Hyperscalers have already outspent most famous US megaprojects

Original Article ↗ Hacker News Discussion ↗

What’s Being Compared and How

  • Thread discusses a chart claiming hyperscaler/datacenter capex rivals or exceeds famous US “megaprojects” (railroads, interstate, Apollo, Manhattan Project, ISS, Marshall Plan, F‑35).
  • Several argue this mixes unlike things: decades-long public programs vs a short, still-ongoing private buildout; infrastructure vs targeted military/science projects.
  • Others say adjusting by GDP or time window changes the story; historical GDP estimates are noisy, so numbers have “wiggle room.”

Railroads, Infrastructure, and Durability

  • Railroads and highways are cited as long-lived, broadly enabling infrastructure with century-scale lifetimes and low marginal costs.
  • Some push back that not all track survived; many lines went bankrupt or were abandoned, and US rail expansion was driven partly by geopolitics and land subsidies, not just ROI.

AI Datacenters, GPUs, and Depreciation

  • Core contrast: rail/roads/fiber last 20–100+ years; GPUs and AI hardware are treated as 3–6 year assets, and performance-per-watt improvements incentivize rapid replacement.
  • Disagreement on whether GPUs are “shovels/consumables” or closer to structural steel in a bridge; all agree they are capital-intensive and power-hungry.

Economic Impact, ROI, and Bubble Risk

  • Many commenters worry about an AI/compute bubble analogous to 19th‑century railroad manias: massive capex “ahead of demand,” unclear profits, possible painful correction.
  • One cited survey claims most firms report positive GenAI ROI, but others note it measures perceived, not measured, returns and may be distorted by internal pressure.
  • Concern that private firms need fast payoff unlike public megaprojects, increasing systemic risk if returns disappoint.

AI Capabilities vs Hype

  • Enthusiasts see LLMs as already transformative (natural-language interfaces, code generation, translation, knowledge access).
  • Skeptics focus on hallucinations, weak disruption of search, questionable productivity gains, maintainability of “vibe-coded” software, and lack of a clear path to “true” AI.

Energy, Security, and Alternative Uses

  • Debate over whether datacenter buildout drives new power generation (especially renewables) vs mostly gas/coal.
  • Some note GPUs are also critical for simulations, scientific computing, robotics, and potentially warfare; others say classification/security constraints limit reuse of public-cloud GPUs.

Values and Priorities

  • Several lament that comparable or larger sums aren’t going to climate, space, or public infrastructure.
  • Others see large AI capex as a standard speculative cycle; market will eventually reprice if expectations are wrong.