Demand Usage Wrong
What If the AI Boom Looks Like the Fiber Glut? The Case for Doubt
The Lindsay Clancy trial has stretched into a sixth day of jury deliberations, with the defense attorney demanding removal of a holdout juror accused of ignoring reasonable doubt instructions — a demand the judge refused. The case, involving charges connected to the deaths of Clancy's children and centering on claims of postpartum psychosis, has become a painful public reckoning with the limits of legal frameworks for mental illness. Senator Ron Wyden is pressing the NSA to warn Americans about VPN security risks, specifically the concern that some providers with opaque ownership structures may be logging user traffic and sharing it with foreign intelligence services — a reminder that privacy tools can themselves become vectors of compromise. A whistleblower has claimed ICE skipped background checks in its rush to hire thousands of new officers, an institutional failure with direct humanitarian and political implications.
The most confident consensus claim in technology right now — the one most worth stress-testing — is the investment thesis underlying the AI infrastructure buildout: that falling AI token prices will drive explosive usage growth, and that usage growth will eventually justify the massive capital commitments already made. This is the standard technology diffusion model, and Goldman Sachs, multiple market analysts, and the dominant investment narrative all embrace it.
The assumption embedded in this thesis is that AI usage is highly price elastic — that enormous latent demand exists, currently suppressed only by cost. But that assumption may be wrong. Non-users may be non-users for reasons other than price: distrust of AI outputs, workflows that don't benefit from AI assistance, or cognitive overhead of integration that outweighs productivity gains. The historical parallel that warrants attention is the fiber optic buildout of the late 1990s. The investment thesis then was also demand-led: bandwidth will grow exponentially, so lay fiber ahead of that demand. Bandwidth demand did grow exponentially — eventually. But the timing mismatch between infrastructure investment and actual demand drove the dot-com crash, leaving billions of dollars of fiber sitting dark for years before demand caught up.
The counterargument is that the AI adoption curve is structurally different from fiber: every knowledge worker can see right now what AI assistance does for their work, and enterprise software integration is happening in months rather than years, suggesting a steeper adoption curve. The signal to watch — the leading indicator that would indicate the consensus is wrong — is enterprise AI contract renewal data. If companies that purchased AI licenses find that productivity gains do not justify the cost and begin quietly not renewing or downgrading, the usage-growth-justifies-infrastructure thesis starts to crack. Microsoft's Copilot renewal figures, expected in October, represent the clearest near-term window into whether enterprise AI is delivering on its productivity promise.