper (sometimes 8–12%), implying underappreciated risk.
• It draws a rhetorical comparison to conditions before the 2008 crisis, without claiming an identical system-wide banking failure.

Hyperscalers Pulling Back (and Projects Stalling)

• The episode points to deferrals/cancellations and quieter slowdowns by major providers, including claims that planned capacity expansion has been reduced.
• It uses a failed/unstable mega-project example (a large “AI campus” initiative) to illustrate how:
  • Leadership turnover
  • Lack of anchor tenants
  • Supply-chain misunderstandings can derail massive proposals.

Broader Alternatives Discussed

• Whether smaller/local models could reduce centralized infrastructure needs over time.
• The possibility of underwater data centers (already used in China) as a more efficient cooling approach.
• Even if data centers remain necessary, they may need to become smaller and less politically intrusive.

Overall Conclusion

The episode frames the current moment as a reality-check: large-scale data center investments are meeting practical limits (power, supply chains, financing, and community resistance) before returns are proven.

It does not claim data centers will disappear, but argues the industry may need to rethink:

• How much gets built
• Whether the payoff matches the scale of spending

Presenters / Contributors (as mentioned)

• Togo (host)
• Cold Fusion (channel/brand credited by host; no separate individual besides Togo)
• Revolut (sponsor; promotional mention only, no individual credited)