The murky economics of the data-centre investment boom

Original Article ↗ Hacker News Discussion ↗

Short-Term Incentives & Bubble Logic

  • Several comments frame the boom as classic “IBGYBG” behavior: executives, investors, and politicians reap short-term rewards (promotions, stock pops, “jobs created”) while long-term risks are discounted.
  • Data center approval is politically easy, money is abundant, and most actors are optimizing over a 3–5 year horizon, not over the life of the assets.

Circular Financing & Risk Concentration

  • Multiple posts highlight “circular” deals: AI companies pre-commit to enormous cloud spend; cloud providers borrow to buy GPUs; chipmakers invest back into the AI companies.
  • Examples cited include multi-hundred-billion or even trillion-scale commitments that far exceed current AI revenues, raising fears of Enron-style optics and manufactured growth.
  • Concern that when this unwinds, solid businesses will be dragged down alongside fragile ones, causing broader financial damage.

Company-Specific Debates

  • Debate over Google: some argue it’s uniquely insulated by ad cash flow and TPU economics; others see its AI unit economics as similar to peers and note heavy losses in non-ad ventures.
  • Oracle is viewed skeptically: dependent on loss-making AI customers, deeply borrowing for capex, and now revealed to have thin margins on GPU rentals.
  • There is anxiety around OpenAI’s massive envisioned buildout versus modest revenue, and around GPU vendors investing heavily in their own largest customers.

Tasmania & Siting of Data Centers

  • A highly valued Australian “AI data centre” startup with a Bitcoin-mining past and controversial founders is used as a bubble case study.
  • Thread disputes whether Tasmania is a good site: strong hydro power and renewables vs. limited transmission capacity and fragile international connectivity (few submarine cables).

Profitability, Real Demand & AGI Bets

  • Many see no clear path to sustainable AI profits beyond ads and premium cloud features; current token prices are viewed as artificially low and investor-subsidized.
  • Critics question whether end-user value (beyond spam, “slop,” and novelty) justifies the capex.
  • Others point to real compute shortages, unreliable major providers, and expect massive demand growth as AI tools permeate office work (e.g., spreadsheet agents) and as specialized inference ASICs emerge.
  • AGI/superintelligence is treated by some firms as a Pascal-style wager: overspend now to avoid missing a possibly transformative technology.

Cloud vs On-Prem & Post-Boom Assets

  • Discussion notes that cloud was always more expensive per unit than on-prem; for large, cash-rich companies, shifting back to owned or colo data centers can improve margins.
  • Rough sense that about half of current capex is in relatively durable infrastructure (land, buildings, power, cooling) and half in rapidly obsoleting GPUs, unlike the dark-fiber era where the long-lived asset dominated.