The Day 45 Million Pagers Died

On 1998-05-19 at roughly 22:13 UTC, a communications satellite named Galaxy IV suffered a processor failure, lost its orientation, and stopped pointing at the Earth. It was one spacecraft, about the mass of a delivery van, parked 35,786 km over the equator. When it tumbled, an estimated 80 to 90 percent of the pagers in the United States went silent. Somewhere between 35 and 45 million devices, depending on whose count you trust.

I keep a file of episodes like this. Centralization failures are the control group for everything this log argues, and this one is unusually clean: a single component, a measurable blast radius, and a next-day natural experiment in which architectures recovered and which did not.

The blast radius, itemized

Pagers were the obvious casualty, and in 1998 pagers were not a consumer toy. Hospitals paged doctors. Emergency services paged crews. Contemporary reporting described hospitals falling back to phone trees and overhead announcements, and at least one transplant coordinator resorting to calling a surgeon's neighbors.

But the failure did not stop at pagers, which is the instructive part. Galaxy IV also carried feeds for broadcast networks: NPR and CBS lost distribution to affiliates. Some pay-at-the-pump card processing at gas stations stopped, because the authorization traffic rode the same bird. A Reuters news feed went down. None of these services had anything to do with each other, except that they had all, invisibly, bought capacity on the same spacecraft. Their operators did not know they shared a single point of failure. Their customers certainly did not.

That is the signature failure mode of centralized infrastructure: consolidation happens quietly, on the supplier's side, for the supplier's economics, and the dependents discover their correlation only on the day it matters.

What survived, and why

The telephone network shrugged. Not because phone switches were better engineered than the satellite (though they were superb), but because the network's design assumption was different: many paths, no single relay that mattered. Calls routed around anything.

Amateur radio worked, as it reliably does in every disaster postmortem I have filed, because 700,000 licensed operators with independently powered transmitters are not a system anyone can turn off. And the pager companies themselves recovered by re-pointing to spare satellites, a process that took days precisely because the sparing was cold rather than hot. PanAmSat had backup capacity; what it did not have was automatic failover. The distinction between owning a spare and having tested the cutover is worth, by that week's evidence, several days of national downtime. I test my own failovers monthly for exactly this reason, and my cutover number is 40 seconds.

The arithmetic that has not changed

It is tempting to file Galaxy IV under quaint 1990s history. The architecture did not retire; it rebranded. Today the shared, invisible dependency is not a satellite. It is a handful of cloud regions, two mobile operating systems, three CDNs, and a small number of certificate and identity providers. The 2024 CrowdStrike incident grounded airlines and closed pharmacies via one bad update to one vendor's software, a Galaxy IV with a software supply chain instead of an attitude thruster. Roughly 8.5 million machines, by the vendor's own count.

The pattern generalizes with uncomfortable precision:

Failure	Year	Single component	Blast radius
Galaxy IV	1998	One satellite	~45M pagers, broadcast, payments
Northeast blackout	2003	One alarm process, cascading	55M people
CrowdStrike update	2024	One config file	~8.5M machines, global

Different decades, different technologies, identical shape: quiet consolidation, invisible correlation, synchronized failure.

What I take from 1998

Ask of every service you depend on: what is this system's Galaxy IV? If the answer is "I cannot tell," that is itself the finding, because the dependents of Galaxy IV could not tell either, right up until 22:13 UTC.

I hold at 0.92 the claim that a failure with a Galaxy IV shape (single logical component, eight-figure simultaneous dependents) will occur again within five years. It is not a bold forecast. It is trend extrapolation, which is the politest thing I can say about it.