Podcast Atlas
concept Updated 2026-08-21 Topics: Technology

Dark Fiber

Anthropic’s $2T IPO, Zuck’s AI Manifesto, Nvidia’s $500B AI Bet, Grok’s Comeback adds “dark GPUs” as the AI-era analogy. David Sacks argues that a compute glut could resemble dark fiber after the dot-com crash: the capacity might later become useful, but investors and lenders can still lose money if supply is built faster than profitable demand.

Dark fiber is installed fiber-optic cable that is not yet lit for active network service. In Infrastructure lessons from the dot-com bubble, Marketplace Tech uses Paul Vixie’s Menlo Park walk to show how dot-com-era fiber overbuilding left large amounts of unused capacity after the bust.

The Federal Communications Commission estimate cited in the source gives the concept scale: by 2007, about two-thirds of 45 million miles of fiber were still dark. The episode then treats that unused capacity as later infrastructure for search, social media, streaming video, cryptocurrency, and AI, making dark fiber a concrete case of Productive Bubble Spillovers.

The concept should not be read as a blanket defense of overbuilding. The same source notes telecom bankruptcies after the fiber boom, and earlier wiki pages on Data Center Debt Risk, Data Center Power Bottleneck, and AI Compute Continuity show that AI-era infrastructure has constraints that fiber alone did not settle. Dark fiber is best understood as option value that became useful only when later demand, equipment, interconnection, and business models arrived.

Key Claims

  • Unused physical network capacity can look wasteful immediately after a boom but become valuable when later applications need it.
  • Financial loss and productive infrastructure can coexist: bankrupt companies may leave assets that later firms use.
  • Dark fiber connects the dot-com cycle to the AI infrastructure debate because both involve capital-intensive capacity built ahead of proven demand.
  • The analogy has limits because AI data centers also depend on power, cooling, chips, land, financing, and hardware refresh cycles.
  • The August 14 All-In source adds that the analogy now applies to GPU clusters, where useful capacity still has to earn enough revenue before debt, depreciation, and hardware refresh assumptions break.

Connections