Maximum earthquake magnitude (mx) is a critical parameter in seismic hazard and risk analysis. However, some recent large earthquakes have shown that most of the existing methods for estimating mx are inadequate. Moreover, mx itself is ill‐defined because its meaning largely depends on the context, and it usually cannot be inferred using existing data without associating it with a time interval. In this study, we use probable maximum earthquake magnitude within a time period of interest, mp(T), to replace mx. The term mp(T) contains not only the information of magnitude limit but also the occurrence rate of the extreme events. We estimate mp(T) for circum‐Pacific subduction zones using tapered Gutenberg–Richter (TGR) distributions. The estimation of the two TGR parameters, β‐value and corner magnitude (mc), is performed using the maximum‐likelihood method with the constraint from tectonic moment rate. To populate the TGR, the rates of smaller earthquakes are needed. We apply the Whole Earth model, a high‐resolution global estimate of the rate of m≥5 earthquakes, to estimate these rates. The uncertainties of mp(T) are calculated using Monte‐Carlo simulation. Our results show that most of the circum‐Pacific subduction zones can generate m≥8.5 earthquakes over a 250‐year interval, m≥8.8 earthquakes over a 500‐year interval, and m≥9.0 earthquakes over a 10,000‐year interval. For the Cascadia subduction zone, we include the 10,000‐year paleoseismic record based on turbidite studies to supplement the limited instrumental earthquake data. Our results show that over a 500‐year period, m≥8.8 earthquakes are expected in this zone; over a 1000‐year period, m≥9.0 earthquakes are expected; and over a 10,000‐year period, m≥9.3 earthquakes are expected.