Microsoft cancels leases for AI data centers, analyst says

Original Article ↗ Hacker News Discussion ↗

Scale and nature of Microsoft’s lease cancellations

  • Commenters note “a couple hundred megawatts” is large but still modest relative to Microsoft’s stated ~$80B AI capex.
  • Confusion over the article’s wording: most conclude Microsoft was renting third‑party data center capacity and is backing out of some leases, not shutting down its own facilities.
  • Several speculate this is normal reallocation: pulling back from older/overpriced or overseas leases toward owned US capacity or better‑sited facilities, not an AI retreat.
  • Others stress this could simply reflect over-ordering during peak hype, now trimmed as demand and pricing become clearer.

Media reliability and claims of FUD

  • Strong criticism of Bloomberg’s past reporting (spy chip, subpoena stories, fake press release) leads some to frame this as potential market-moving FUD, especially around Nvidia.
  • Others push back: this story is based on a TD Cowen equity research note also cited elsewhere; bad or sensational reporting ≠ deliberate manipulation.
  • There is broader concern that inaccurate or context‑free headlines (“scaling back”, “no value”) distort nuanced moves like capacity rebalancing.

OpenAI, Oracle, and Stargate dynamics

  • TD Cowen notes OpenAI workloads shifting to Oracle; plus an OpenAI–SoftBank JV and “Stargate” plans to build huge independent capacity.
  • Some see this as OpenAI gradually reducing reliance on Microsoft, forcing Microsoft to emphasize inference for many customers rather than owning all training capacity.
  • Others counter that Microsoft still holds rights of first refusal for OpenAI compute and that external builds are upside if they drive more Azure inference.

AI capex, Nvidia, and bubble questions

  • DeepSeek’s lower-cost model and massive capex headlines prompt questions about sustainability of GPU demand and whether an “AI bubble” is starting to deflate.
  • Debate over shorting Nvidia: some argue its CUDA/software moat and lack of credible competition make shorting suicidal; others think cheaper training and overbuild will erode margins.
  • Several note a pattern of rumors and “hot takes” around accelerators and chipmakers that can whipsaw retail investors.

Is AI delivering real value?

  • Nadella’s remarks are read as: current AI hasn’t yet produced “Industrial Revolution”‑level productivity (e.g., 10% global growth), but could over time.
  • Thread splits between:
    • Heavy users who say LLMs are now indispensable (coding assistance, config generation, personal tasks).
    • Skeptics who find them unreliable, hype-driven, or only marginally better than search—especially for complex, high‑stakes or legacy‑system work.
  • Many expect broad adoption to be slowed by messy organizations, data quality, regulation, and unclear economics, even if the underlying tools keep improving.