We evaluate the potential performance of the ShakeAlert earthquake early warning system for M 9 megathrust earthquakes in the Pacific Northwest (PNW) using synthetic seismograms from 30 simulated M 9 earthquake scenarios on the Cascadia subduction zone. The timeliness and accuracy of source estimates and effectiveness of ShakeAlert alert contours are evaluated with a station‐based alert classification scheme using an alert threshold equal to the target threshold. We develop a population‐based alert classification method by aligning a population grid with Voronoi diagrams computed from the station locations for each scenario. Using raster statistics, we estimate the PNW population that would receive timely accurate alerts during an offshore M 9 earthquake. We also examine the range of expected warning times with respect to the spatial distribution of the population. Results show that most of the population in our evaluation region could receive alerts with positive warning times for an alert threshold of modified Mercalli intensity (MMI) III, but that late and missed alerts increase because the alert threshold is increased. An average of just under 60% of the population would be alerted for MMI V prior to the arrival of threshold level shaking. Large regions of late and missed alerts for thresholds MMI IV and V are caused by delays in alert updates, inaccurate FinDer source estimates, and undersized alert contours due to magnitude underestimation. We also investigate an alerting strategy where ShakeAlert sends out an alert to the entire evaluation region when the system detects at least an M 8 earthquake along the coast. Because large magnitude offshore earthquakes are rare in Cascadia, overalerting is most likely to occur from an overestimated M 7+ on the Gorda plate. With appropriate criteria to minimize overalerting, this strategy may eliminate all missed and late alerts except at sites close to the epicenter.